If so, why would Apple need to do it online, except to "leak" the machine's performance to the public. However, judging from the unremarkable results, that's not very likely.
I don't know about that, they could leak the info to manage expectations.
You can twist the facts if you want but it is ten percent faster with half the sockets. Mind you thatis ten percent faster while running beta hardware and software.
It's still comparing a 3-year old top-of-the-line computer with a brand new top of the line computer. Honestly, people with Mac Pros already have enough space that the size won't make a difference; I mean the Mac Pro is already sitting there. A 10% increase in power over 3-year-old tech is not exactly mindblowing. I mean every year a new Macbook is released, Apple calls it "50% more powerful than last year's" and things like that - not "10% more powerful than 3 generations ago".
I suspect that you are simply looking for ways to be negative here. If nothing else this highlights just how well performance will scale on future machines. A process shrink and more thermally efficient design of the processor could lead to another GHz of performance in the same machine. This Mac Pro is a design built for the future.
It's still comparing a 3-year old top-of-the-line computer with a brand new top of the line computer. Honestly, people with Mac Pros already have enough space that the size won't make a difference; I mean the Mac Pro is already sitting there. A 10% increase in power over 3-year-old tech is not exactly mindblowing. I mean every year a new Macbook is released, Apple calls it "50% more powerful than last year's" and things like that - not "10% more powerful than 3 generations ago".
First you don't really have enough information to render a final verdict here. Second if you believe this year's macbook pro is 50% faster, you are out of your mind. The major gains of that magnitude have been accompanied by core count changes. Going from 2 to 4 cores is huge. As for the mac pro, it depends on pricing and if the extra GPGPU power is really leveraged. Assuming we're looking just at X86 performance, much of it comes down to price. If it's 10% faster in the same price range as we had in 2010, that is disappointing. They are only offering a single socket version. Other oems can go further, but Apple doesn't seem to be addressing that market. They also never offered Tesla co-processors, although those were arguably an even smaller niche. I can spot certain things I dislike, but I wouldn't form an entire opinion of a future machine based solely on that.
You folks are delusional if you think Intel's Xeon processors have jumped in performance to trounce the 2010 Mac Pro. They haven't. Intel has hit a wall and is already rearchitecting their CPU systems to follow AMD's path.
Well I have to agree that Intels incremental improvements in basic ALU operations has stalled a bit but let's not forget this CPU is running very slow and is still beating the old technology. Given a process shrink and other thermal management improvements this machine will be able to run much faster. Intel is expected to have a six core Haswell based chip later this year running at 130 watts and 3.6 GHz if I remember right.
The point is this machine could see a nice performance boost with a new generation of Xeons Gains in basic CPU performance aren't gone but it does require clock rate increase from here on out. This highlights another thing, people running software that is not heavily multithreaded may be better off with fewer cores. A six core machine will be very interesting when benchmarked with real apps. This assumes getting higher clock rates out of the six core implementation.
So while you are right that Intels performance increases from one generation to another has slowed considerably we still have a couple of generations of process shrinks ahead of us that should lead to much faster machines. The age of significant core improvements is gone so we are back to clock rate scaling and process shrinks.
Sorry, but the tradeoff with the Single CPU is we finally get Dual FirePro GPGPUs and together those retail for > $3k. Even if AMD cuts the price by $500 you still have $2500 in GPGPUs now.
The highest-end W9000 FirePros retail for $3k each. Two W5000s would be $800-900 for both worst case. I agree that Apple has quite rightly replaced a second Intel CPU with a much better value AMD OpenCL processor and it's up to developers to take advantage of it now.
This Mac Pro will most certainly cost much more than today's current and underpowered single socket Model at $2500+. Instead, this 12 core without the FirePro would be around the midpoint between the $2500 and $3800 models with old GPGPU technology. In short, $3150 + the new GPGPUs which will put this product well past $5k.
I don't think the entry model needs to cost more. There's not much reason to force expensive GPUs on everyone, especially people who will want this as a server. The highest-end model will be very expensive but they can make options that span all the way from $2-2.5k to $15k.
What you will get is continued expansion of power and life of a system because more and more of the OS and Application installed base is moving to leveraging OpenCL and GCD across the board. Adobe has recognized that in their Premiere Pro and CS Suite by working closely with AMD using OpenCL 1.2.
Yeah, all it would take is for some CPU intensive software to transition to OpenCL and the benefits will be very clear to see. All software that deals with image processing should be moving in this direction as well as mathematics software - Mathematica has already added support:
I really wish it wasn't always CUDA and OpenCL together. It's like the situation with DirectX and OpenGL. One of them is only compatible with a single vendor and the other is a standard. It's just the maturity of each API I suppose and what they are capable of but it really messes things up.
Wow a lot of people on MR freaking out....geez why don't we wait for some reviews and benchmarks once the machine has actually been released?
Agreed. This is pre-production hardware (with what is most likely a pre-production processor) running a beta version of the OS. There's no reason to take these results too seriously.
Agreed. This is pre-production hardware (with what is most likely a pre-production processor) running a beta version of the OS. There's no reason to take these results too seriously.
Pre-production / engineering sample geekbench scores have leaked before. They've changed somewhat with live hardware, but not that much. I think it's important to note that many of these users wouldn't have purchased something at the cost of the top spec 12 core. I would wait and what this one is like. If the machine is unsuitable due to tco or performance, you can always look at alternatives. I don't think they intend to throw the cost to performance ratio further out of balance. That would be a really bad move.
Pre-production / engineering sample geekbench scores have leaked before. They've changed somewhat with live hardware, but not that much. I think it's important to note that many of these users wouldn't have purchased something at the cost of the top spec 12 core. I would wait and what this one is like. If the machine is unsuitable due to tco or performance, you can always look at alternatives. I don't think they intend to throw the cost to performance ratio further out of balance. That would be a really bad move.
Personally I'd like to see how a six core machine performs. I don't know what wattage range Apple has designed the machine for but a 130 watt processor ought to be rather impressive. This is especially the case if they can turbo up to 4GHz or more.
Cinebench only goes up 6% due to the architecture. The biggest gain comes from core count increases.
Here's the result for dual 2.6GHz E5-2670 Sandy Bridge:
Per processor, that's 11.25. The core-count per CPU there is 8 and 2.6GHz, the E5-2697 v2 will be 12-core and 2.7GHz. So, expected Cinebench score for the new top-end Mac Pro would be:
1.06 x 11.25 x 12/8 x 2.7/2.6 = 18.5
The current 12-core gets 16 so it's an increase of about 15%. As I say though, potentially $1200 cheaper for that level of CPU performance because it's just a single CPU.
A Dell costs $7500 with dual E5-2687W with a score of 25 and the v2 will be 10-core at 3.4GHz vs 3.1GHz so 45% faster = score of 36.5.
If the old Mac Pro was updated with dual CPU, they'd have used the IB equivalent of E5-2665 x 2, which in SB is 8-core 2.4GHz. This would have hit the same $6200 price. This SB score is here:
It's about 19-20. Intel hasn't increased the clocks here much, if at all for IB so 6% improvement from the architecture with 10-cores vs 8, a $6200 Mac Pro would have scored 26 with dual CPUs. Not double the new Mac Pro.
This is similar to what they did with the 2012 refresh, which is bring the higher power in at lower price points rather than try making machines with the fastest CPUs at $7k+ price points because the volume of buyers isn't there. The volume buyers for those expensive CPUs are for blade servers.
Personally I'd like to see how a six core machine performs. I don't know what wattage range Apple has designed the machine for but a 130 watt processor ought to be rather impressive. This is especially the case if they can turbo up to 4GHz or more.
It could be the base model depending on how they structure it. Intel is likely to maintain a quad core in their Xeon E and EP lineups for Ivy. I suspect it depends on how Apple is going to organize and price the line. The really big gap is around the 12 core level. That could be an $1800 chip. The typical upper mid range Xeon 1600 chip that Apple is likely to use for their 6 core dropped to $600 mid westmere and stayed with that through Sandy. I suspect it will still be around $600 retail. That is higher than Apple has used in base machine so far, but I suspect they have a lot of leverage in the base model. Just breaking down what went into it suggested that the margins per unit were extremely high. I'm also not reading too much into the assembled in the US thing. I don't think they would move that if it drove the cost way up. The initial thread had too many clueless comments that were more based on preconceived notions rather than case studies. (Edit: yeah that's somewhat preconceived too, not feeling like linking buzz word articles that discuss "insourcing", but I don't think Apple views the bulk of their potential market as being in the $5-10k realm).
Quote:
Originally Posted by Marvin
This benchmark actually might be pretty accurate. Here's Anandtech's comparison of E3 Sandy Bridge to Ivy Bridge:
Cinebench only goes up 6% due to the architecture. The biggest gain comes from core count increases.
Same thing as I've stated without the nice collection of references. It should have been obvious to anyone. In both mainstream chips and EP versions, performance per clock cycle hasn't risen more than 15% even in favorable testing conditions. That would be why we disagreed regarding Moore's law a few months back (not digging up the thread). The increasing core counts have driven some pricing higher at the cpu level. As for gpus, I'm not not exactly sure what we'll see there. I haven't looked up the chips. I'm not as convinced as everyone else that it will be dual W9000s if AMD is sticking to their current pricing structure. Those came out mid last year. With extended cycles, pricing does sometimes drops very significantly on workstation variants. A good portion of the pricing is also due to extra driver development. There's a frequent debate over whether it's worth it and how most workstation cards are the same chips with different settings. Beyond guessing at the basic available chips, I'm unsure what Apple is buying. They're branded as firepros, but I think it's too early to read into a lot of this stuff. Assuming they stick to $2500 and up, they do need something better at the low end of that. There's no extra storage or memory capacity (allowing for lower density ram) to entice anyone, so they do need real differentiation.
Quote:
This is similar to what they did with the 2012 refresh, which is bring the higher power in at lower price points rather than try making machines with the fastest CPUs at $7k+ price points because the volume of buyers isn't there. The volume buyers for those expensive CPUs are for blade servers.
They've gone that route since at least 2011 with mainstream quad core options and greater emphasis on viable APUs.
Same thing as I've stated without the nice collection of references. It should have been obvious to anyone. In both mainstream chips and EP versions, performance per clock cycle hasn't risen more than 15% even in favorable testing conditions. That would be why we disagreed regarding Moore's law a few months back (not digging up the thread).
Moore's Law is about transistor count so they aren't far off that as they've gone from 6-core chips to 12-core in 3 years. We knew they were a year behind in their Xeon line. I'd have preferred if they moved to Haswell but AMD isn't keeping up with them on the CPU side so there's less incentive.
The real problem, which couldn't have been obvious until you knew the clock speeds and power consumption is performance-per-watt. The desktop/laptop chips all have IGPs so the Xeon chips should be able to move up further. Previously the Mac Pro used 2x $1440 95W (190W total) CPUs, now it will use 1x$2057 130W CPU and be around 15% faster.
Say that you get Y units of performance from the 190W chips, you now get 1.15Y units from 130W so performance-per-watt has improved 70% in 3 years, which is really the important factor and where ARM is doing better than Intel. At least the price is a bit lower.
Intel is increasing the core count ok but they are using lower clock speeds. They are dynamic, which complicates things but if these had 3.4GHz base clock and 3.8GHz Turbo, rather than 2.7GHz with 3.3GHz Turbo, it would have been what I expected from the single CPU.
I'm not as convinced as everyone else that it will be dual W9000s if AMD is sticking to their current pricing structure.
They might not be branded as W9000, these are custom parts. They just have the same spec. They are using the FirePro brand and they'll need some identifier to compare them though so I don't see the problem with them calling them W9000. I think we can be pretty sure they can't have the full $6k retail price.
We don't have all the facts, but there are a few things that can be deduced from what Apple has shown us about the Tube Mac Pro.
1. There must be 2 GPUs. Probability: High. When Apple decided to eliminate PCIe slots and use Thunderbolt, they looked at how many Thunderbolt ports to include, and decided 4 was too few, 6 sufficient. They probably didn't want to have some data-only Thunderbolt ports, so all are capable of driving a monitor. There is also an HDMI port, for a total of 7 display outputs. While the machine may not be able to drive all 7 simultaneously, by Apple design standards, any one should work.
The architecture would therefore be a single Xeon, 32 PCIe lanes to the two GPUs, 24 lanes to the Thunderbolt controllers (4 per TB connector), 6 displayport channels from the GPUs to the Thunderbolt controllers. The additional HDMI display channel could in theory come from the Xeon for some Xeon chips, but it's more likely that it's always connected to one of the GPUs, since it must be for the 12-core.
The allocation of display output bandwidth to DisplayPort output channels is done inside the GPUs. A FirePro GPU with several DisplayPort connectors can support high (i.e. 2560x1600) resolution on any connector, but not on all at once (personal experience). Apple states that the Tube can support up to 3 4K displays.
This design also limits which GPUs can be used, as the lower-end cards can't support 4 connections. It's possible, but unlikely, that Apple could have a low-end configuration with two different GPUs, one driving 3, the other 4 of the possible outputs.
2. By eliminating PCIe slots, and probably discontinuing the current Mac Pro (and assuming there is no new mid-level machine with PCIe slots), Apple will eliminate the market for 3rd-party video cards with MacOS drivers.
3. The GPUs are not upgradable. Probability: very high. The reason is simple: the GPUs are thermally mounted to the central heat sink, and Apple would never consider replacing thermal grease an end-user option. In fact, unless Apple has carefully designed it (which they certainly have done in the past), it won't be easy for a technician to replace a GPU card.
Also, since Thunderbolt 2.0 is limited to 4 PCIe lanes, an external high-performance GPU isn't an option. All modern GPUs use 16-lane PCIe slots.
4. The use of workstation-class FirePro GPUs will make it possible to run software like Solidworks on a machine as configured from the Apple store. There certainly will be users who want to use the Windows or Linux drivers.
5. The reason for speculation that the W9000 is used in the high-end configuration is that it's the only FirePro GPU with 6GB GDDR5 memory.
Seeing as how Apple states up front that the new Mac Pros use dual workstation-class GPUs as standard, the probability is beyond high…
Quote:
Originally Posted by davida
It's possible, but unlikely, that Apple could have a low-end configuration with two different GPUs, one driving 3, the other 4 of the possible outputs.
Not likely at all. See above. Apple is pushing this as what it is, a paradigm shift in the way we approach workstation computing. Why force more work on the CPU when a pair of (GP)GPUs can do some of that work faster & easier. Hence the push for developers to really start thinking about OpenCL at the latest WWDC.
Quote:
Originally Posted by davida
3. The GPUs are not upgradable. Probability: very high. The reason is simple: the GPUs are thermally mounted to the central heat sink, and Apple would never consider replacing thermal grease an end-user option. In fact, unless Apple has carefully designed it (which they certainly have done in the past), it won't be easy for a technician to replace a GPU card.
I do not think they are using thermal paste in this design. Look at the pics on the Apple website. Note the black metal brackets & knurled machine screws on the CPU & GPU cards. I believe they are using these items to clamp the chips directly to the Thermal Core, with a pad of thermally-conductive material between the chips & said Core. This also makes sense from an assembly standpoint; these three PCBs need to slot into the backplane AND align to the Thermal Core properly, thermal paste would be messy and could lead to unwanted damage to the PCBs while the worker is trying to assemble it all. Much easier to slot the PCBs into the backplane, drop that sub-assembly over the Thermal Core (which has the conductive pads already in place) and screw down the brackets/machine screws. Also much easier for a technician to replace a PCB if needed.
Moore's Law is about transistor count so they aren't far off that as they've gone from 6-core chips to 12-core in 3 years. We knew they were a year behind in their Xeon line. I'd have preferred if they moved to Haswell but AMD isn't keeping up with them on the CPU side so there's less incentive.
The real problem, which couldn't have been obvious until you knew the clock speeds and power consumption is performance-per-watt. The desktop/laptop chips all have IGPs so the Xeon chips should be able to move up further. Previously the Mac Pro used 2x $1440 95W (190W total) CPUs, now it will use 1x$2057 130W CPU and be around 15% faster.
Say that you get Y units of performance from the 190W chips, you now get 1.15Y units from 130W so performance-per-watt has improved 70% in 3 years, which is really the important factor and where ARM is doing better than Intel. At least the price is a bit lower.
Intel is increasing the core count ok but they are using lower clock speeds. They are dynamic, which complicates things but if these had 3.4GHz base clock and 3.8GHz Turbo, rather than 2.7GHz with 3.3GHz Turbo, it would have been what I expected from the single CPU.
I was partly referring to some of your performance projections. In some tiers they have added more performance than others, but keeping up with the highest core counts meant some price escalation. As for Haswell, Ivy wasn't originally predicted to hit 12 cores. I think they'll do okay. AMD isn't keeping up with them, but opterons still represent a good value with software that scales in a near linear manner with core counts, as AMD has much cheaper . They certainly aren't terrible, and they have regularly shown up in some form in HPC solutions as indicated by the top500. Their current top supercomputer should make Wizard happy, as it employs Xeon Phis. Number 2 is using Tesla solutions and opterons. Anyway from what you've read, has ARM really been on top of power efficiency at the server level? They don't have anything really aimed at workstations, but they do make server chips.
Quote:
They might not be branded as W9000, these are custom parts. They just have the same spec. They are using the FirePro brand and they'll need some identifier to compare them though so I don't see the problem with them calling them W9000. I think we can be pretty sure they can't have the full $6k retail price.
I need to find the chip data. I suspect much of is the same as their top gaming card, with more ram, clocked lower, and custom drivers. A lot of the cost of workstation gpus is typically driven by higher margins and a need to recoup R&D expenses on a much smaller number of units. With Macs they had to develop the low level mac drivers either way. It's also possible that the W9000s will be falling in price mid cycle in the near future, given that they have been out for a while. My point was that the price may not be highly aligned with the current $3k+ market price when the mac pro debuts, probably Q4, so yeah I agree with you there. I'm not pleased by the lack of internal storage for numerous reasons. I am unlikely to change my mind about that. It's just that if the machine proves viable for me overall, I won't avoid it due to that one detail
I guess I should clarify this. When I say there must be 2 GPUs, I mean the Thunderbolt-instead-of-PCIe design decision dictates 2 GPUs. And regarding the Apple web page, it states that the Mac Pro has two workstation GPUs with up to 6GB memory, but I hope this isn't the only configuration. It would bump up the minimum system price quite a bit if two W9000s/6GB were the only option.
Another way of looking at it: It's an inflexible design.
I do not think they are using thermal paste in this design. Look at the pics on the Apple website. Note the black metal brackets & knurled machine screws on the CPU & GPU cards. I believe they are using these items to clamp the chips directly to the Thermal Core, with a pad of thermally-conductive material between the chips & said Core. This also makes sense from an assembly standpoint; these three PCBs need to slot into the backplane AND align to the Thermal Core properly, thermal paste would be messy and could lead to unwanted damage to the PCBs while the worker is trying to assemble it all. Much easier to slot the PCBs into the backplane, drop that sub-assembly over the Thermal Core (which has the conductive pads already in place) and screw down the brackets/machine screws. Also much easier for a technician to replace a PCB if needed.
Thermal pads are obsolete, they wouldn't conduct enough heat away from the chips. Thermal paste can and should be a very thin layer. If you want to get an idea of how it's used, see these instructions:
I can assure you that thermal paste is being used, it's a requirement for the Xeon and the GPUs. It's possible they are using a heat spreader like a copper block between the chip and the aluminum core heat sink, but you would still use thermal paste.
The metal brackets with 4 screws are really calibrated springs compressing the GPU against the thermal core.
Thermal pads are obsolete, they wouldn't conduct enough heat away from the chips. Thermal paste can and should be a very thin layer. If you want to get an idea of how it's used, see these instructions:
I can assure you that thermal paste is being used, it's a requirement for the Xeon and the GPUs. It's possible they are using a heat spreader like a copper block between the chip and the aluminum core heat sink, but you would still use thermal paste.
The metal brackets with 4 screws are really calibrated springs compressing the GPU against the thermal core.
I worked for a good number of years in a production facility dealing with last mile wireless networking equipment for businesses, this was when a single wifi radio was about the size of a "Wheel of Time' paperback. Meaning, the units we produced were 4U rack-mounted computers, with the self-same Arctic Silver thermal compound (paste) nestled between the CPU & heat sink. So yeah, I know how to use thermal compound…
In retrospect, I guess they could use some AS, but I like the idea of a modern thermal pad.
However they do it, any tech tearing one of these things down to replace one of the three main chip bearing PCBs is still gonna have to take it almost completely apart to do so, by my figuring that is about 60 pieces (PCB assemblies, chassis parts, fasteners, brackets, etc.) to futz with.
I just hope the folks doing assembly know how to have a light hand with the thermal compound, it is not an item where more is better…
I guess I should clarify this. When I say there must be 2 GPUs, I mean the Thunderbolt-instead-of-PCIe design decision dictates 2 GPUs. And regarding the Apple web page, it states that the Mac Pro has two workstation GPUs with up to 6GB memory, but I hope this isn't the only configuration. It would bump up the minimum system price quite a bit if two W9000s/6GB were the only option.
Another way of looking at it: It's an inflexible design.
Just based on my reading of past Apple specification pages, that 'up to 6 GB memory' seems to certainly indicate there will be GPUs on offer with less than 6GB. I am no expert on these matters but I would have thought the price difference in RAM on the GPUs would not make that much difference these days, so the wording seems odd if that is true. Is it your take that there will be a lower RAM configurations and would that make a large price difference? Also is such GPU RAM upgradeable or is it a one time decision?
Comments
I don't know about that, they could leak the info to manage expectations.
I suspect that you are simply looking for ways to be negative here. If nothing else this highlights just how well performance will scale on future machines. A process shrink and more thermally efficient design of the processor could lead to another GHz of performance in the same machine. This Mac Pro is a design built for the future.
OUCH! If the test score number is near the same when the final 2013 Mac Pro is released, that is simply going to be embarrassing for Apple.
Quote:
Originally Posted by Superbass
It's still comparing a 3-year old top-of-the-line computer with a brand new top of the line computer. Honestly, people with Mac Pros already have enough space that the size won't make a difference; I mean the Mac Pro is already sitting there. A 10% increase in power over 3-year-old tech is not exactly mindblowing. I mean every year a new Macbook is released, Apple calls it "50% more powerful than last year's" and things like that - not "10% more powerful than 3 generations ago".
First you don't really have enough information to render a final verdict here. Second if you believe this year's macbook pro is 50% faster, you are out of your mind. The major gains of that magnitude have been accompanied by core count changes. Going from 2 to 4 cores is huge. As for the mac pro, it depends on pricing and if the extra GPGPU power is really leveraged. Assuming we're looking just at X86 performance, much of it comes down to price. If it's 10% faster in the same price range as we had in 2010, that is disappointing. They are only offering a single socket version. Other oems can go further, but Apple doesn't seem to be addressing that market. They also never offered Tesla co-processors, although those were arguably an even smaller niche. I can spot certain things I dislike, but I wouldn't form an entire opinion of a future machine based solely on that.
Well I have to agree that Intels incremental improvements in basic ALU operations has stalled a bit but let's not forget this CPU is running very slow and is still beating the old technology. Given a process shrink and other thermal management improvements this machine will be able to run much faster. Intel is expected to have a six core Haswell based chip later this year running at 130 watts and 3.6 GHz if I remember right.
The point is this machine could see a nice performance boost with a new generation of Xeons Gains in basic CPU performance aren't gone but it does require clock rate increase from here on out. This highlights another thing, people running software that is not heavily multithreaded may be better off with fewer cores. A six core machine will be very interesting when benchmarked with real apps. This assumes getting higher clock rates out of the six core implementation.
So while you are right that Intels performance increases from one generation to another has slowed considerably we still have a couple of generations of process shrinks ahead of us that should lead to much faster machines. The age of significant core improvements is gone so we are back to clock rate scaling and process shrinks.
The highest-end W9000 FirePros retail for $3k each. Two W5000s would be $800-900 for both worst case. I agree that Apple has quite rightly replaced a second Intel CPU with a much better value AMD OpenCL processor and it's up to developers to take advantage of it now.
I don't think the entry model needs to cost more. There's not much reason to force expensive GPUs on everyone, especially people who will want this as a server. The highest-end model will be very expensive but they can make options that span all the way from $2-2.5k to $15k.
Yeah, all it would take is for some CPU intensive software to transition to OpenCL and the benefits will be very clear to see. All software that deals with image processing should be moving in this direction as well as mathematics software - Mathematica has already added support:
http://www.wolfram.com/mathematica/new-in-8/cuda-and-opencl-support/
I really wish it wasn't always CUDA and OpenCL together. It's like the situation with DirectX and OpenGL. One of them is only compatible with a single vendor and the other is a standard. It's just the maturity of each API I suppose and what they are capable of but it really messes things up.
Quote:
Originally Posted by Rogifan
Wow a lot of people on MR freaking out....geez why don't we wait for some reviews and benchmarks once the machine has actually been released?
Agreed. This is pre-production hardware (with what is most likely a pre-production processor) running a beta version of the OS. There's no reason to take these results too seriously.
Quote:
Originally Posted by Conrail
Agreed. This is pre-production hardware (with what is most likely a pre-production processor) running a beta version of the OS. There's no reason to take these results too seriously.
Pre-production / engineering sample geekbench scores have leaked before. They've changed somewhat with live hardware, but not that much. I think it's important to note that many of these users wouldn't have purchased something at the cost of the top spec 12 core. I would wait and what this one is like. If the machine is unsuitable due to tco or performance, you can always look at alternatives. I don't think they intend to throw the cost to performance ratio further out of balance. That would be a really bad move.
Personally I'd like to see how a six core machine performs. I don't know what wattage range Apple has designed the machine for but a 130 watt processor ought to be rather impressive. This is especially the case if they can turbo up to 4GHz or more.
http://www.tomshardware.com/reviews/xeon-e3-benchmark-review,3258-6.html
Cinebench only goes up 6% due to the architecture. The biggest gain comes from core count increases.
Here's the result for dual 2.6GHz E5-2670 Sandy Bridge:
Per processor, that's 11.25. The core-count per CPU there is 8 and 2.6GHz, the E5-2697 v2 will be 12-core and 2.7GHz. So, expected Cinebench score for the new top-end Mac Pro would be:
1.06 x 11.25 x 12/8 x 2.7/2.6 = 18.5
The current 12-core gets 16 so it's an increase of about 15%. As I say though, potentially $1200 cheaper for that level of CPU performance because it's just a single CPU.
A Dell costs $7500 with dual E5-2687W with a score of 25 and the v2 will be 10-core at 3.4GHz vs 3.1GHz so 45% faster = score of 36.5.
If the old Mac Pro was updated with dual CPU, they'd have used the IB equivalent of E5-2665 x 2, which in SB is 8-core 2.4GHz. This would have hit the same $6200 price. This SB score is here:
It's about 19-20. Intel hasn't increased the clocks here much, if at all for IB so 6% improvement from the architecture with 10-cores vs 8, a $6200 Mac Pro would have scored 26 with dual CPUs. Not double the new Mac Pro.
This is similar to what they did with the 2012 refresh, which is bring the higher power in at lower price points rather than try making machines with the fastest CPUs at $7k+ price points because the volume of buyers isn't there. The volume buyers for those expensive CPUs are for blade servers.
Quote:
Originally Posted by wizard69
Personally I'd like to see how a six core machine performs. I don't know what wattage range Apple has designed the machine for but a 130 watt processor ought to be rather impressive. This is especially the case if they can turbo up to 4GHz or more.
It could be the base model depending on how they structure it. Intel is likely to maintain a quad core in their Xeon E and EP lineups for Ivy. I suspect it depends on how Apple is going to organize and price the line. The really big gap is around the 12 core level. That could be an $1800 chip. The typical upper mid range Xeon 1600 chip that Apple is likely to use for their 6 core dropped to $600 mid westmere and stayed with that through Sandy. I suspect it will still be around $600 retail. That is higher than Apple has used in base machine so far, but I suspect they have a lot of leverage in the base model. Just breaking down what went into it suggested that the margins per unit were extremely high. I'm also not reading too much into the assembled in the US thing. I don't think they would move that if it drove the cost way up. The initial thread had too many clueless comments that were more based on preconceived notions rather than case studies. (Edit: yeah that's somewhat preconceived too, not feeling like linking buzz word articles that discuss "insourcing", but I don't think Apple views the bulk of their potential market as being in the $5-10k realm).
Quote:
Originally Posted by Marvin
This benchmark actually might be pretty accurate. Here's Anandtech's comparison of E3 Sandy Bridge to Ivy Bridge:
http://www.tomshardware.com/reviews/xeon-e3-benchmark-review,3258-6.html
Cinebench only goes up 6% due to the architecture. The biggest gain comes from core count increases.
Same thing as I've stated without the nice collection of references. It should have been obvious to anyone. In both mainstream chips and EP versions, performance per clock cycle hasn't risen more than 15% even in favorable testing conditions. That would be why we disagreed regarding Moore's law a few months back (not digging up the thread). The increasing core counts have driven some pricing higher at the cpu level. As for gpus, I'm not not exactly sure what we'll see there. I haven't looked up the chips. I'm not as convinced as everyone else that it will be dual W9000s if AMD is sticking to their current pricing structure. Those came out mid last year. With extended cycles, pricing does sometimes drops very significantly on workstation variants. A good portion of the pricing is also due to extra driver development. There's a frequent debate over whether it's worth it and how most workstation cards are the same chips with different settings. Beyond guessing at the basic available chips, I'm unsure what Apple is buying. They're branded as firepros, but I think it's too early to read into a lot of this stuff. Assuming they stick to $2500 and up, they do need something better at the low end of that. There's no extra storage or memory capacity (allowing for lower density ram) to entice anyone, so they do need real differentiation.
Quote:
This is similar to what they did with the 2012 refresh, which is bring the higher power in at lower price points rather than try making machines with the fastest CPUs at $7k+ price points because the volume of buyers isn't there. The volume buyers for those expensive CPUs are for blade servers.
They've gone that route since at least 2011 with mainstream quad core options and greater emphasis on viable APUs.
Moore's Law is about transistor count so they aren't far off that as they've gone from 6-core chips to 12-core in 3 years. We knew they were a year behind in their Xeon line. I'd have preferred if they moved to Haswell but AMD isn't keeping up with them on the CPU side so there's less incentive.
The real problem, which couldn't have been obvious until you knew the clock speeds and power consumption is performance-per-watt. The desktop/laptop chips all have IGPs so the Xeon chips should be able to move up further. Previously the Mac Pro used 2x $1440 95W (190W total) CPUs, now it will use 1x$2057 130W CPU and be around 15% faster.
Say that you get Y units of performance from the 190W chips, you now get 1.15Y units from 130W so performance-per-watt has improved 70% in 3 years, which is really the important factor and where ARM is doing better than Intel. At least the price is a bit lower.
Intel is increasing the core count ok but they are using lower clock speeds. They are dynamic, which complicates things but if these had 3.4GHz base clock and 3.8GHz Turbo, rather than 2.7GHz with 3.3GHz Turbo, it would have been what I expected from the single CPU.
They might not be branded as W9000, these are custom parts. They just have the same spec. They are using the FirePro brand and they'll need some identifier to compare them though so I don't see the problem with them calling them W9000. I think we can be pretty sure they can't have the full $6k retail price.
We don't have all the facts, but there are a few things that can be deduced from what Apple has shown us about the Tube Mac Pro.
1. There must be 2 GPUs. Probability: High. When Apple decided to eliminate PCIe slots and use Thunderbolt, they looked at how many Thunderbolt ports to include, and decided 4 was too few, 6 sufficient. They probably didn't want to have some data-only Thunderbolt ports, so all are capable of driving a monitor. There is also an HDMI port, for a total of 7 display outputs. While the machine may not be able to drive all 7 simultaneously, by Apple design standards, any one should work.
The architecture would therefore be a single Xeon, 32 PCIe lanes to the two GPUs, 24 lanes to the Thunderbolt controllers (4 per TB connector), 6 displayport channels from the GPUs to the Thunderbolt controllers. The additional HDMI display channel could in theory come from the Xeon for some Xeon chips, but it's more likely that it's always connected to one of the GPUs, since it must be for the 12-core.
The allocation of display output bandwidth to DisplayPort output channels is done inside the GPUs. A FirePro GPU with several DisplayPort connectors can support high (i.e. 2560x1600) resolution on any connector, but not on all at once (personal experience). Apple states that the Tube can support up to 3 4K displays.
This design also limits which GPUs can be used, as the lower-end cards can't support 4 connections. It's possible, but unlikely, that Apple could have a low-end configuration with two different GPUs, one driving 3, the other 4 of the possible outputs.
2. By eliminating PCIe slots, and probably discontinuing the current Mac Pro (and assuming there is no new mid-level machine with PCIe slots), Apple will eliminate the market for 3rd-party video cards with MacOS drivers.
3. The GPUs are not upgradable. Probability: very high. The reason is simple: the GPUs are thermally mounted to the central heat sink, and Apple would never consider replacing thermal grease an end-user option. In fact, unless Apple has carefully designed it (which they certainly have done in the past), it won't be easy for a technician to replace a GPU card.
Also, since Thunderbolt 2.0 is limited to 4 PCIe lanes, an external high-performance GPU isn't an option. All modern GPUs use 16-lane PCIe slots.
4. The use of workstation-class FirePro GPUs will make it possible to run software like Solidworks on a machine as configured from the Apple store. There certainly will be users who want to use the Windows or Linux drivers.
5. The reason for speculation that the W9000 is used in the high-end configuration is that it's the only FirePro GPU with 6GB GDDR5 memory.
That and the FLOPs match up when you consider two of them.
If by 'high' you mean 100%.
Quote:
Originally Posted by davida
1. There must be 2 GPUs. Probability: High.
Seeing as how Apple states up front that the new Mac Pros use dual workstation-class GPUs as standard, the probability is beyond high…
Quote:
Originally Posted by davida
It's possible, but unlikely, that Apple could have a low-end configuration with two different GPUs, one driving 3, the other 4 of the possible outputs.
Not likely at all. See above. Apple is pushing this as what it is, a paradigm shift in the way we approach workstation computing. Why force more work on the CPU when a pair of (GP)GPUs can do some of that work faster & easier. Hence the push for developers to really start thinking about OpenCL at the latest WWDC.
Quote:
Originally Posted by davida
3. The GPUs are not upgradable. Probability: very high. The reason is simple: the GPUs are thermally mounted to the central heat sink, and Apple would never consider replacing thermal grease an end-user option. In fact, unless Apple has carefully designed it (which they certainly have done in the past), it won't be easy for a technician to replace a GPU card.
I do not think they are using thermal paste in this design. Look at the pics on the Apple website. Note the black metal brackets & knurled machine screws on the CPU & GPU cards. I believe they are using these items to clamp the chips directly to the Thermal Core, with a pad of thermally-conductive material between the chips & said Core. This also makes sense from an assembly standpoint; these three PCBs need to slot into the backplane AND align to the Thermal Core properly, thermal paste would be messy and could lead to unwanted damage to the PCBs while the worker is trying to assemble it all. Much easier to slot the PCBs into the backplane, drop that sub-assembly over the Thermal Core (which has the conductive pads already in place) and screw down the brackets/machine screws. Also much easier for a technician to replace a PCB if needed.
Quote:
Originally Posted by Marvin
Moore's Law is about transistor count so they aren't far off that as they've gone from 6-core chips to 12-core in 3 years. We knew they were a year behind in their Xeon line. I'd have preferred if they moved to Haswell but AMD isn't keeping up with them on the CPU side so there's less incentive.
The real problem, which couldn't have been obvious until you knew the clock speeds and power consumption is performance-per-watt. The desktop/laptop chips all have IGPs so the Xeon chips should be able to move up further. Previously the Mac Pro used 2x $1440 95W (190W total) CPUs, now it will use 1x$2057 130W CPU and be around 15% faster.
Say that you get Y units of performance from the 190W chips, you now get 1.15Y units from 130W so performance-per-watt has improved 70% in 3 years, which is really the important factor and where ARM is doing better than Intel. At least the price is a bit lower.
Intel is increasing the core count ok but they are using lower clock speeds. They are dynamic, which complicates things but if these had 3.4GHz base clock and 3.8GHz Turbo, rather than 2.7GHz with 3.3GHz Turbo, it would have been what I expected from the single CPU.
I was partly referring to some of your performance projections. In some tiers they have added more performance than others, but keeping up with the highest core counts meant some price escalation. As for Haswell, Ivy wasn't originally predicted to hit 12 cores. I think they'll do okay. AMD isn't keeping up with them, but opterons still represent a good value with software that scales in a near linear manner with core counts, as AMD has much cheaper . They certainly aren't terrible, and they have regularly shown up in some form in HPC solutions as indicated by the top500. Their current top supercomputer should make Wizard happy, as it employs Xeon Phis. Number 2 is using Tesla solutions and opterons. Anyway from what you've read, has ARM really been on top of power efficiency at the server level? They don't have anything really aimed at workstations, but they do make server chips.
Quote:
They might not be branded as W9000, these are custom parts. They just have the same spec. They are using the FirePro brand and they'll need some identifier to compare them though so I don't see the problem with them calling them W9000. I think we can be pretty sure they can't have the full $6k retail price.
I need to find the chip data. I suspect much of is the same as their top gaming card, with more ram, clocked lower, and custom drivers. A lot of the cost of workstation gpus is typically driven by higher margins and a need to recoup R&D expenses on a much smaller number of units. With Macs they had to develop the low level mac drivers either way. It's also possible that the W9000s will be falling in price mid cycle in the near future, given that they have been out for a while. My point was that the price may not be highly aligned with the current $3k+ market price when the mac pro debuts, probably Q4, so yeah I agree with you there. I'm not pleased by the lack of internal storage for numerous reasons. I am unlikely to change my mind about that. It's just that if the machine proves viable for me overall, I won't avoid it due to that one detail
Quote:
Originally Posted by Tallest Skil
If by 'high' you mean 100%.
I guess I should clarify this. When I say there must be 2 GPUs, I mean the Thunderbolt-instead-of-PCIe design decision dictates 2 GPUs. And regarding the Apple web page, it states that the Mac Pro has two workstation GPUs with up to 6GB memory, but I hope this isn't the only configuration. It would bump up the minimum system price quite a bit if two W9000s/6GB were the only option.
Another way of looking at it: It's an inflexible design.
Quote:
I do not think they are using thermal paste in this design. Look at the pics on the Apple website. Note the black metal brackets & knurled machine screws on the CPU & GPU cards. I believe they are using these items to clamp the chips directly to the Thermal Core, with a pad of thermally-conductive material between the chips & said Core. This also makes sense from an assembly standpoint; these three PCBs need to slot into the backplane AND align to the Thermal Core properly, thermal paste would be messy and could lead to unwanted damage to the PCBs while the worker is trying to assemble it all. Much easier to slot the PCBs into the backplane, drop that sub-assembly over the Thermal Core (which has the conductive pads already in place) and screw down the brackets/machine screws. Also much easier for a technician to replace a PCB if needed.
Thermal pads are obsolete, they wouldn't conduct enough heat away from the chips. Thermal paste can and should be a very thin layer. If you want to get an idea of how it's used, see these instructions:
http://www.arcticsilver.com/instructions.htm
I can assure you that thermal paste is being used, it's a requirement for the Xeon and the GPUs. It's possible they are using a heat spreader like a copper block between the chip and the aluminum core heat sink, but you would still use thermal paste.
The metal brackets with 4 screws are really calibrated springs compressing the GPU against the thermal core.
Quote:
Originally Posted by davida
Thermal pads are obsolete, they wouldn't conduct enough heat away from the chips. Thermal paste can and should be a very thin layer. If you want to get an idea of how it's used, see these instructions:
http://www.arcticsilver.com/instructions.htm
I can assure you that thermal paste is being used, it's a requirement for the Xeon and the GPUs. It's possible they are using a heat spreader like a copper block between the chip and the aluminum core heat sink, but you would still use thermal paste.
The metal brackets with 4 screws are really calibrated springs compressing the GPU against the thermal core.
I worked for a good number of years in a production facility dealing with last mile wireless networking equipment for businesses, this was when a single wifi radio was about the size of a "Wheel of Time' paperback. Meaning, the units we produced were 4U rack-mounted computers, with the self-same Arctic Silver thermal compound (paste) nestled between the CPU & heat sink. So yeah, I know how to use thermal compound…
In retrospect, I guess they could use some AS, but I like the idea of a modern thermal pad.
However they do it, any tech tearing one of these things down to replace one of the three main chip bearing PCBs is still gonna have to take it almost completely apart to do so, by my figuring that is about 60 pieces (PCB assemblies, chassis parts, fasteners, brackets, etc.) to futz with.
I just hope the folks doing assembly know how to have a light hand with the thermal compound, it is not an item where more is better…
Just based on my reading of past Apple specification pages, that 'up to 6 GB memory' seems to certainly indicate there will be GPUs on offer with less than 6GB. I am no expert on these matters but I would have thought the price difference in RAM on the GPUs would not make that much difference these days, so the wording seems odd if that is true. Is it your take that there will be a lower RAM configurations and would that make a large price difference? Also is such GPU RAM upgradeable or is it a one time decision?