I'd still want them in RAID 10 or 5 because the point of RAID isn't data backup, it's system resilience. Doesn't matter if you needed the data on them or not - if one goes down, so does your system until it's sorted.
Correct. I'm having to bin a pair of FirePro W9Ks simply because Autodesk and Pix4D have gone all in with CUDA. Shame, the cards are still very competitive elsewhere, but Autodesk mesh manipulation is phenomenal and Pix4D is simply the gold standard for photogrammetry.
Presumably he means Robot Structural Analysis Professional, which is a structural engineering focused software (it is used by architects, nothing to do with robots), rather than a rendering focused one, like most of Autodesk's programs.
ReMake and ReCap. They work brilliantly with comparatively large data sets and are surprisingly simple to learn and use once you get the hang of the Autodesk implementation of mesh interpolation.
And as for visuals? Stunning. It may not seem important, but when an Engineering department is begging for funding, a little "ooh" and "ah" goes a long, long way with folks who don't necessarily understand what is being done but they like pretty pictures. :-)
AMD is really stuck between a rock and a hard place for the GPU market. Their architecture is way more competitive in compute than gaming but most of the compute stuff is CUDA so they're buying NVIDIA no matter what.
I say "link" meaning link to us this magical system that doesn't exist. Just suppose for the moment the MI25 was released and available for purchase in shipping systems (it isn't). One MI25 has 24.6 TFLOPS of FP16. To get 3 PFLOPS you'd need 122 of them. Put 8 of them in a node. So you need 16 nodes. Now if you can find us a "full pre-built" cluster of 16 nodes containing at least 122 MI25s and all the processors, power supplies, cooling, and interconnect necessary for less than $150K I'm gonna be impressed. Such a thing would be in the several millions of dollars, once the MI25 becomes available from server makers.
You mean the AMD Instinct MI25 that isn't even available for purchase? I mean, since there isn't a listed price, I suppose you could pretend the MI25 cost less than 1% of a DGX-1 and get a 3PFlops system.
Volta is really shaping up to be interesting. I've got a lower end Pascal GPU to get me through the year, but I'm really going to step it up when GV104 lands.
I think a distinction of some kind should be made regarding the Server's FP16 performance. DGX-1V has only 240 TF of general purpose FP16 compute, right? It's deep learning workloads only where that shoots up 4x to 960. It seems like there's be a useful distinction between deep learning FLOPS and general purpose ones, unless I'm mistaken.
You are absolutely correct, and the chart has been updated to reflect that. I did note specialized tensor performance separately when PCIe V100 was announced, but the NVIDIA DGX product pages and datasheet charts simply use 'FP16 TFLOPS' in comparing Volta-based DGX vs Pascal. While DGX systems are focused on deep learning, at a glance that label makes it appear that 960 TFLOPS is general FP16 performance, and so that deep learning compute clarification should definitely be made clear.
Yeah, and from their own benchmarks, that 4x performance boost translates to only a less than 2x boost in actual deep learning workloads. In one benchmark, NVIDIA claims the DGX-1V trains about 2.4 times faster than the DGX-1P. Based on core count and clock differences, using standard FP16 instructions the expectation would be about 1.5 times faster. So, assuming there isn't any reduction in bandwidth bottlenecks with the V100, the Tensor cores seem to be providing about a 1.6 times speed up over standard FP16 instructions in practical use.
What does this mean for the eventual release of consumer-focussed Volta-based GeForce products? Can we expect GeForce 11 cards before Xmas 2017, or is it more likely to be some time in 2018?
Doubtful. V100 isn't a consumer product, even if V102 is probably largely similar. I suspect they'll sell all of these that they can make for the next few months.
The DGX-1 (Pascal) was released 2 months before consumer GPU but looking at the lack of rumours (testings boards...), I'm not really expecting anything before CES 2018. Moreover, the GTX 1180/2080 will ship with the GV104, which is a different die than the V100/V102, so the timing between them doesn't have to be correlated.
Memory contract prices and available volume makes it unlikely for a release (at least any with meaningful supply) in 2017. 2018Q1 is the likely candidate for consumer releases barring any issues.
During NV's latest earnings call, JHH stated: "Volta for gaming, we haven't announced anything [...] But for the holiday season for the foreseeable future, I think Pascal is just unbeatable."[1][2] In his words, Volta-based GeForce products are not expected in 2017. Like others mentioned, V100 timeline doesn't necessarily mean anything much with respect to consumer Volta products.
They are not rushing with gaming Volta boards because Vega performed much worse than expected. If Vega 64 had beat 1080Ti gaming Volta samples would probably have already be sent to reviewers.
Look at the turnaround from the P100 to the V100. And the V100 is a more complex and much bigger chip than the consumer GPUs will be. NVIDIA could have brought consumer Volta in 2017 if they wanted to.
And they possibly could have a couple revisions of the chip floating around at various stages in the pipeline and when they decide when they want to bring it to market they choose an appropriate revision to mass produce.
In the light of the Vega release, Nvidia has postponed all consumer Volta products to 2018. And even then, I expect we'll see high-end around March and mid-end God knows when. Could be anything between June and holidays season.
When was it ever postponed? Who says consumer volta (presumably GV104) was ever intended for 2017?
Nvidia has never (in recent history at least, to my knowledge) launched two flagship parts that "replace" each other in the same year. The Titans get updated about once a year, and a G"X" 104/102 generally comes out once a year. The ONLY exception was the GTX 780 and 780 Ti, which both came out in 2013, but that doesn't really count since it was the same chip (GK110), the GTX 780 was just very disabled. I don't think we've ever seen a flagship level x80 or x80 Ti replaced in the same year it launched, so it seems kind of strange to expect Nvidia to do it now with a brand new architecture.
Where're the CPUs on that board? The 8 copper heat sinks are presumably the 8x V100 chips. The only other sinks I see are the 4 at the front; and those look way too small to cool a CPU even aside from the spec table saying there're only 2 CPUs not 4 in the system.
Like the Pascal-based DGX-1s, the CPUs are on its own board, connected by IB EDR.[1][2] Photos of other components were not shown in the CCDS PR photos.
Is there a disconnect in the general public understanding of this tech?
For example....
Quote "At CCDS, these AI supercomputers will continue to be used in training deep neural networks for the purpose of evaluating medical images and scans"
O.K. but once the data is in hand, will doctors be able to access the power of all that training with general purpose software on any standard PC?
I was very impressed with the power of A.I. that can be had for free recently when I downloaded a trial copy of iZotope RX6
An offline dualcore 35 Watt Sandy Bridge can do audio processing now that was impossible just a year ago (Go watch some Youtube Videos)
Once the training data is in hand, end users can access that data with properly coded software on any PC without any need for these new systems
The initial cost for these systems (as I understand it) is to get the training data needed
But once the data is available, a general purpose PC can access that training data in the middle of nowhere and without an Internet connection as long as they have the software capable of properly using that data
Is that correct?
In other words, will the true power of all this training data be available to the masses with properly coded software to make a better World, or will greedy Corporations hoard the data so they can play God and profit at the expense of everyone else ?
You'll have to ask them what they'll do. I assume they will sell it on the open market just like any other good or service. They are spending money to develop a system that they think people will find valuable. That is, it will be cheaper and/or higher quality than current methods. So I am guessing that any interested medical institution would have to pay for the use of the trained neural network just like they'd have to pay for the use of a trained doctor or medical billing software. I'm not sure why you think this labor and enterprise would be "greedy" for looking for a return on their investment.
Using the trained networks will definitely require a lot less number crunching power - depending on how many images and at what resolution are needed. This will probably still require modern GPUs, but not 8xGV100.
Obligatory Ethereum mining reference - that's about the equivalent of 140 GTX1060's in memory bandwidth. But at a cost of about 4x buying that many 1060's :)
......with orders of magnitude of computational power in hand, the stalwart whiteboard keeps watch, proud in knowing it still is needed to keep the world on track.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
48 Comments
Back to Article
Wardrive86 - Thursday, September 7, 2017 - link
Almost a PFlop of FP16 for $150 k.Despoiler - Thursday, September 7, 2017 - link
That's pretty pathetic. Full pre-built systems with AMD Instinct hardware will give you 3 PFlops for the same price.ddriver - Thursday, September 7, 2017 - link
But it's AMD, so you won't get to claim superiority on grounds of brand premium.Evidently, all people care nowadays is 16 bit computing. Meh...
4 drives in raid 0, talk about major data loss when one of those gives.
Space Jam - Thursday, September 7, 2017 - link
Those drives are more so a cache than for permanent storage in any proper system.Spunjji - Wednesday, September 13, 2017 - link
I'd still want them in RAID 10 or 5 because the point of RAID isn't data backup, it's system resilience. Doesn't matter if you needed the data on them or not - if one goes down, so does your system until it's sorted.mldude - Friday, September 29, 2017 - link
you can reconfigure them in raid10 or raid5.verl - Thursday, September 7, 2017 - link
OTOH, most packages have CUDA support, but not OpenCL.johnnycanadian - Thursday, September 7, 2017 - link
Correct. I'm having to bin a pair of FirePro W9Ks simply because Autodesk and Pix4D have gone all in with CUDA. Shame, the cards are still very competitive elsewhere, but Autodesk mesh manipulation is phenomenal and Pix4D is simply the gold standard for photogrammetry.etobler - Friday, September 8, 2017 - link
"Autodesk mesh manipulation", curious what Autodesk app you're referring to. Please do let me know.Santoval - Friday, September 8, 2017 - link
Presumably he means Robot Structural Analysis Professional, which is a structural engineering focused software (it is used by architects, nothing to do with robots), rather than a rendering focused one, like most of Autodesk's programs.etobler - Friday, September 8, 2017 - link
Thank you, but as far as I know, Robot Structural Analysis Professional doesn't have mesh manipulation tools.johnnycanadian - Friday, September 8, 2017 - link
ReMake and ReCap. They work brilliantly with comparatively large data sets and are surprisingly simple to learn and use once you get the hang of the Autodesk implementation of mesh interpolation.And as for visuals? Stunning. It may not seem important, but when an Engineering department is begging for funding, a little "ooh" and "ah" goes a long, long way with folks who don't necessarily understand what is being done but they like pretty pictures. :-)
notashill - Thursday, September 7, 2017 - link
AMD is really stuck between a rock and a hard place for the GPU market. Their architecture is way more competitive in compute than gaming but most of the compute stuff is CUDA so they're buying NVIDIA no matter what.Yojimbo - Thursday, September 7, 2017 - link
"That's pretty pathetic. Full pre-built systems with AMD Instinct hardware will give you 3 PFlops for the same price. "Link?
Yojimbo - Thursday, September 7, 2017 - link
I say "link" meaning link to us this magical system that doesn't exist. Just suppose for the moment the MI25 was released and available for purchase in shipping systems (it isn't). One MI25 has 24.6 TFLOPS of FP16. To get 3 PFLOPS you'd need 122 of them. Put 8 of them in a node. So you need 16 nodes. Now if you can find us a "full pre-built" cluster of 16 nodes containing at least 122 MI25s and all the processors, power supplies, cooling, and interconnect necessary for less than $150K I'm gonna be impressed. Such a thing would be in the several millions of dollars, once the MI25 becomes available from server makers.jordanclock - Thursday, September 7, 2017 - link
You mean the AMD Instinct MI25 that isn't even available for purchase? I mean, since there isn't a listed price, I suppose you could pretend the MI25 cost less than 1% of a DGX-1 and get a 3PFlops system.Shadow7037932 - Monday, September 11, 2017 - link
Oh look, a product that dosen't even existDrumsticks - Thursday, September 7, 2017 - link
Volta is really shaping up to be interesting. I've got a lower end Pascal GPU to get me through the year, but I'm really going to step it up when GV104 lands.I think a distinction of some kind should be made regarding the Server's FP16 performance. DGX-1V has only 240 TF of general purpose FP16 compute, right? It's deep learning workloads only where that shoots up 4x to 960. It seems like there's be a useful distinction between deep learning FLOPS and general purpose ones, unless I'm mistaken.
Nate Oh - Thursday, September 7, 2017 - link
You are absolutely correct, and the chart has been updated to reflect that. I did note specialized tensor performance separately when PCIe V100 was announced, but the NVIDIA DGX product pages and datasheet charts simply use 'FP16 TFLOPS' in comparing Volta-based DGX vs Pascal. While DGX systems are focused on deep learning, at a glance that label makes it appear that 960 TFLOPS is general FP16 performance, and so that deep learning compute clarification should definitely be made clear.Nate Oh - Thursday, September 7, 2017 - link
Oops, I forgot to say thanks, but thanks Drumsticks :)Drumsticks - Thursday, September 7, 2017 - link
Of course. Will be looking forward to any future coverage y'all have for us on Volta :)Yojimbo - Thursday, September 7, 2017 - link
Yeah, and from their own benchmarks, that 4x performance boost translates to only a less than 2x boost in actual deep learning workloads. In one benchmark, NVIDIA claims the DGX-1V trains about 2.4 times faster than the DGX-1P. Based on core count and clock differences, using standard FP16 instructions the expectation would be about 1.5 times faster. So, assuming there isn't any reduction in bandwidth bottlenecks with the V100, the Tensor cores seem to be providing about a 1.6 times speed up over standard FP16 instructions in practical use.colonelclaw - Thursday, September 7, 2017 - link
What does this mean for the eventual release of consumer-focussed Volta-based GeForce products? Can we expect GeForce 11 cards before Xmas 2017, or is it more likely to be some time in 2018?A5 - Thursday, September 7, 2017 - link
Doubtful. V100 isn't a consumer product, even if V102 is probably largely similar. I suspect they'll sell all of these that they can make for the next few months.Taiki - Thursday, September 7, 2017 - link
The DGX-1 (Pascal) was released 2 months before consumer GPU but looking at the lack of rumours (testings boards...), I'm not really expecting anything before CES 2018.Moreover, the GTX 1180/2080 will ship with the GV104, which is a different die than the V100/V102, so the timing between them doesn't have to be correlated.
limitedaccess - Thursday, September 7, 2017 - link
Memory contract prices and available volume makes it unlikely for a release (at least any with meaningful supply) in 2017. 2018Q1 is the likely candidate for consumer releases barring any issues.Nate Oh - Thursday, September 7, 2017 - link
During NV's latest earnings call, JHH stated: "Volta for gaming, we haven't announced anything [...] But for the holiday season for the foreseeable future, I think Pascal is just unbeatable."[1][2] In his words, Volta-based GeForce products are not expected in 2017. Like others mentioned, V100 timeline doesn't necessarily mean anything much with respect to consumer Volta products.[1] https://www.pcgamer.com/nvidias-next-gen-volta-gam...
[2] https://seekingalpha.com/article/4097782-nvidia-nv...
Santoval - Friday, September 8, 2017 - link
They are not rushing with gaming Volta boards because Vega performed much worse than expected. If Vega 64 had beat 1080Ti gaming Volta samples would probably have already be sent to reviewers.jordanclock - Friday, September 8, 2017 - link
I think you're really underestimating the turnaround on next GPU generations.Yojimbo - Friday, September 8, 2017 - link
Look at the turnaround from the P100 to the V100. And the V100 is a more complex and much bigger chip than the consumer GPUs will be. NVIDIA could have brought consumer Volta in 2017 if they wanted to.And they possibly could have a couple revisions of the chip floating around at various stages in the pipeline and when they decide when they want to bring it to market they choose an appropriate revision to mass produce.
bug77 - Thursday, September 7, 2017 - link
In the light of the Vega release, Nvidia has postponed all consumer Volta products to 2018. And even then, I expect we'll see high-end around March and mid-end God knows when. Could be anything between June and holidays season.Drumsticks - Thursday, September 7, 2017 - link
When was it ever postponed? Who says consumer volta (presumably GV104) was ever intended for 2017?Nvidia has never (in recent history at least, to my knowledge) launched two flagship parts that "replace" each other in the same year. The Titans get updated about once a year, and a G"X" 104/102 generally comes out once a year. The ONLY exception was the GTX 780 and 780 Ti, which both came out in 2013, but that doesn't really count since it was the same chip (GK110), the GTX 780 was just very disabled. I don't think we've ever seen a flagship level x80 or x80 Ti replaced in the same year it launched, so it seems kind of strange to expect Nvidia to do it now with a brand new architecture.
vladx - Thursday, September 7, 2017 - link
Exactly, people claiming that stuff above about Volta are mostly AMD fans looking to direct attention from AMD's failures to Nvidia.Hurr Durr - Thursday, September 7, 2017 - link
Weird strategy, considering how much nV is steamrolling them.Zingam - Thursday, September 14, 2017 - link
After years of AMD I got NVIDIA once again and I was hailed by critical driver bugs immediately. Whose fans and whose failures?DanNeely - Thursday, September 7, 2017 - link
Where're the CPUs on that board? The 8 copper heat sinks are presumably the 8x V100 chips. The only other sinks I see are the 4 at the front; and those look way too small to cool a CPU even aside from the spec table saying there're only 2 CPUs not 4 in the system.Nate Oh - Thursday, September 7, 2017 - link
Like the Pascal-based DGX-1s, the CPUs are on its own board, connected by IB EDR.[1][2] Photos of other components were not shown in the CCDS PR photos.[1] http://www.anandtech.com/show/10229/nvidia-announc...
[2] http://images.anandtech.com/doci/10229/DGX1Parts.j...
Nate Oh - Thursday, September 7, 2017 - link
Whoops, I didn't meant to say IB EDR connects the CPU board to the GPU board.Arbie - Friday, September 8, 2017 - link
Bet you wish this forum had "edit" buttons...Bullwinkle J Moose - Thursday, September 7, 2017 - link
Is there a disconnect in the general public understanding of this tech?For example....
Quote
"At CCDS, these AI supercomputers will continue to be used in training deep neural networks for the purpose of evaluating medical images and scans"
O.K. but once the data is in hand, will doctors be able to access the power of all that training with general purpose software on any standard PC?
I was very impressed with the power of A.I. that can be had for free recently when I downloaded a trial copy of iZotope RX6
An offline dualcore 35 Watt Sandy Bridge can do audio processing now that was impossible just a year ago (Go watch some Youtube Videos)
Once the training data is in hand, end users can access that data with properly coded software on any PC without any need for these new systems
The initial cost for these systems (as I understand it) is to get the training data needed
But once the data is available, a general purpose PC can access that training data in the middle of nowhere and without an Internet connection as long as they have the software capable of properly using that data
Is that correct?
In other words, will the true power of all this training data be available to the masses with properly coded software to make a better World, or will greedy Corporations hoard the data so they can play God and profit at the expense of everyone else ?
Yojimbo - Thursday, September 7, 2017 - link
You'll have to ask them what they'll do. I assume they will sell it on the open market just like any other good or service. They are spending money to develop a system that they think people will find valuable. That is, it will be cheaper and/or higher quality than current methods. So I am guessing that any interested medical institution would have to pay for the use of the trained neural network just like they'd have to pay for the use of a trained doctor or medical billing software. I'm not sure why you think this labor and enterprise would be "greedy" for looking for a return on their investment.MrSpadge - Friday, September 8, 2017 - link
Using the trained networks will definitely require a lot less number crunching power - depending on how many images and at what resolution are needed. This will probably still require modern GPUs, but not 8xGV100.gfkBill - Friday, September 8, 2017 - link
Obligatory Ethereum mining reference - that's about the equivalent of 140 GTX1060's in memory bandwidth. But at a cost of about 4x buying that many 1060's :)Ironchef3500 - Friday, September 8, 2017 - link
holy shitsparkuss - Friday, September 8, 2017 - link
I just enjoy the picture composition,......with orders of magnitude of computational power in hand, the stalwart whiteboard keeps watch, proud in knowing it still is needed to keep the world on track.
dagnamit - Wednesday, September 13, 2017 - link
Yes, but can it run Crysis?Zingam - Thursday, September 14, 2017 - link
No!fmttech - Sunday, July 29, 2018 - link
looks like NVIDIA DGX is breakthrough technology looking forward to what's next. https://www.future-micro.ca