Another is Intels new supper chip with built in Infiniband. This chip would be exactly what Apple would need for a cluster machine.
Inifiniband uses channel bonding for higher bandwidth like they could use with PCIe and it has over 100ns latency vs 8ns for Thunderbolt (across 7 devices no less).
I don't think the connection bandwidth of a single port would be insufficient, we're talking about 1.25GBytes/s. Most of the world's fastest supercomputers use GigE (1 Gbit/s) as an interconnect:
"Gigabit Ethernet is still the most-used internal system interconnect technology (223 systems, down from 230 systems), due to its widespread use at industrial customers, followed by InfiniBand technology with 213 systems, up from 208 systems. However, InfiniBand-based systems account for almost twice as much performance (28.7 Petaflop/s) than Gigabit Ethernet ones (14.2 Petaflop/s)."
That would be a great way to saturate the Thunderbolt bus and make it utterly unusable for anything that you would require a cluster in the first place.
Read/Write speeds would be back in the Commodore 64 era.
Consider every cluster ever built in the past. Were the connections between nodes faster or slower than Thunderbolt?
I'll give you a hint. The internodal connections within those previous clusters were slower--substantially slower--than Thunderbolt. The notion that the Thunderbolt bus will be saturated by a Thunderbolt-based cluster of Macs is ridiculous on its face.
Admittedly info is scarce but I don't think TB would do the job due to the lack of node support and other features. At least on installations of any size. Something as simple as sending a packet of info a hop or two beyond the neighbors would be interesting with TB. Unless of course you implement a simple network / hub arrangement but I'm not convinced that TB can work over a hub.
Quote:
Originally Posted by Marvin
Inifiniband uses channel bonding for higher bandwidth like they could use with PCIe and it has over 100ns latency vs 8ns for Thunderbolt (across 7 devices no less).
Seven devices would make a nice small cluster. With today's computing hardware you would not want to sneeze at that however how would one do a larger cluster.
Quote:
I don't think the connection bandwidth of a single port would be insufficient, we're talking about 1.25GBytes/s. Most of the world's fastest supercomputers use GigE (1 Gbit/s) as an interconnect:
Eventually bandwidth becomes an issue no matter what generation of cluster technology you use. This is one of the reasons you see so many different topologies implemented. For a small scale cluster though I think you are right in that the bandwidth is OK for today. I'm just not convinced though that TB has the support for the networking or communications for clusters of any size.
Quote:
"Gigabit Ethernet is still the most-used internal system interconnect technology (223 systems, down from 230 systems), due to its widespread use at industrial customers, followed by InfiniBand technology with 213 systems, up from 208 systems. However, InfiniBand-based systems account for almost twice as much performance (28.7 Petaflop/s) than Gigabit Ethernet ones (14.2 Petaflop/s)."
Ethernet is every where. Right now I'd have to believe it would be far easier to implement clustering over Ethernet than it would be over Thunderbolt. All the hardware is there already as is a strong software base.
Let's say though that Apples desire here is more limited in that they don't even want to support hubs or large systems. Maybe their idea of a cluster is less than six machines with direct TB to TB port interfacing. Such a small easy to administer cluster would be very useful, the question though is how well does TB support such use. This is the question in my mind as TB seems to be very I/O oriented with thin to non existent higher level protocols.
Consider every cluster ever built in the past. Were the connections between nodes faster or slower than Thunderbolt?
I'll give you a hint. The internodal connections within those previous clusters were slower--substantially slower--than Thunderbolt. The notion that the Thunderbolt bus will be saturated by a Thunderbolt-based cluster of Macs is ridiculous on its face.
The potential for interconnect saturation is greater today than in the past. It is very possible for each node these days to have eight processors and sixteen threads running. That is from a single socket node, back in the day you had two processors in two different sockets. Beyond that if you have storage or anything else running on the same port you end up with even more congestion.
On the other hand I suspect that it would be rather economical for Apple to implement multiple TB ports That is nothing seems to prevent a Mac Pro replacement from having 4 or more TB ports.
I think the point is...given Thunderbolt and X-Grid/Open CL and in the context of the customer Apple serves
Single user with several machines, family users of Mac, small business users, small studios, solo artists etc, Mac students...
If you could daisy chain '7' (is that the magic number for Thunderbolt then?) Macs together for extra compute oomph then that seems like the Apple demographic to me.
Instead of throwing out your old dinosaur Mac Pro...just put it in chain (wirelessly?), keep your old thunderbolt iMac and have a dual monitor iMac set up with 8 cores, 16 threads hyperthreading a 3D render's asss.
7 Macs is a pretty powerful solution to a small scale operation I would have thought. eg an iMac and even 4 Mac Mini nodes via thunderbolt X-Gridded?
I'm surprised nobody has tried this with Pov-Ray or something and a modest '7 or less' Mac cluster.
Interconnect is always going to be a potential bottleneck. Computers always have a bottleneck. HD? Memory? CPU? Bandwidth? IO? GPU? The computer press have taken turns blaming one thing or the other over the years. For me the current laggard is monitor resolutions. But what do I know.
While iPad/iPhone is making massive inroads into the Enterprise and Macs are popular amongst 'bring your own machine to work' types...the return of the 'X-Serve' maybe premature for now. Maybe localised clusters of 'working group' Macs could serve that need rather than a ward drobe rack of X-Serves.
If you could daisy chain '7' (is that the magic number for Thunderbolt then?)
The 7 limit might not apply to the grid. That's for dummy slaves. Each active computer has an active controller so it should allow unlimited chaining. They can also run an arbitrary protocol over it like TCP/IP so it would behave much the same way as 10Gbps ethernet.
Quote:
Originally Posted by Lemon Bon Bon.
I'm surprised nobody has tried this with Pov-Ray or something and a modest '7 or less' Mac cluster.
They'd need to write a PCI driver/kernel extension. That takes a special type of developer. I'm sure someone will try it eventually though. They're already building external GPUs so this sort of thing can't be too far off.
Uh, nope. Which is exactly why the iMac operates AS a display. You want to upgrade, do it. Use the iMac as a display with the hardware turned off. They've been able to do that for years.
That's a workaround, and not a terribly flexible one, either. It would make more sense for Apple to try to sell two Thunderbolt displays to those who need them. Some users want two like displays, and some prefer one large one and one smaller one for pallets, so using an old iMac as a display doesn't work. And what about those who need to run a wide color gamut display but do not need a Xeon workstation? Should they just "buy from someone else" as you put it? That's a winning business model!
Quote:
And for almost 15 years, Apple's marketshare has increased and profits skyrocketed. Obviously the hole doesn't even exist.
Flogging this one again, eh? I'll file this is in the "Apple is profitable so their hardware and business model are perfect" file.
Quote:
Odd that you'd find Apple fans on an Apple fan forum, huh?
So buy from someone else.
If I wasn't an Apple fan I wouldn't own a Mac Pro, two Minis, and two Power Mac G4s, with a MacBook on the way as soon as the news ones go on sale.
Interestingly, I bought both my Power Macs new from Apple. I bought my Mac Pro used and upgraded the processor myself. Why? Because I'll pay a little extra for an Apple tower, but not crazy stupid extra. Apple lost a sale to me because of their inability or unwillingness to offer a desktop solution that fits my needs.
In any event, I'd venture to guess that Apple doesn't sell many desktops because of their cool designs. People buy Mac desktops because Mac OS X is a superior OS. I'd also hazard to guess that what bugs people about "PCs" is Windows, not the hardware. Some of the hardware one the other side is brilliant, far better than Apple's hardware IMO, and not just because it's cheaper. But I must say that I love the Mac Pro's design, both functional and cosmetic, that tower is is work of art. It could use a little tweaking so it can use SSDs without modifications, but that's nitpicking. In fact I'm now modding my Mac Pro's optical bay for SSDs based on a design I saw on eBay; the guy gutted an optical drive, riveted a couple aluminum rails on it, and it looks sweet! I can't wait to get a couple of SSDs to try it out.
We'll hand it to HP -- "power without the tower" is pretty fab. And so is that drop-dead gorgeous 27-inch IPS panel dominating the front of its Z1 workstation. Designed to handle stresses normally thrown exclusively at floor-sitting wind tunnels, the Z1 offers up Intel Core i3 or Xeon server-class CPUs, NVIDIA Quadro graphics, USB 3.0 sockets, a slot-loading Blu-ray writer, SSD / RAID options, support for over a billion colors and a seductive starting tag of just $1,899. Intrigued? Hit the source link to get your order in.
I'm willing to say Mac Pros replacement will have more than two TB ports. Almost seems like a given.
Yeah, the Mac Pro has 2x x16 + 2x x4 = 40 PCI lanes. Say they run an internal GPU over 16 lanes, that leaves 24 lanes with 4 per port = 6 Thunderbolt ports. Some GPUs ship with 6 Mini-DP connectors so it seems like a good number of display outputs to have. A single GPU likely couldn't run more than 6 displays on its own anyway.
<p> I can't see Apple doing a more accessible iMac. (They once made a song and dance about it's access...) Now it seems 'lockdown' to help the churn.</p><p> </p>
I'm not sure what makes you say that. The new Mini was an improvement over the old, it is still far from perfect but no where as near as bad as the old. Same thing for Apples Laptops, servicing the new MBPs is far easier today than say with a 2007 model.
As to tablets and the AIRs it is the nature of the beast.
<p> Mind you, apart from the HD failure...I had little reason to get into one other than that.</p><p> </p><p> Lemon Bon Bon. :)</p>
Is that not reason enough? Seriously why make a machine where a high service rate device is hard to get to?
I will say one thing here for people to contemplate, iMacs very design keeps it out of many businesses.
Comments
Another is Intels new supper chip with built in Infiniband. This chip would be exactly what Apple would need for a cluster machine.
Inifiniband uses channel bonding for higher bandwidth like they could use with PCIe and it has over 100ns latency vs 8ns for Thunderbolt (across 7 devices no less).
I don't think the connection bandwidth of a single port would be insufficient, we're talking about 1.25GBytes/s. Most of the world's fastest supercomputers use GigE (1 Gbit/s) as an interconnect:
"Gigabit Ethernet is still the most-used internal system interconnect technology (223 systems, down from 230 systems), due to its widespread use at industrial customers, followed by InfiniBand technology with 213 systems, up from 208 systems. However, InfiniBand-based systems account for almost twice as much performance (28.7 Petaflop/s) than Gigabit Ethernet ones (14.2 Petaflop/s)."
http://www.top500.org/lists/2011/11/press-release
That would be a great way to saturate the Thunderbolt bus and make it utterly unusable for anything that you would require a cluster in the first place.
Read/Write speeds would be back in the Commodore 64 era.
Consider every cluster ever built in the past. Were the connections between nodes faster or slower than Thunderbolt?
I'll give you a hint. The internodal connections within those previous clusters were slower--substantially slower--than Thunderbolt. The notion that the Thunderbolt bus will be saturated by a Thunderbolt-based cluster of Macs is ridiculous on its face.
Inifiniband uses channel bonding for higher bandwidth like they could use with PCIe and it has over 100ns latency vs 8ns for Thunderbolt (across 7 devices no less).
Seven devices would make a nice small cluster. With today's computing hardware you would not want to sneeze at that however how would one do a larger cluster.
I don't think the connection bandwidth of a single port would be insufficient, we're talking about 1.25GBytes/s. Most of the world's fastest supercomputers use GigE (1 Gbit/s) as an interconnect:
Eventually bandwidth becomes an issue no matter what generation of cluster technology you use. This is one of the reasons you see so many different topologies implemented. For a small scale cluster though I think you are right in that the bandwidth is OK for today. I'm just not convinced though that TB has the support for the networking or communications for clusters of any size.
"Gigabit Ethernet is still the most-used internal system interconnect technology (223 systems, down from 230 systems), due to its widespread use at industrial customers, followed by InfiniBand technology with 213 systems, up from 208 systems. However, InfiniBand-based systems account for almost twice as much performance (28.7 Petaflop/s) than Gigabit Ethernet ones (14.2 Petaflop/s)."
http://www.top500.org/lists/2011/11/press-release
Ethernet is every where. Right now I'd have to believe it would be far easier to implement clustering over Ethernet than it would be over Thunderbolt. All the hardware is there already as is a strong software base.
Let's say though that Apples desire here is more limited in that they don't even want to support hubs or large systems. Maybe their idea of a cluster is less than six machines with direct TB to TB port interfacing. Such a small easy to administer cluster would be very useful, the question though is how well does TB support such use. This is the question in my mind as TB seems to be very I/O oriented with thin to non existent higher level protocols.
Consider every cluster ever built in the past. Were the connections between nodes faster or slower than Thunderbolt?
I'll give you a hint. The internodal connections within those previous clusters were slower--substantially slower--than Thunderbolt. The notion that the Thunderbolt bus will be saturated by a Thunderbolt-based cluster of Macs is ridiculous on its face.
The potential for interconnect saturation is greater today than in the past. It is very possible for each node these days to have eight processors and sixteen threads running. That is from a single socket node, back in the day you had two processors in two different sockets. Beyond that if you have storage or anything else running on the same port you end up with even more congestion.
On the other hand I suspect that it would be rather economical for Apple to implement multiple TB ports That is nothing seems to prevent a Mac Pro replacement from having 4 or more TB ports.
Single user with several machines, family users of Mac, small business users, small studios, solo artists etc, Mac students...
If you could daisy chain '7' (is that the magic number for Thunderbolt then?) Macs together for extra compute oomph then that seems like the Apple demographic to me.
Instead of throwing out your old dinosaur Mac Pro...just put it in chain (wirelessly?), keep your old thunderbolt iMac and have a dual monitor iMac set up with 8 cores, 16 threads hyperthreading a 3D render's asss.
7 Macs is a pretty powerful solution to a small scale operation I would have thought. eg an iMac and even 4 Mac Mini nodes via thunderbolt X-Gridded?
I'm surprised nobody has tried this with Pov-Ray or something and a modest '7 or less' Mac cluster.
Interconnect is always going to be a potential bottleneck. Computers always have a bottleneck. HD? Memory? CPU? Bandwidth? IO? GPU? The computer press have taken turns blaming one thing or the other over the years. For me the current laggard is monitor resolutions. But what do I know.
While iPad/iPhone is making massive inroads into the Enterprise and Macs are popular amongst 'bring your own machine to work' types...the return of the 'X-Serve' maybe premature for now. Maybe localised clusters of 'working group' Macs could serve that need rather than a ward drobe rack of X-Serves.
Skynet on the rise, eh? *(sticks tongue out.)
Lemon Bon Bon.
If you could daisy chain '7' (is that the magic number for Thunderbolt then?)
The 7 limit might not apply to the grid. That's for dummy slaves. Each active computer has an active controller so it should allow unlimited chaining. They can also run an arbitrary protocol over it like TCP/IP so it would behave much the same way as 10Gbps ethernet.
I'm surprised nobody has tried this with Pov-Ray or something and a modest '7 or less' Mac cluster.
They'd need to write a PCI driver/kernel extension. That takes a special type of developer. I'm sure someone will try it eventually though. They're already building external GPUs so this sort of thing can't be too far off.
?Thunderbolt? ?is just generally flaky?
In what way?
The iMac has two Thunderbolt ports. I imagine all Macs will eventually have at least two and many will have more.
I don't expect the next Mac Pro to have fewer than four, and if Apple is serious about it, they'll have six or eight.
Six or eight thunderbolt ports? Hell, the mac pro only has five USB ports.
You will never see an Apple computer with more than two thunderbolt ports.
You will never see an Apple computer with more than two thunderbolt ports.
Honestly, I doubt that.
Uh, nope. Which is exactly why the iMac operates AS a display. You want to upgrade, do it. Use the iMac as a display with the hardware turned off. They've been able to do that for years.
That's a workaround, and not a terribly flexible one, either. It would make more sense for Apple to try to sell two Thunderbolt displays to those who need them. Some users want two like displays, and some prefer one large one and one smaller one for pallets, so using an old iMac as a display doesn't work. And what about those who need to run a wide color gamut display but do not need a Xeon workstation? Should they just "buy from someone else" as you put it? That's a winning business model!
And for almost 15 years, Apple's marketshare has increased and profits skyrocketed. Obviously the hole doesn't even exist.
Flogging this one again, eh? I'll file this is in the "Apple is profitable so their hardware and business model are perfect" file.
Odd that you'd find Apple fans on an Apple fan forum, huh?
So buy from someone else.
If I wasn't an Apple fan I wouldn't own a Mac Pro, two Minis, and two Power Mac G4s, with a MacBook on the way as soon as the news ones go on sale.
Interestingly, I bought both my Power Macs new from Apple. I bought my Mac Pro used and upgraded the processor myself. Why? Because I'll pay a little extra for an Apple tower, but not crazy stupid extra. Apple lost a sale to me because of their inability or unwillingness to offer a desktop solution that fits my needs.
In any event, I'd venture to guess that Apple doesn't sell many desktops because of their cool designs. People buy Mac desktops because Mac OS X is a superior OS. I'd also hazard to guess that what bugs people about "PCs" is Windows, not the hardware. Some of the hardware one the other side is brilliant, far better than Apple's hardware IMO, and not just because it's cheaper. But I must say that I love the Mac Pro's design, both functional and cosmetic, that tower is is work of art. It could use a little tweaking so it can use SSDs without modifications, but that's nitpicking. In fact I'm now modding my Mac Pro's optical bay for SSDs based on a design I saw on eBay; the guy gutted an optical drive, riveted a couple aluminum rails on it, and it looks sweet! I can't wait to get a couple of SSDs to try it out.
http://www.engadget.com/2012/04/17/h...ation-pricing/
iMac Pro, please
http://www.engadget.com/2012/04/17/h...ation-pricing/
I think HP and the Z1 deserve some credit for this.
Return serve, Apple?
Lemon Bon Bon.
We'll hand it to HP -- "power without the tower" is pretty fab. And so is that drop-dead gorgeous 27-inch IPS panel dominating the front of its Z1 workstation. Designed to handle stresses normally thrown exclusively at floor-sitting wind tunnels, the Z1 offers up Intel Core i3 or Xeon server-class CPUs, NVIDIA Quadro graphics, USB 3.0 sockets, a slot-loading Blu-ray writer, SSD / RAID options, support for over a billion colors and a seductive starting tag of just $1,899. Intrigued? Hit the source link to get your order in.
Lemon Bon Bon.
Honestly, I doubt that.
I'm willing to say Mac Pros replacement will have more than two TB ports. Almost seems like a given.
I'm willing to say Mac Pros replacement will have more than two TB ports. Almost seems like a given.
Yeah, the Mac Pro has 2x x16 + 2x x4 = 40 PCI lanes. Say they run an internal GPU over 16 lanes, that leaves 24 lanes with 4 per port = 6 Thunderbolt ports. Some GPUs ship with 6 Mini-DP connectors so it seems like a good number of display outputs to have. A single GPU likely couldn't run more than 6 displays on its own anyway.
iMac Pro, please
http://www.engadget.com/2012/04/17/h...ation-pricing/
I loved the way it can be open for user access to internal parts but I'm not seeing Apple doing something like this
I can't see Apple doing a more accessible iMac. (They once made a song and dance about it's access...) Now it seems 'lockdown' to help the churn.
Mind you, apart from the HD failure...I had little reason to get into one other than that.
Lemon Bon Bon.
As to tablets and the AIRs it is the nature of the beast. Is that not reason enough? Seriously why make a machine where a high service rate device is hard to get to?
I will say one thing here for people to contemplate, iMacs very design keeps it out of many businesses.