Nothing, some people just want a consumer tower so they don't have to buy an iMac or Macbook Pro - a quad-core i7 + 680-ish GPU on the cheap.
A Mini can't dissipate that much heat without sounding like a vacuum cleaner though so it would have to be bigger and/or it would have to use some advanced cooling.
The Haswell Mini should at least satisfy people looking for a cheap, decent gaming box that also has a fast CPU.
Marvin why do you always associate a more rational mini with a consumer tower? It really makes no sense at all. The concept of the Mini has its place in Apples lineups but that doesn't mean that we need a machine like today's Mini. Apple really should focus on these issues to make the Mini more attractive and to recover from slipping sales;
Make the Mini more accessible! It is almost laughable that they even have a model that they call a server because servicing in the hard drives on the machine is a joke.
Beef up the power supply and cooling to handle a bit more wattage. In the past the idea here was to support a discrete GPU of respectable performance. With the advent of Haswell and follow on chips the need for descrete GPU chips isn't overwhelming but it would be nice to be able to buy a Mini without bottom of the barrel chips driving it.
Make a PCI Express based Solid State storage subsystem a part of the design. This ought to be easy in the Mini as they don't use all the lanes available to them anyways. Apple is the only company with the capability to do this at a rational cost point and such a Mini would be very nice performance wise. Give us two slots for PCI Express based storage and they can throw out support for magnetic drives.
Support more RAM in the machines.
This isn't asking a lot really. With the right approach the machine might even end up thinner which always seems to be a win at Apple. For example getting rid of mechanical drives saves on cooling needs, and power supply size requirements. This can lead to a thinner Mini. Less power used in storage means that more of the cooling capability can be steared to keeping the processor cool. This means we might be able to bump the processor by ten watts or so which can lead to far better performance.
In any event I'm not sure why you always seem to equate the need for an improved Mini with a desire for an XMac. They really are two different levels of hardware. As to solid state memory we have seen clearly what such memory can do for performance in the likes of the Mac Book AIRs. At this point SSD storage suitable for the Mini isn't that far out of the affordability range even for Apple hardware.
Quote:
Originally Posted by Marvin
Nothing, some people just want a consumer tower so they don't have to buy an iMac or Macbook Pro - a quad-core i7 + 680-ish GPU on the cheap.
A Mini can't dissipate that much heat without sounding like a vacuum cleaner though so it would have to be bigger and/or it would have to use some advanced cooling.
The Haswell Mini should at least satisfy people looking for a cheap, decent gaming box that also has a fast CPU.
You are right to an extent about the Minis cooling capacity but you should realize that removal of the hard drives significantly alters the cooling burden. If your primary heat load is one device then you can optimize for that. The thing here is that with Haswell and follow on chips even a ten watt gain in processor capability can lead to large performance deltas.
You pop up the word gaming again to imply that the desire for decent GPU performance is only linked to that. Sadly that makes you look pretty foolish. For the most part people dont buy Apple hardware for gaming desires so im not sure why you even bother beating that bush.
While somewhat reserved I do believe that Haswell can lead to a vastly improved Mini but I also believe that an architecture change can take Haswell far further in the Mini. I really just want to see the Mini get the same engineering respect that the laptop line gets. That means breaking new ground in a desktop computer. I'd also like to see the Mini become the good value it was in the past, it is far from that today.
It is an old architecture that isn't the good value it has been in the past! That is the simplest way to put it!
In a nut shell it needs to be modernized for better performance so that we don't have the massive gap between the Mini and Apples other desktop the Mac Pro. In a nutshell they need to look at what they have done with the laptops and transfer some of that tech to the Mini. One item to go for would be deletion of the internal hard drives in favor of solid state storage modules. This would lower the thermal burden allowing for either a higher performance processor or the return of a discrete GPU.
If Apple and Intel would get OpenCL working on Intels GPUs the need for a discrete GPU would melt away. If that isn't to be then they really need to support a discrete GPU so that modern software doing GPGPU computing doesn't suffer so much on the machine. Frankly I'd rather see Intel and Apple get off their butts in this regard, especially with Haswell and the follow ons near.
Other things I'd like to see include more TB ports, expanded RAM capability and a bundling of a keyboard. All of this without a price hike. If you are wondering yes I believe it is doable. If Haswell mobile lives up to billing and Apple leverages some of those new features it should actually be a bit cheaper to build a Mini while increasing performance. The addition of flash would impact the cost a bit in the wrong direction but Apple will have to do something sooner or later as the speed differential between secondary storage and the processor is so great now that it is almost a joke. If people think such a machine is impossible all I have to say is look at the laptop line up.
To sum it up the Mini is a stagnate platform, the concept hasn't changed in close to a decade now. A modern platform demands a rethink of what a base computer ought to be designed around.
You say a gaming box is that all this haswell processor is good for? The MM has more functions than gaming i am sure.
Oh of course. The mini is a fully functioning computer as well as HTPC. The thing will Haswell is that while the cpu speed is only supposed to be a 10-15% upgrade, the GPU is supposed to be a lot more. A 650M? Doubtable. A 640M with GDDR5 like in the iMac now? Hopefully with enough memory.
Marvin why do you always associate a more rational mini with a consumer tower? It really makes no sense at all.
...
In a nut shell it needs to be modernized for better performance so that we don't have the massive gap between the Mini and Apples other desktop the Mac Pro.
It always comes round to being about the space the iMac is in.
With the advent of Haswell and follow on chips the need for descrete GPU chips isn't overwhelming but it would be nice to be able to buy a Mini without bottom of the barrel chips driving it.
If NVidia chips are only 25-30% faster, it's not appropriate to say they bottom of the barrel. It has been a 100% difference in many cases in the past.
Make a PCI Express based Solid State storage subsystem a part of the design. Give us two slots for PCI Express based storage and they can throw out support for magnetic drives.
I don't get this obsession with PCIe storage. It's not going to make a huge difference to real world performance and is just going to restrict storage options. mSATA is fine if you want blade storage but the Mini is a low price option. PCIe storage would just raise the entry price.
You pop up the word gaming again to imply that the desire for decent GPU performance is only linked to that. Sadly that makes you look pretty foolish. For the most part people dont buy Apple hardware for gaming desires so im not sure why you even bother beating that bush.
It doesn't matter what the GPU is used for, the Haswell chip should perform adequately for it. OpenCL runs on the CPU and in some cases faster than the GPU. I tested a Core i7 with a 650M with LuxMark and the CPU is double the speed of the GPU. It'll vary between different tests but the Mini has a low power limit by design so they ramp the CPU and GPU up and down to maintain the power limit. They designed it to be a low power machine and you want to make it bigger, more expensive, hotter, noisier and less power efficient because as you've said many times, you want a consumer tower in the $1000-2000 range, which there isn't a market for any more.
It always comes round to being about the space the iMac is in.
If NVidia chips are only 25-30% faster, it's not appropriate to say they bottom of the barrel. It has been a 100% difference in many cases in the past.
I don't get this obsession with PCIe storage. It's not going to make a huge difference to real world performance and is just going to restrict storage options. mSATA is fine if you want blade storage but the Mini is a low price option.
Apple seems to still use SATA. I think the obsession was due to single ssds approaching the bandwidth limit of SATA III. Has Apple actually used mSATA? I thought they used their own tweaked implementation. Also regarding speed differences, you might note that he referred to what is actually supported and what actually works, not so much benchmarks on things that do run. NAND has dropped considerably in price. One factor is how much storage is required for the base mini. Apple's cto pricing can be quite high, but it is what it is, so which is why I emphasize the base amount.
Quote:
I tested a Core i7 with a 650M with LuxMark and the CPU is double the speed of the GPU. It'll vary between different tests but the Mini has a low power limit by design so they ramp the CPU and GPU up and down to maintain the power limit. They designed it to be a low power machine and you want to make it bigger, more expensive, hotter, noisier and less power efficient because as you've said many times, you want a consumer tower in the $1000-2000 range, which there isn't a market for any more.
Consumer variants of Kepler were supposedly less focused on computation than Fermi. I would have to look up all the details, but the tech sites reported this consistently. Even then your results puzzle me. I don't really care for LuxMark or Lux Render (which it's based on), but I still find it surprising. The general strength of GPGPU is in highly parallel computing, like you would have with render buckets, assuming the memory on the video card is enough to fully contain the information being processed. That was a complaint when NVidia released iray (not to mention botched implementation). Dealing with a scene of decent size can generate too much data for most gpu to hold considering that even television productions deal with GB of texture data.
It always comes round to being about the space the iMac is in.
If NVidia chips are only 25-30% faster, it's not appropriate to say they bottom of the barrel. It has been a 100% difference in many cases in the past.
For a given generation of chips, The Mini has always received bottom of the barrel. I'm not sure how you can deny this. It really isn't a case of What NVidia can do, it is a matter of the mini being designed with the lowest performance chips of the entire Mac line up each year.
I don't get this obsession with PCIe storage. It's not going to make a huge difference to real world performance and is just going to restrict storage options. mSATA is fine if you want blade storage but the Mini is a low price option. PCIe storage would just raise the entry price.
SSD would not have a significant impact on the Minis price. Sure there would be a slight increase in costs. but as we have seen on the AIRs the cost of the storage modules isn't a huge factor in pricing for Apple. The AIRs literally have replaced the old Mac Books at the low end.
The thing with PCIe storage is this, SATA is a dead end. It is just barely capable of keeping up with second generation solid state drives. As such it isn't the interface to design a long term platform around. Going PCIe is all about designing for the future.
SSDs can get as hot as if not hotter than HDDs so it doesn't change anything as far as cooling is concerned.
That depends upon the technology implemented in the SSD. Everything about the SSD industry is changing extremely fast, One of those things that is changing is the power profile for such storage.
It doesn't matter what the GPU is used for, the Haswell chip should perform adequately for it. OpenCL runs on the CPU and in some cases faster than the GPU. I tested a Core i7 with a 650M with LuxMark and the CPU is double the speed of the GPU.
Sadly this is meaningless as you can always find examples of software that runs to advantage in one place and not another. It is most interesting that the top 5 OpenCL processors, benchmark wise, are AMD GPU's. Even with that reality you have to recognize that benchmarks are often focused on specific types of compute uses. In the end the lack of Intel OpenCL GPU drivers is a demonstration of a lack of commitment and nothing more. Apps should be free to distribute their compute loads as they see fit.
It'll vary between different tests but the Mini has a low power limit by design so they ramp the CPU and GPU up and down to maintain the power limit. They designed it to be a low power machine and you want to make it bigger, more expensive, hotter, noisier and less power efficient because as you've said many times, you want a consumer tower in the $1000-2000 range, which there isn't a market for any more.
Is it really that difficult for you to separate what I say about the Mini from what I'd like to see in an XMac? You constantly do this and I really don't know why. I'm not talking about blowing out the Mini to handle 300 watts of GPU and CPU power, I'm talking about being able to add 10 to 20 watts to its capability. That might not sound like much but with the likes of Haswell it could be a huge step up in performance.
Suppose they got rid of the server last year and went with a dual-core i7 and the 640M with 512 MB of GDDR5 for $799. Could that have been feasible?
Possibly. It really depends upon how the wattages add up. The big problem is that I wouldn't pay $800 these days for a dual core machine. I look at it this way they managed to put much more into the Mac Book Pros.
I think the obsession was due to single ssds approaching the bandwidth limit of SATA III.
Only in sequential reads and in some cases writes. Random reads/writes are nowhere near it. With 500MB/s sequential writes, you'd fill up a 256GB SSD in 8.5 minutes so there's no urgency to use a faster interface.
Even then your results puzzle me. I don't really care for LuxMark or Lux Render (which it's based on), but I still find it surprising. The general strength of GPGPU is in highly parallel computing, like you would have with render buckets, assuming the memory on the video card is enough to fully contain the information being processed. That was a complaint when NVidia released iray (not to mention botched implementation). Dealing with a scene of decent size can generate too much data for most gpu to hold considering that even television productions deal with GB of texture data.
It could just be poor optimisation with LuxRender but the results are:
Core-i7 3615QM : NVidia 650M
basic scene = 5678K rays/sec : 1510K rays/sec
medium scene = 3746K rays/sec : 1387K rays/sec
complex scene = 2231K rays/sec : 1082K rays/sec
The memory never exceeded the GPU memory but in every case, the i7 was at least twice as fast.
CLBenchmark (only runs under Windows) showed that in some cases the GPU is ahead and in others, the CPU. I don't know why the raytrace failed on the 650M as it has the latest drivers:
In the OpenCL Galaxies demo, the i7 scores 195FPS/62GFLOPs, the 650M scores 479FPS/161GFLOPs.
Adobe CS apps show massive gains for just the NVidia GPU as it's CUDA-only but if it was OpenCL, it would run vector code on the CPU too.
A dedicated GPU is not necessary for decent levels of computation or gaming. As long as the IGP is powerful enough, when it's combined with a powerful CPU, it will run OpenCL code just fine. With more computation units like the 40 in Haswell vs 16 in Ivy Bridge, parallel tasks should run faster.
Power has to be taken into consideration too. A 250W GPU might perform 5x faster than a 95W CPU. When you factor in the power draw, it's not as impressive.
There's no question that GPU computation is better in many cases but it's not as impressive with low power GPUs.
the lack of Intel OpenCL GPU drivers is a demonstration of a lack of commitment
Intel had their own compute framework going for a while but they seem to be on board with OpenCL. It was an afterthought with Ivy Bridge but the Haswell IGP should be OpenCL 1.2 compliant and I expect it to ship with OpenCL drivers.
I'm talking about being able to add 10 to 20 watts to its capability. That might not sound like much but with the likes of Haswell it could be a huge step up in performance.
It depends what you consider to be a huge step up. You only get about 5GFLOPs per watt from a GPU so 10-20W won't double the performance. You can always justify adding a little more to something but is it going to make it any more compelling to the target audience? The target audience for the Mini is people who want a low-price desktop and server.
Let's say they made a Mini with a quad i7 and a 750M Kepler GPU with a 105W PSU and it was a bit bigger with accessible drive slots at the back. Would you buy it for $999 and if so, what would you be using the GPU for?
So the GPU is really the main factor with this Haswell processor.
Power saving too as they can shut down parts more efficiently. They ran a GPU test quite well at 8W. During idle or low usage, they'd be able to scale it down pretty low so a laptop might get an extra few hours of battery life.
To sum it up the Mini is a stagnate platform, the concept hasn't changed in close to a decade now. A modern platform demands a rethink of what a base computer ought to be designed around.
Given that the mini was released in 2005 and updated form factor in 2010 it's a wonder that you don't think that the iPhone, being a whole 2 years younger, isn't also a stagnate [sic] platform given that the concept hasn't changed in a while now either.
Gee...the MBP must be an even more stagnate [sic] platform too given the laptop concept has been around since the 80's.
Ow. Rolling your eyes really hard hurts.
For everyone that doesn't use the GPU heavily the current mini is an f-ing powerhouse for the form factor. When the Haswell update happens then the Mini will handle moderate GPU tasks.
"Don’t look now, but the new Mac minis are getting comparable to the last gen Xserve and 2010 Mac Pros as far as benchmarks. Tech progress marches on."
The biggest limitation is 16GB RAM. If you need 32+GB and a better GPU then the cheapest option today is a refurb 2010 MP on the Apple site for $1819.
Given that a tricked out mini is $1449 (16GB, Fusion, 2.6Ghz) that's not a bad deal even it is an older Nahelem CPU. It's anywhere from only $600-$700 more depending on who you source for aftermarket memory and SSDs if you don't buy that off the apple store...which I wouldn't.
For everyone that doesn't use the GPU heavily the current mini is an f-ing powerhouse for the form factor. When the Haswell update happens then the Mini will handle moderate GPU tasks.
I cannot wait for that. I am so psyched for Haswell. I hope I haven't given the impression that I think the 2012 Mac mini is terrible, I just think it is not good enough of an update for me to replace my current 2011 Mac mini. The Haswell GPU and quad-core processor I feel will be worthy of my money.
Only in sequential reads and in some cases writes. Random reads/writes are nowhere near it. With 500MB/s sequential writes, you'd fill up a 256GB SSD in 8.5 minutes so there's no urgency to use a faster interface.
It was an incomplete explanation on my part. I was thinking of my explanation in the thread regarding mac pros and the potential for 2TB SSDs. I don't expect them soon, but it made me think of their use for scratch data or files that require quick access where a larger volume would be essential. I was saying that future ssd generations would further eat away from the raid on a chip solutions as sizes grow to become appropriate for an increasing number of usage scenarios. A good internal raid card can run you $1000. Enterprise or standard drives are just a matter of how large the volume and what configuration you want to use, but it can get quite expensive and you generally don't want to fill more than half their capacity to ensure the required level of performance. SSDs will continue to leverage against such solutions as their capacities increase. This isn't to be confused with longer term storage or larger storage solutions, and I guess it's a bit less relevant to the mini. Since 2008 or so, SSDs have become significantly faster at peak speeds. With HDDs the SATA bus limit was a non issue for single drives. No one suggested drag racing them, which is really what you're suggesting with that 8.5 minute figure. 111111111111111111111..... Okay I clearly didn't sleep well last night if it's mid day and I find that funny.
Quote:
It could just be poor optimisation with LuxRender but the results are:
Core-i7 3615QM : NVidia 650M
basic scene = 5678K rays/sec : 1510K rays/sec
medium scene = 3746K rays/sec : 1387K rays/sec
complex scene = 2231K rays/sec : 1082K rays/sec
The memory never exceeded the GPU memory but in every case, the i7 was at least twice as fast.
CLBenchmark (only runs under Windows) showed that in some cases the GPU is ahead and in others, the CPU. I don't know why the raytrace failed on the 650M as it has the latest drivers:
In the OpenCL Galaxies demo, the i7 scores 195FPS/62GFLOPs, the 650M scores 479FPS/161GFLOPs.
Adobe CS apps show massive gains for just the NVidia GPU as it's CUDA-only but if it was OpenCL, it would run vector code on the CPU too.
A dedicated GPU is not necessary for decent levels of computation or gaming. As long as the IGP is powerful enough, when it's combined with a powerful CPU, it will run OpenCL code just fine. With more computation units like the 40 in Haswell vs 16 in Ivy Bridge, parallel tasks should run faster.
Adobe CS varies. Premiere and After Effects are CUDA. They started to migrate those earlier. Photoshop which moved later went the OpenCL route. A lot of these things seem to be used to increase responsiveness where realtime interaction would have been impossible with uncached data in an X86 implementation. Apps like Capture One have gone the OpenCL route to counter this as they moved away from cached low res previews with version 4 (2007 or 2008). DaVinci Resolve can use OpenCL for playback, but only on Macs. Windows requires CUDA, which is fine as NVidia tends to dominate professional graphics on the Windows side. Mari requires CUDA. I don't really use that one at all, but I'm familiar with it. The same goes for Resolve. I still have Resolve Lite installed.
Quote:
Power has to be taken into consideration too. A 250W GPU might perform 5x faster than a 95W CPU. When you factor in the power draw, it's not as impressive.
Barefeats used to post benchmarks on this kind of stuff. Those aren't really realistic use cases for that software, but I chose an example that is somewhat of a mass market application. GPU acceleration can be quite significant. Prior to capable gpus, companies like this were often more conservative on the tools they would build. Capture One is another example. I suspect Hasselblad will eventually do the same thing with Phocus. Generating high resolution previews that involve debayering, gamma correction, and a bunch of other adjustments can soak up a lot of cpu cycles.
Quote:
Not true, I don't see a quad-core i7 Air or 13" Macbook Pro. Even the iMac base models are all quad-core i5. The $799 Mini is quad-core i7.
I don't think intel will drop the use of dual cores too soon. They have focused much more on gpu improvements. Core count is likely to be a lower priority in low tdp chips until the gpu is considered up to snuff in these solutions. I suspect the low end of the discrete gpu market will not last unless this turns out to be a bubble.
Quote:
Intel had their own compute framework going for a while but they seem to be on board with OpenCL. It was an afterthought with Ivy Bridge but the Haswell IGP should be OpenCL 1.2 compliant and I expect it to ship with OpenCL drivers.
I wasn't aware of it. OpenCL isn't a bad thing. CUDA is only used in so many things because it was available in a stable form at an earlier date.
Quote:
It depends what you consider to be a huge step up. You only get about 5GFLOPs per watt from a GPU so 10-20W won't double the performance. You can always justify adding a little more to something but is it going to make it any more compelling to the target audience? The target audience for the Mini is people who want a low-price desktop and server.
That is a good question. I suspect highly parallel processes will become typically allocated to the gpu in the near future. I'm not even completely focused on the Mini here. I'm more curious what it will mean for things like the ipad, which may be able to take on additional types of applications.
Quote:
Let's say they made a Mini with a quad i7 and a 750M Kepler GPU with a 105W PSU and it was a bit bigger with accessible drive slots at the back. Would you buy it for $999 and if so, what would you be using the GPU for?
Power saving too as they can shut down parts more efficiently. They ran a GPU test quite well at 8W. During idle or low usage, they'd be able to scale it down pretty low so a laptop might get an extra few hours of battery life.
I noticed you mentioned Kepler, as Maxwell will be at least 2014. They could always return to AMD, although I think at the very least the mac pro needs at least one decent NVidia option given the presence of CUDA software. It doesn't necessarily have to be direct through Apple. It just has to be stable.
Prior to capable gpus, companies like this were often more conservative on the tools they would build.
Certainly the image processing apps seem to have jumped on it in a big way and that's really where the GPU shines with so many units working in parallel. Haswell will hopefully put in a good showing with more than doubling their execution units. With shared memory, Intel might even be able to shuttle the processing to the most appropriate parts of the CPU. They could process the first couple of cycles of a process on both the CPU and IGP and whichever was faster, use it for the rest of the task. OS X should really be able to do that too but context switching won't be as easy without shared memory.
I wasn't aware of it.
It used to be called Intel Ct but it looks like they might have renamed it to Cilk:
That's why I was wondering about the report of the AMD 7000 series drivers for the Mac Pro but it makes sense when they are the same GPUs as last year.
It's worded like it's a way of using PCIe directly (SATA bandwidth comes from available PCI lanes anyway) while maintaining backward compatibility. At least that is how your article describes it. I'm not sure if that is really the case, but it's worded that way.
Quote:
The SATA Express specification will enable development of new devices that can utilize the PCIe interface and maintain compatibility with existing SATA applications. The technology, SATA-IO claims, will provide a cost-effective means to increase device interface speeds to 8Gb/s and 16Gb/s.
Quote:
I reckon non-volatile RAM will be where we end up anyway:
At this point, ReRAM looks to cost several times as much as NAND flash. To make ReRAM memory, you have to take a process as complex as NAND manufacturing and then put another layer on top of that, he said.
Of course ram has a far higher cost per GB than solid state storage. The two don't have to be in line with one another to find a practical use. The first thing that came to mind was Apple's sleep states.
Quote:
Certainly the image processing apps seem to have jumped on it in a big way and that's really where the GPU shines with so many units working in parallel. Haswell will hopefully put in a good showing with more than doubling their execution units. With shared memory, Intel might even be able to shuttle the processing to the most appropriate parts of the CPU. They could process the first couple of cycles of a process on both the CPU and IGP and whichever was faster, use it for the rest of the task. OS X should really be able to do that too but context switching won't be as easy without shared memory.
It used to be called Intel Ct but it looks like they might have renamed it to Cilk:
Image processing apps have always used things like tiles and bucket systems, so they are a good example. They're also an area of familiarity to me, which is why I tend to reference them. What I would like to see is a really good set of editing tools that can work with data as close to its raw form as possible in something like 16 bit half float regardless of whether we're talking about still images or video data. Given that these things are often processed in tiles or buckets, they are a good area for gpu computation. 3D rendering is the same. It's just that the mind boggling amounts of data can be difficult with the ram limitations.
Quote:
A while ago it seemed like Intel wanted to dismiss GPU computation entirely and NVidia made fun of them when they said GPUs were only 14x faster:
Intel's claims there sound like posturing, but "only 14 times faster" is a somewhat silly thing to say unless NVidia's 14x faster solution was much more costly or difficult to implement. NVidia was first to market with a stable solution, and they show up frequently in a number of top ranked HPC solutions. Note top 500. It varies. Sometimes they're more prevalent in that list than others, but they do provide a viable solution to highly parallel number crunching.
Quote:
That probably encouraged them to focus more on their IGPs because they really seem to dislike NVidia.
AMD is renaming their GPUs this year too. It's the 7000 series rebranded as the 8000 series:
That's why I was wondering about the report of the AMD 7000 series drivers for the Mac Pro but it makes sense when they are the same GPUs as last year.
You may remember the litigation between the two companies a few years back. They are clearly in competition. Intel also wants to ensure that they don't eventually lose the low end of the market to ARM, so a balanced strategy is obviously needed. I could see the death of a greater range of low end discrete gpus. I don't know what will happen to the higher end ones without the ability to leverage low margin high volume areas to distribute development costs.
"At this point, ReRAM looks to cost several times as much as NAND flash. To make ReRAM memory, you have to take a process as complex as NAND manufacturing and then put another layer on top of that, he said."
Of course ram has a far higher cost per GB than solid state storage.
You can imagine having a 256GB SSD blade with 16GB ReRAM connected to a fast memory interface. It would effectively give you 256GB of RAM for about $350 but it would depend on how full the storage was - minimum would be 16GB. When loading data off the SSD, it would be extremely fast and with it being non-volatile, it maintains the exact OS state during restarts.
Given that these things are often processed in tiles or buckets, they are a good area for gpu computation. 3D rendering is the same. It's just that the mind boggling amounts of data can be difficult with the ram limitations.
Heterogeneous computing is really the best way forward for everything as can get rid of the memory limits. Certain kinds of processing work faster on the CPU as the earlier benchmarks showed so both need to be used in the right situation. It's good that OpenCL has taken off the way it has. I remember the days of Altivec and hardly anyone supported it.
IIntel's claims there sound like posturing, but "only 14 times faster" is a somewhat silly thing to say unless NVidia's 14x faster solution was much more costly or difficult to implement.
Intel said NVidia took what they wrote out of context but they didn't really. They just picked out the most favourable stat. At one point Intel says they tested 14 compute kernels and the 230W GTX 280 was on average 2.5x faster than the 130W Core i7-960. One important point they made is that developers tend to not optimise for CPUs as much as GPUs. This will partly be because they have to write the whole thing in vectorised OpenCL for it to work on a GPU.
They talk about the transcendental hardware in GPUs being used to speed up GPU acceleration. This probably led them to implement those in hardware with Sandy Bridge:
"There are other improvements within the EU. Transcendental math is handled by hardware in the EU and its performance has been sped up considerably. Intel told us that sine and cosine operations are several orders of magnitude faster now than they were in current HD Graphics."
These things all happened around the same time. Intel's push towards faster IGPs, Crystalwell being directly compared to NVidia GPUs not AMD, their Phi hardware aimed at the Tesla market and OpenCL support all seems to be motivated by a desire to compete with NVidia. Let's hope NVidia doesn't stop challenging them because as we can see with AMD without that challenge, Intel slows back down and makes very small improvements year after year.
I think this is Intel's strategy with Haswell. NVidia won over quite a few laptop manufacturers with Kepler. If Haswell matches the 650M while also offering huge power savings, it could be game over for NVidia in the low-end market because NVidia's GPUs wouldn't offer a significant enough advantage. Even the 21.5" iMac just has the 640M and 650M. I could easily see them going with the Haswell IGP.
I have mentioned this briefly before, though are we expecting Haswell to match the 640M or even the 650M with DDR3 or GDDR5? If it matches the latter, great. If not, eh... : /
are we expecting Haswell to match the 640M or even the 650M with DDR3 or GDDR5? If it matches the latter, great. If not, eh... : /
Intel seems to think it will match the 650M otherwise they'd probably compare it with a lower model:
[VIDEO]
The problem with a demo like Dirt 3 is they are running at 1080p, high quality no AA, which the 650M runs at 80FPS. Even if the Haswell was running at 640M speed ~64FPS, you can't tell the difference. They did get 40FPS in Heaven vs 20FPS on the Ivy Bridge:
The Ivy Bridge laptop in the video actually looks like a Macbook Air - they taped over the name.
Unigine Heaven can be configured with different settings though and on Notebookcheck, the HD 4000 scores 10FPS with the 640M around 24FPS. Intel claims they've doubled performance so that only really puts it at 640M class as the 650M gets 30FPS.
However, their test was in an Ultrabook with a 17W test chip, which will ship at 15W. So I reckon the 15W Haswell destined for the MBA will match the 640M (around double last year's Air) and the Haswell destined for the Mini and Macbook Pro will come closer to the 650M as it has the higher power limit - the highest ones go up to 57W now but they will likely use the 47W models in the Mini and MBP.
They should only use DDR3 memory for the expanded part of the memory. The Crystalwell part has 128MB of GDDR5. It won't be as good as a 650M GT with 512MB-1GB GDDR5 but the real-world difference should be negligible and it uses less power.
Whatever it comes out to, anything between the 640M and 650M will be a good enough GPU for the entry level machines - Apple uses the 640M on the entry iMac. It should get OpenCL 1.2 support from launch as well as OpenGL 4:
Comments
Nothing, some people just want a consumer tower so they don't have to buy an iMac or Macbook Pro - a quad-core i7 + 680-ish GPU on the cheap.
A Mini can't dissipate that much heat without sounding like a vacuum cleaner though so it would have to be bigger and/or it would have to use some advanced cooling.
The Haswell Mini should at least satisfy people looking for a cheap, decent gaming box that also has a fast CPU.
Marvin why do you always associate a more rational mini with a consumer tower? It really makes no sense at all. The concept of the Mini has its place in Apples lineups but that doesn't mean that we need a machine like today's Mini. Apple really should focus on these issues to make the Mini more attractive and to recover from slipping sales;
Make the Mini more accessible! It is almost laughable that they even have a model that they call a server because servicing in the hard drives on the machine is a joke.
Beef up the power supply and cooling to handle a bit more wattage. In the past the idea here was to support a discrete GPU of respectable performance. With the advent of Haswell and follow on chips the need for descrete GPU chips isn't overwhelming but it would be nice to be able to buy a Mini without bottom of the barrel chips driving it.
Make a PCI Express based Solid State storage subsystem a part of the design. This ought to be easy in the Mini as they don't use all the lanes available to them anyways. Apple is the only company with the capability to do this at a rational cost point and such a Mini would be very nice performance wise. Give us two slots for PCI Express based storage and they can throw out support for magnetic drives.
Support more RAM in the machines.
This isn't asking a lot really. With the right approach the machine might even end up thinner which always seems to be a win at Apple. For example getting rid of mechanical drives saves on cooling needs, and power supply size requirements. This can lead to a thinner Mini. Less power used in storage means that more of the cooling capability can be steared to keeping the processor cool. This means we might be able to bump the processor by ten watts or so which can lead to far better performance.
In any event I'm not sure why you always seem to equate the need for an improved Mini with a desire for an XMac. They really are two different levels of hardware. As to solid state memory we have seen clearly what such memory can do for performance in the likes of the Mac Book AIRs. At this point SSD storage suitable for the Mini isn't that far out of the affordability range even for Apple hardware.
Quote:
Originally Posted by Marvin
Nothing, some people just want a consumer tower so they don't have to buy an iMac or Macbook Pro - a quad-core i7 + 680-ish GPU on the cheap.
A Mini can't dissipate that much heat without sounding like a vacuum cleaner though so it would have to be bigger and/or it would have to use some advanced cooling.
The Haswell Mini should at least satisfy people looking for a cheap, decent gaming box that also has a fast CPU.
You are right to an extent about the Minis cooling capacity but you should realize that removal of the hard drives significantly alters the cooling burden. If your primary heat load is one device then you can optimize for that. The thing here is that with Haswell and follow on chips even a ten watt gain in processor capability can lead to large performance deltas.
You pop up the word gaming again to imply that the desire for decent GPU performance is only linked to that. Sadly that makes you look pretty foolish. For the most part people dont buy Apple hardware for gaming desires so im not sure why you even bother beating that bush.
While somewhat reserved I do believe that Haswell can lead to a vastly improved Mini but I also believe that an architecture change can take Haswell far further in the Mini. I really just want to see the Mini get the same engineering respect that the laptop line gets. That means breaking new ground in a desktop computer. I'd also like to see the Mini become the good value it was in the past, it is far from that today.
It is an old architecture that isn't the good value it has been in the past! That is the simplest way to put it!
In a nut shell it needs to be modernized for better performance so that we don't have the massive gap between the Mini and Apples other desktop the Mac Pro. In a nutshell they need to look at what they have done with the laptops and transfer some of that tech to the Mini. One item to go for would be deletion of the internal hard drives in favor of solid state storage modules. This would lower the thermal burden allowing for either a higher performance processor or the return of a discrete GPU.
If Apple and Intel would get OpenCL working on Intels GPUs the need for a discrete GPU would melt away. If that isn't to be then they really need to support a discrete GPU so that modern software doing GPGPU computing doesn't suffer so much on the machine. Frankly I'd rather see Intel and Apple get off their butts in this regard, especially with Haswell and the follow ons near.
Other things I'd like to see include more TB ports, expanded RAM capability and a bundling of a keyboard. All of this without a price hike. If you are wondering yes I believe it is doable. If Haswell mobile lives up to billing and Apple leverages some of those new features it should actually be a bit cheaper to build a Mini while increasing performance. The addition of flash would impact the cost a bit in the wrong direction but Apple will have to do something sooner or later as the speed differential between secondary storage and the processor is so great now that it is almost a joke. If people think such a machine is impossible all I have to say is look at the laptop line up.
To sum it up the Mini is a stagnate platform, the concept hasn't changed in close to a decade now. A modern platform demands a rethink of what a base computer ought to be designed around.
Quote:
Originally Posted by marvfox
What is wrong with the design?
You say a gaming box is that all this haswell processor is good for? The MM has more functions than gaming i am sure.
Oh of course. The mini is a fully functioning computer as well as HTPC. The thing will Haswell is that while the cpu speed is only supposed to be a 10-15% upgrade, the GPU is supposed to be a lot more. A 650M? Doubtable. A 640M with GDDR5 like in the iMac now? Hopefully with enough memory.
It always comes round to being about the space the iMac is in.
If NVidia chips are only 25-30% faster, it's not appropriate to say they bottom of the barrel. It has been a 100% difference in many cases in the past.
I don't get this obsession with PCIe storage. It's not going to make a huge difference to real world performance and is just going to restrict storage options. mSATA is fine if you want blade storage but the Mini is a low price option. PCIe storage would just raise the entry price.
SSDs can get as hot as if not hotter than HDDs so it doesn't change anything as far as cooling is concerned.
It doesn't matter what the GPU is used for, the Haswell chip should perform adequately for it. OpenCL runs on the CPU and in some cases faster than the GPU. I tested a Core i7 with a 650M with LuxMark and the CPU is double the speed of the GPU. It'll vary between different tests but the Mini has a low power limit by design so they ramp the CPU and GPU up and down to maintain the power limit. They designed it to be a low power machine and you want to make it bigger, more expensive, hotter, noisier and less power efficient because as you've said many times, you want a consumer tower in the $1000-2000 range, which there isn't a market for any more.
Quote:
Originally Posted by Marvin
It always comes round to being about the space the iMac is in.
If NVidia chips are only 25-30% faster, it's not appropriate to say they bottom of the barrel. It has been a 100% difference in many cases in the past.
I don't get this obsession with PCIe storage. It's not going to make a huge difference to real world performance and is just going to restrict storage options. mSATA is fine if you want blade storage but the Mini is a low price option.
Apple seems to still use SATA. I think the obsession was due to single ssds approaching the bandwidth limit of SATA III. Has Apple actually used mSATA? I thought they used their own tweaked implementation. Also regarding speed differences, you might note that he referred to what is actually supported and what actually works, not so much benchmarks on things that do run. NAND has dropped considerably in price. One factor is how much storage is required for the base mini. Apple's cto pricing can be quite high, but it is what it is, so which is why I emphasize the base amount.
Quote:
I tested a Core i7 with a 650M with LuxMark and the CPU is double the speed of the GPU. It'll vary between different tests but the Mini has a low power limit by design so they ramp the CPU and GPU up and down to maintain the power limit. They designed it to be a low power machine and you want to make it bigger, more expensive, hotter, noisier and less power efficient because as you've said many times, you want a consumer tower in the $1000-2000 range, which there isn't a market for any more.
Consumer variants of Kepler were supposedly less focused on computation than Fermi. I would have to look up all the details, but the tech sites reported this consistently. Even then your results puzzle me. I don't really care for LuxMark or Lux Render (which it's based on), but I still find it surprising. The general strength of GPGPU is in highly parallel computing, like you would have with render buckets, assuming the memory on the video card is enough to fully contain the information being processed. That was a complaint when NVidia released iray (not to mention botched implementation). Dealing with a scene of decent size can generate too much data for most gpu to hold considering that even television productions deal with GB of texture data.
The thing with PCIe storage is this, SATA is a dead end. It is just barely capable of keeping up with second generation solid state drives. As such it isn't the interface to design a long term platform around. Going PCIe is all about designing for the future. That depends upon the technology implemented in the SSD. Everything about the SSD industry is changing extremely fast, One of those things that is changing is the power profile for such storage. Sadly this is meaningless as you can always find examples of software that runs to advantage in one place and not another. It is most interesting that the top 5 OpenCL processors, benchmark wise, are AMD GPU's. Even with that reality you have to recognize that benchmarks are often focused on specific types of compute uses. In the end the lack of Intel OpenCL GPU drivers is a demonstration of a lack of commitment and nothing more. Apps should be free to distribute their compute loads as they see fit. Is it really that difficult for you to separate what I say about the Mini from what I'd like to see in an XMac? You constantly do this and I really don't know why. I'm not talking about blowing out the Mini to handle 300 watts of GPU and CPU power, I'm talking about being able to add 10 to 20 watts to its capability. That might not sound like much but with the likes of Haswell it could be a huge step up in performance.
Possibly. It really depends upon how the wattages add up. The big problem is that I wouldn't pay $800 these days for a dual core machine. I look at it this way they managed to put much more into the Mac Book Pros.
So the GPU is really the main factor with this Haswell processor. Interesting thought.
Only in sequential reads and in some cases writes. Random reads/writes are nowhere near it. With 500MB/s sequential writes, you'd fill up a 256GB SSD in 8.5 minutes so there's no urgency to use a faster interface.
It could just be poor optimisation with LuxRender but the results are:
Core-i7 3615QM : NVidia 650M
basic scene = 5678K rays/sec : 1510K rays/sec
medium scene = 3746K rays/sec : 1387K rays/sec
complex scene = 2231K rays/sec : 1082K rays/sec
The memory never exceeded the GPU memory but in every case, the i7 was at least twice as fast.
CLBenchmark (only runs under Windows) showed that in some cases the GPU is ahead and in others, the CPU. I don't know why the raytrace failed on the 650M as it has the latest drivers:
In the OpenCL Galaxies demo, the i7 scores 195FPS/62GFLOPs, the 650M scores 479FPS/161GFLOPs.
Adobe CS apps show massive gains for just the NVidia GPU as it's CUDA-only but if it was OpenCL, it would run vector code on the CPU too.
A dedicated GPU is not necessary for decent levels of computation or gaming. As long as the IGP is powerful enough, when it's combined with a powerful CPU, it will run OpenCL code just fine. With more computation units like the 40 in Haswell vs 16 in Ivy Bridge, parallel tasks should run faster.
Power has to be taken into consideration too. A 250W GPU might perform 5x faster than a 95W CPU. When you factor in the power draw, it's not as impressive.
There's no question that GPU computation is better in many cases but it's not as impressive with low power GPUs.
Not true, I don't see a quad-core i7 Air or 13" Macbook Pro. Even the iMac base models are all quad-core i5. The $799 Mini is quad-core i7.
You can configure it with an SSD already - a 256GB SSD adds $300.
Intel had their own compute framework going for a while but they seem to be on board with OpenCL. It was an afterthought with Ivy Bridge but the Haswell IGP should be OpenCL 1.2 compliant and I expect it to ship with OpenCL drivers.
It depends what you consider to be a huge step up. You only get about 5GFLOPs per watt from a GPU so 10-20W won't double the performance. You can always justify adding a little more to something but is it going to make it any more compelling to the target audience? The target audience for the Mini is people who want a low-price desktop and server.
Let's say they made a Mini with a quad i7 and a 750M Kepler GPU with a 105W PSU and it was a bit bigger with accessible drive slots at the back. Would you buy it for $999 and if so, what would you be using the GPU for?
Power saving too as they can shut down parts more efficiently. They ran a GPU test quite well at 8W. During idle or low usage, they'd be able to scale it down pretty low so a laptop might get an extra few hours of battery life.
Quote:
Originally Posted by wizard69
To sum it up the Mini is a stagnate platform, the concept hasn't changed in close to a decade now. A modern platform demands a rethink of what a base computer ought to be designed around.
Given that the mini was released in 2005 and updated form factor in 2010 it's a wonder that you don't think that the iPhone, being a whole 2 years younger, isn't also a stagnate [sic] platform given that the concept hasn't changed in a while now either.
Gee...the MBP must be an even more stagnate [sic] platform too given the laptop concept has been around since the 80's.
Ow. Rolling your eyes really hard hurts.
For everyone that doesn't use the GPU heavily the current mini is an f-ing powerhouse for the form factor. When the Haswell update happens then the Mini will handle moderate GPU tasks.
"Don’t look now, but the new Mac minis are getting comparable to the last gen Xserve and 2010 Mac Pros as far as benchmarks. Tech progress marches on."
http://blog.macminicolo.net/post/34175374589/impressions-of-the-2012-mac-mini-updated
The biggest limitation is 16GB RAM. If you need 32+GB and a better GPU then the cheapest option today is a refurb 2010 MP on the Apple site for $1819.
Given that a tricked out mini is $1449 (16GB, Fusion, 2.6Ghz) that's not a bad deal even it is an older Nahelem CPU. It's anywhere from only $600-$700 more depending on who you source for aftermarket memory and SSDs if you don't buy that off the apple store...which I wouldn't.
I cannot wait for that. I am so psyched for Haswell. I hope I haven't given the impression that I think the 2012 Mac mini is terrible, I just think it is not good enough of an update for me to replace my current 2011 Mac mini. The Haswell GPU and quad-core processor I feel will be worthy of my money.
Quote:
Originally Posted by Marvin
Only in sequential reads and in some cases writes. Random reads/writes are nowhere near it. With 500MB/s sequential writes, you'd fill up a 256GB SSD in 8.5 minutes so there's no urgency to use a faster interface.
It was an incomplete explanation on my part. I was thinking of my explanation in the thread regarding mac pros and the potential for 2TB SSDs. I don't expect them soon, but it made me think of their use for scratch data or files that require quick access where a larger volume would be essential. I was saying that future ssd generations would further eat away from the raid on a chip solutions as sizes grow to become appropriate for an increasing number of usage scenarios. A good internal raid card can run you $1000. Enterprise or standard drives are just a matter of how large the volume and what configuration you want to use, but it can get quite expensive and you generally don't want to fill more than half their capacity to ensure the required level of performance. SSDs will continue to leverage against such solutions as their capacities increase. This isn't to be confused with longer term storage or larger storage solutions, and I guess it's a bit less relevant to the mini. Since 2008 or so, SSDs have become significantly faster at peak speeds. With HDDs the SATA bus limit was a non issue for single drives. No one suggested drag racing them, which is really what you're suggesting with that 8.5 minute figure. 111111111111111111111..... Okay I clearly didn't sleep well last night if it's mid day and I find that funny.
Quote:
It could just be poor optimisation with LuxRender but the results are:
Core-i7 3615QM : NVidia 650M
basic scene = 5678K rays/sec : 1510K rays/sec
medium scene = 3746K rays/sec : 1387K rays/sec
complex scene = 2231K rays/sec : 1082K rays/sec
The memory never exceeded the GPU memory but in every case, the i7 was at least twice as fast.
CLBenchmark (only runs under Windows) showed that in some cases the GPU is ahead and in others, the CPU. I don't know why the raytrace failed on the 650M as it has the latest drivers:
In the OpenCL Galaxies demo, the i7 scores 195FPS/62GFLOPs, the 650M scores 479FPS/161GFLOPs.
Adobe CS apps show massive gains for just the NVidia GPU as it's CUDA-only but if it was OpenCL, it would run vector code on the CPU too.
A dedicated GPU is not necessary for decent levels of computation or gaming. As long as the IGP is powerful enough, when it's combined with a powerful CPU, it will run OpenCL code just fine. With more computation units like the 40 in Haswell vs 16 in Ivy Bridge, parallel tasks should run faster.
Adobe CS varies. Premiere and After Effects are CUDA. They started to migrate those earlier. Photoshop which moved later went the OpenCL route. A lot of these things seem to be used to increase responsiveness where realtime interaction would have been impossible with uncached data in an X86 implementation. Apps like Capture One have gone the OpenCL route to counter this as they moved away from cached low res previews with version 4 (2007 or 2008). DaVinci Resolve can use OpenCL for playback, but only on Macs. Windows requires CUDA, which is fine as NVidia tends to dominate professional graphics on the Windows side. Mari requires CUDA. I don't really use that one at all, but I'm familiar with it. The same goes for Resolve. I still have Resolve Lite installed.
Quote:
Power has to be taken into consideration too. A 250W GPU might perform 5x faster than a 95W CPU. When you factor in the power draw, it's not as impressive.
Barefeats used to post benchmarks on this kind of stuff. Those aren't really realistic use cases for that software, but I chose an example that is somewhat of a mass market application. GPU acceleration can be quite significant. Prior to capable gpus, companies like this were often more conservative on the tools they would build. Capture One is another example. I suspect Hasselblad will eventually do the same thing with Phocus. Generating high resolution previews that involve debayering, gamma correction, and a bunch of other adjustments can soak up a lot of cpu cycles.
Quote:
Not true, I don't see a quad-core i7 Air or 13" Macbook Pro. Even the iMac base models are all quad-core i5. The $799 Mini is quad-core i7.
I don't think intel will drop the use of dual cores too soon. They have focused much more on gpu improvements. Core count is likely to be a lower priority in low tdp chips until the gpu is considered up to snuff in these solutions. I suspect the low end of the discrete gpu market will not last unless this turns out to be a bubble.
Quote:
Intel had their own compute framework going for a while but they seem to be on board with OpenCL. It was an afterthought with Ivy Bridge but the Haswell IGP should be OpenCL 1.2 compliant and I expect it to ship with OpenCL drivers.
I wasn't aware of it. OpenCL isn't a bad thing. CUDA is only used in so many things because it was available in a stable form at an earlier date.
Quote:
It depends what you consider to be a huge step up. You only get about 5GFLOPs per watt from a GPU so 10-20W won't double the performance. You can always justify adding a little more to something but is it going to make it any more compelling to the target audience? The target audience for the Mini is people who want a low-price desktop and server.
That is a good question. I suspect highly parallel processes will become typically allocated to the gpu in the near future. I'm not even completely focused on the Mini here. I'm more curious what it will mean for things like the ipad, which may be able to take on additional types of applications.
Quote:
Let's say they made a Mini with a quad i7 and a 750M Kepler GPU with a 105W PSU and it was a bit bigger with accessible drive slots at the back. Would you buy it for $999 and if so, what would you be using the GPU for?
Power saving too as they can shut down parts more efficiently. They ran a GPU test quite well at 8W. During idle or low usage, they'd be able to scale it down pretty low so a laptop might get an extra few hours of battery life.
I noticed you mentioned Kepler, as Maxwell will be at least 2014. They could always return to AMD, although I think at the very least the mac pro needs at least one decent NVidia option given the presence of CUDA software. It doesn't necessarily have to be direct through Apple. It just has to be stable.
SATA Express should be in place in plenty of time:
http://www.sata-io.org/technology/sataexpress.asp
http://www.computerworld.com/s/article/9235229/Speedy_8Gbit_16Gbit_SATA_Express_systems_coming_this_year
I reckon non-volatile RAM will be where we end up anyway:
http://www.desktopreview.com/default.asp?newsID=2016&News=Toshiba+reRAM+Non-Volatile+RAM
Certainly the image processing apps seem to have jumped on it in a big way and that's really where the GPU shines with so many units working in parallel. Haswell will hopefully put in a good showing with more than doubling their execution units. With shared memory, Intel might even be able to shuttle the processing to the most appropriate parts of the CPU. They could process the first couple of cycles of a process on both the CPU and IGP and whichever was faster, use it for the rest of the task. OS X should really be able to do that too but context switching won't be as easy without shared memory.
It used to be called Intel Ct but it looks like they might have renamed it to Cilk:
http://software.intel.com/en-us/intel-cilk-plus
A while ago it seemed like Intel wanted to dismiss GPU computation entirely and NVidia made fun of them when they said GPUs were only 14x faster:
http://blogs.nvidia.com/2010/06/gpus-are-only-up-to-14-times-faster-than-cpus-says-intel/
That probably encouraged them to focus more on their IGPs because they really seem to dislike NVidia.
AMD is renaming their GPUs this year too. It's the 7000 series rebranded as the 8000 series:
http://www.techpowerup.com/178404/AMD-Rebrands-Radeon-HD-7000-Series-GPUs-to-HD-8000-for-OEMs.html
That's why I was wondering about the report of the AMD 7000 series drivers for the Mac Pro but it makes sense when they are the same GPUs as last year.
Quote:
Originally Posted by Marvin
SATA Express should be in place in plenty of time:
http://www.sata-io.org/technology/sataexpress.asp
http://www.computerworld.com/s/article/9235229/Speedy_8Gbit_16Gbit_SATA_Express_systems_coming_this_year
It's worded like it's a way of using PCIe directly (SATA bandwidth comes from available PCI lanes anyway) while maintaining backward compatibility. At least that is how your article describes it. I'm not sure if that is really the case, but it's worded that way.
Quote:
The SATA Express specification will enable development of new devices that can utilize the PCIe interface and maintain compatibility with existing SATA applications. The technology, SATA-IO claims, will provide a cost-effective means to increase device interface speeds to 8Gb/s and 16Gb/s.
Quote:
I reckon non-volatile RAM will be where we end up anyway:
http://www.desktopreview.com/default.asp?newsID=2016&News=Toshiba+reRAM+Non-Volatile+RAM
It might be a little while.
Quote:
At this point, ReRAM looks to cost several times as much as NAND flash. To make ReRAM memory, you have to take a process as complex as NAND manufacturing and then put another layer on top of that, he said.
Of course ram has a far higher cost per GB than solid state storage. The two don't have to be in line with one another to find a practical use. The first thing that came to mind was Apple's sleep states.
Quote:
Certainly the image processing apps seem to have jumped on it in a big way and that's really where the GPU shines with so many units working in parallel. Haswell will hopefully put in a good showing with more than doubling their execution units. With shared memory, Intel might even be able to shuttle the processing to the most appropriate parts of the CPU. They could process the first couple of cycles of a process on both the CPU and IGP and whichever was faster, use it for the rest of the task. OS X should really be able to do that too but context switching won't be as easy without shared memory.
It used to be called Intel Ct but it looks like they might have renamed it to Cilk:
http://software.intel.com/en-us/intel-cilk-plus
Image processing apps have always used things like tiles and bucket systems, so they are a good example. They're also an area of familiarity to me, which is why I tend to reference them. What I would like to see is a really good set of editing tools that can work with data as close to its raw form as possible in something like 16 bit half float regardless of whether we're talking about still images or video data. Given that these things are often processed in tiles or buckets, they are a good area for gpu computation. 3D rendering is the same. It's just that the mind boggling amounts of data can be difficult with the ram limitations.
Quote:
A while ago it seemed like Intel wanted to dismiss GPU computation entirely and NVidia made fun of them when they said GPUs were only 14x faster:
http://blogs.nvidia.com/2010/06/gpus-are-only-up-to-14-times-faster-than-cpus-says-intel/
Intel's claims there sound like posturing, but "only 14 times faster" is a somewhat silly thing to say unless NVidia's 14x faster solution was much more costly or difficult to implement. NVidia was first to market with a stable solution, and they show up frequently in a number of top ranked HPC solutions. Note top 500. It varies. Sometimes they're more prevalent in that list than others, but they do provide a viable solution to highly parallel number crunching.
Quote:
That probably encouraged them to focus more on their IGPs because they really seem to dislike NVidia.
AMD is renaming their GPUs this year too. It's the 7000 series rebranded as the 8000 series:
http://www.techpowerup.com/178404/AMD-Rebrands-Radeon-HD-7000-Series-GPUs-to-HD-8000-for-OEMs.html
That's why I was wondering about the report of the AMD 7000 series drivers for the Mac Pro but it makes sense when they are the same GPUs as last year.
You may remember the litigation between the two companies a few years back. They are clearly in competition. Intel also wants to ensure that they don't eventually lose the low end of the market to ARM, so a balanced strategy is obviously needed. I could see the death of a greater range of low end discrete gpus. I don't know what will happen to the higher end ones without the ability to leverage low margin high volume areas to distribute development costs.
A hybrid solution would be great though:
http://www.engadget.com/2012/06/14/chou-university-builds-hybrid-nand-reram-unit/
You can imagine having a 256GB SSD blade with 16GB ReRAM connected to a fast memory interface. It would effectively give you 256GB of RAM for about $350 but it would depend on how full the storage was - minimum would be 16GB. When loading data off the SSD, it would be extremely fast and with it being non-volatile, it maintains the exact OS state during restarts.
Heterogeneous computing is really the best way forward for everything as can get rid of the memory limits. Certain kinds of processing work faster on the CPU as the earlier benchmarks showed so both need to be used in the right situation. It's good that OpenCL has taken off the way it has. I remember the days of Altivec and hardly anyone supported it.
The paper they wrote is here:
http://www.hwsw.hu/kepek/hirek/2010/06/p451-lee.pdf
Intel said NVidia took what they wrote out of context but they didn't really. They just picked out the most favourable stat. At one point Intel says they tested 14 compute kernels and the 230W GTX 280 was on average 2.5x faster than the 130W Core i7-960. One important point they made is that developers tend to not optimise for CPUs as much as GPUs. This will partly be because they have to write the whole thing in vectorised OpenCL for it to work on a GPU.
They talk about the transcendental hardware in GPUs being used to speed up GPU acceleration. This probably led them to implement those in hardware with Sandy Bridge:
http://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/5
"There are other improvements within the EU. Transcendental math is handled by hardware in the EU and its performance has been sped up considerably. Intel told us that sine and cosine operations are several orders of magnitude faster now than they were in current HD Graphics."
These things all happened around the same time. Intel's push towards faster IGPs, Crystalwell being directly compared to NVidia GPUs not AMD, their Phi hardware aimed at the Tesla market and OpenCL support all seems to be motivated by a desire to compete with NVidia. Let's hope NVidia doesn't stop challenging them because as we can see with AMD without that challenge, Intel slows back down and makes very small improvements year after year.
I think this is Intel's strategy with Haswell. NVidia won over quite a few laptop manufacturers with Kepler. If Haswell matches the 650M while also offering huge power savings, it could be game over for NVidia in the low-end market because NVidia's GPUs wouldn't offer a significant enough advantage. Even the 21.5" iMac just has the 640M and 650M. I could easily see them going with the Haswell IGP.
Intel seems to think it will match the 650M otherwise they'd probably compare it with a lower model:
[VIDEO]
The problem with a demo like Dirt 3 is they are running at 1080p, high quality no AA, which the 650M runs at 80FPS. Even if the Haswell was running at 640M speed ~64FPS, you can't tell the difference. They did get 40FPS in Heaven vs 20FPS on the Ivy Bridge:
http://www.notebookcheck.net/Intel-Keynote-Haswell-Dragon-Assistant-Atom-and-more.81667.0.html
The Ivy Bridge laptop in the video actually looks like a Macbook Air - they taped over the name.
Unigine Heaven can be configured with different settings though and on Notebookcheck, the HD 4000 scores 10FPS with the 640M around 24FPS. Intel claims they've doubled performance so that only really puts it at 640M class as the 650M gets 30FPS.
However, their test was in an Ultrabook with a 17W test chip, which will ship at 15W. So I reckon the 15W Haswell destined for the MBA will match the 640M (around double last year's Air) and the Haswell destined for the Mini and Macbook Pro will come closer to the 650M as it has the higher power limit - the highest ones go up to 57W now but they will likely use the 47W models in the Mini and MBP.
They should only use DDR3 memory for the expanded part of the memory. The Crystalwell part has 128MB of GDDR5. It won't be as good as a 650M GT with 512MB-1GB GDDR5 but the real-world difference should be negligible and it uses less power.
Whatever it comes out to, anything between the 640M and 650M will be a good enough GPU for the entry level machines - Apple uses the 640M on the entry iMac. It should get OpenCL 1.2 support from launch as well as OpenGL 4:
http://www.anandtech.com/show/6355/intels-haswell-architecture/12
"Haswell's processor graphics extends API support to DirectX 11.1, OpenCL 1.2 and OpenGL 4.0."
This might mean the next OS X has OpenGL 4 support as it will be supported on every machine.