NVIDIA Formally Announces PCIe Tesla V100: Available Later This Yearby Nate Oh on June 20, 2017 11:00 AM EST
Similar to last year, at this year's International Supercomputing Conference (ISC) NVIDIA has announced and detailed a PCI Express version of their latest Tesla GPU accelerator, the Volta-based V100. The conference itself runs from June 19 to 22, and with several speakers from NVIDIA scheduled for events tomorrow, NVIDIA is set to outline its next-generation efforts in HPC and deep learning with Volta.
With Volta discussed and described at their GPU Technology Conference in mid-May, NVIDIA upped the ante in terms of both features and reticle size: V100 is 815mm2 of custom TSMC 12FFN silicon, chock full of tensor cores and unified L1 cache per SM, along with many more fundamental – and as of yet not fully revealed – microarchitectural changes.
Like the previous Pascal iteration, the Tesla V100 PCIe offers a more traditional form factor as opposed to NVIDIA’s own mezzanine-type SXM2 form factor. This allows vendors to drop Tesla cards in traditional PCIe systems, making the cards far more accessible to server builders who don't want to build around NVIDIA's SXM2 connector or carrier board. The tradeoff being that the PCIe cards have a lower 250W TDP, and they don't get NVLink, instead relying on just PCIe.
|NVIDIA Tesla Family Specification Comparison|
|Memory Clock||1.75Gbps HBM2||1.75Gbps HBM2||1.4Gbps HBM2||1.4Gbps HBM2|
|Memory Bus Width||4096-bit||4096-bit||4096-bit||4096-bit|
|Half Precision||30 TFLOPS||28 TFLOPS||21.2 TFLOPS||18.7 TFLOPS|
|Single Precision||15 TFLOPS||14 TFLOPS||10.6 TFLOPS||9.3 TFLOPS|
|Double Precision||7.5 TFLOPS
|120 TFLOPS||112 TFLOPS||N/A||N/A|
|GPU||GV100 (815mm2)||GV100 (815mm2)||GP100 (610mm2)||GP100 (610mm2)|
|Form Factor||Mezzanine (SXM2)||PCIe||Mezzanine (SXM2)||PCIe|
|Manufacturing Process||TSMC 12nm FFN||TSMC 12nm FFN||TSMC 16nm FinFET||TSMC 16nm FinFET|
On the surface, the addition of tensor cores is the most noticeable change. To recap, tensor cores can be liked to a series of unified ALUs that are able to multiply two 4x4 FP16 matrices together and subsequently add that product to an FP16 or FP32 4x4 matrix in a fused multiply add operation, as opposed to conventional FP32 or FP64 CUDA cores. In the end, this means that for very specific kinds (and specifically programmed) workloads, Volta can take advantage of the 100+ TFLOPS capability that NVIDIA has tossed into the mix.
As for the specific specifications of the PCIe Tesla V100, it's similarly configured to the SXM2 version, getting the same number of CUDA cores and memory capacity, however operating at a lower clockspeed in-line with its reduced 250W TDP. Based on NVIDIA's throughput figures, this puts the PCIe card's boost clock at around 1370MHz, 85MHz (~6%) slower than the SXM2 version.
Interestingly, unlike the Tesla P100 family, NVIDIA isn't offering a second-tier PCIe card based on salvaged chips; so this generation doesn't have an equivalent to the 12GB PCIe Tesla P100. NVIDIA's experience with GP100/interposer/HBM2 assembly as well as continuing production of HBM2 has likely reduced the need for memory-salvaged parts.
Finally, PCIe-based Tesla V100 accelerators are “expected to be available later this year from NVIDIA reseller partner and manufacturers,” including Hewlett Packard Enterprise, which will offer three different PCIe Volta systems.
Post Your CommentPlease log in or sign up to comment.
View All Comments
Drumsticks - Tuesday, June 20, 2017 - linkThe V100 is an incredible part. I wonder just how many will actually be available though. It can't be easy to manufacture something over 800mm area.
I also wonder how this will lead into Volta having cards. The P100 was faster than the initial run off Pascal. I wonder if we will actually be seeing the performance flag being moved next year, or if a small card like the 750 ti, or just a more efficient flagship. Think 1080 Ti, but less power. Maybe Vega won't actually have to compete with a GV102 until late 2018ish.
Qwertilot - Tuesday, June 20, 2017 - linkBest guess - and it obviously is a real guess, but they are very predictable - is a repeat of what they did with Pascal. So an 1170/80 with the 1180 a little ahead of the 1080ti in performance while using rather less power.
Then a Volta titan (V102 without all the compute stuff) with an 1180ti coming later.
Drumsticks - Tuesday, June 20, 2017 - linkThat's a fair point. The last time we had an arch revisdion without a major node change (780 Ti to 980), absolute performance didn't move a ton, maybe 10 to 20%, but obviously efficiency was way up. I wonder what we'll be seeing from Volta on the gaming side. From a TF efficiency perspective, the 980 was about 400GF below the 780 Ti, but about 10-30% faster. It'll be interesting to see if they can move that mark again with Volta.
FreckledTrout - Tuesday, June 20, 2017 - linkWith what they will charge 800mm wont matter even if they end having to toss half them in the trash. The DGX Station with 4 of these was quoted to cost $69,000.
andychow - Wednesday, June 28, 2017 - linkI was just about to ask: How are they making a 815mm2 chip? I thought the max was 700mm2, on Intel special fabs, and TSMC GloBo could do 610mm2 max. I'm not even talking about defects, but physical limitations like reticle size. Where are these being manufactured?
stanleyipkiss - Tuesday, June 20, 2017 - linkHow are the passively cooling a 250 W card?
nathanddrews - Tuesday, June 20, 2017 - linkBecause the rack environment such a card is used in sounds like an aircraft runway.
ddriver - Tuesday, June 20, 2017 - linkThere are fans on the rackmount case that pull air through the entire chassis.
kaesden - Tuesday, June 20, 2017 - linkits not really passively cooled when installed in a system. The card itself has no active cooling elements, but its put into an environment with loads of airflow from the server chassis. Servers don't give a f- about noise.
Gothmoth - Tuesday, June 20, 2017 - linknvidia should start producing x86 cpus too.
i know i know.. it is not that easy blah blah.. just saying that i would find it nice to have more competition.