Update 05/24: NVIDIA has since reached out to us, informing us that their previous statement about 32GB cards replacing 16GB cards was in error, and that the 16GB V100 SKUs will remain in production. For more details, please see this article.

When first launched last year, the original Tesla V100 shipped with 16GB of HBM2 memory. Now just a little less than a year into its lifetime, NVIDIA is announcing that their workhorse server accelerator is getting a memory capacity bump in the form of new 32GB SKUs, effectively immediately.

For the last couple of years now, NVIDIA has been relying on 4GB (4-Hi) HBM2 memory stacks for their Tesla P100 and Tesla V100 products, as this was the first HBM2 memory to be ready in reasonable commercial volumes. Now that Samsung and SK Hynix have a better grip on HBM2 manufacturing, 8GB (8Hi) HBM2 stacks are far more readily available and reliable. As a result, the conditions are right for NVIDIA to finally give their Tesla cards a long-awaited memory upgrade.

This upgrade will be across the entire Tesla V100 family – all SKUs for SMX and PCIe cards are getting 32GB SKUs. The Tesla V100’s specifications otherwise remain identical, with the same GPU and memory clocks along with the same TDPs. We’re also told that the mechanical specifications are identical as well, which would mean that the 8-Hi stacks won’t cause any cooling problems due to changes in memory stack height.

NVIDIA Tesla/Titan Family Specification Comparison
  Tesla V100
Tesla V100
Titan V
Tesla P100
CUDA Cores 5120 5120 5120 3584
Tensor Cores 640 640 640 N/A
Core Clock ? ? 1200MHz 1328MHz
Boost Clock 1455MHz 1370MHz 1455MHz 1480MHz
Memory Clock 1.75Gbps HBM2 1.75Gbps HBM2 1.7Gbps HBM2 1.4Gbps HBM2
Memory Bus Width 4096-bit 4096-bit 3072-bit 4096-bit
Memory Bandwidth 900GB/sec 900GB/sec 653GB/sec 720GB/sec
12GB 16GB
L2 Cache 6MB 6MB 4.5MB 4MB
Half Precision 30 TFLOPS 28 TFLOPS 27.6 TFLOPS 21.2 TFLOPS
Single Precision 15 TFLOPS 14 TFLOPS 13.8 TFLOPS 10.6 TFLOPS
Double Precision 7.5 TFLOPS 7 TFLOPS 6.9 TFLOPS 5.3 TFLOPS
Tensor Performance
(Deep Learning)
GPU GV100 GV100 GV100 GP100
Transistor Count 21B 21B 21.1B 15.3B
TDP 300W 250W 250W 300W
Form Factor Mezzanine (SXM2) PCIe PCIe Mezzanine (SXM2)
Cooling Passive Passive Active Passive
Manufacturing Process TSMC 12nm FFN TSMC 12nm FFN TSMC 12nm FFN TSMC 16nm FinFET
Architecture Volta Volta Volta Pascal

It should be noted that this upgrade is a wholesale replacement of the existing 16GB versions in NVIDIA’s product stack, as NVIDIA won’t be retaining the 16GB versions now that the 32GB versions are out. So all new cards sold by NVIDIA going forward will be the 32GB cards, and OEMs will be making the same transition in Q2 as their 16GB stocks are depleted and replaced with 32GB cards.

NVIDIA’s Tesla V100-equipped systems will also be getting the same upgrade. Both the DGX-1 server and DGX Station will now ship with the 32GB cards; and indeed these are the only capacities that will be offered for those systems going forward. And while OEMs are a bit farther behind, ultimately I expect they’ll make the same move since the 16GB cards are going away anyhow. Meanwhile NVIDIA isn’t officially talking about pricing here for either the new V100 cards or the updated DGX systems. However as the new parts are essentially drop-in replacements in their respective ongoing lines, there aren’t any signs right now that pricing is changing.

As far as workloads go, since the specifications of the Tesla V100 aren’t changing outside of memory capacity, the benefit and performance impact of this upgrade is almost entirely rooted in dataset size. Workloads that rely on datasets that couldn’t previously fit entirely within a Tesla V100 card should see a sizable boost, as will any applications that can benefit from more local memory for caching purposes. However in pure compute-bound scenarios, the new cards shouldn’t perform any differently than the outgoing models.

Meanwhile, as for NVIDIA's workstation-oriented Titan V, nothing is being announced at this time. The 12GB card could easily be bumped to 24GB using the 8-Hi stacks, however it's not clear if NVIDIA is in a hurry to do so. In practice it will almost certainly come down to whether NVIDIA wants to keep GV100 GPU production to a single line for efficiency reasons, or if they'll operate two lines as part of a broader product segmentation strategy.

Lastly, this upgrade means that NVIDIA has finally caught up with arch-rival AMD in total memory capacity. AMD has been shipping a 32GB FirePro S9170 card using 16x2GB GDDR5 modules for the better part of the last 3 years now, a mark that NVIDIA couldn’t catch up to until 8-Hi HBM2 was ready. The Tesla V100 has always been significantly more powerful than the W9100 regardless, but this finally brings NVIDIA to parity in the one area where they were trailing AMD in the server space.

Comments Locked


View All Comments

  • Yojimbo - Tuesday, March 27, 2018 - link

    Before the Summit and Sierra supercomputers go live their basic computing unit is already out of date?! Seems a little bit weird. Perhaps these supercomputers were always getting the 32 GB variant and NVIDIA didn't want to say anything about it and hear people complain about why can't they also get 32 GB.

    I went to the Summit web site and did a quick calculation on their listed system memory numbers and figured that to reach their >10 PB of total system memory for 4600 total nodes there needs to be at least 27 GB of HBM memory per GPU, so perhaps it's true that the Summit GPUs have 32 GB of HBM2.
  • ronnybrendel - Wednesday, March 28, 2018 - link

    512*4600 (RAM) + 16*6*4600 (HBM) + 1600*4600 (NVMe disk) = 10156800 GB.
    No need for 32GB.
    I don't think Summit has 32GB cards, which is a shame.
    I wonder if they knew that this was coming.
    On the other hand for the price tag of the whole machine (< $300M), the cards must have been cheap.
  • Yojimbo - Wednesday, March 28, 2018 - link

    10,156,800 GB is only 9.686 PB
  • Lord of the Bored - Wednesday, March 28, 2018 - link

    "only" 9.686 PB
  • Yojimbo - Thursday, March 29, 2018 - link

    Did you read the entire thread?
  • mode_13h - Wednesday, March 28, 2018 - link

    Ah, perhaps that's why Titan V has remained in stock, lately... they had all that excess inventory of 4 GB HMB2 dies to burn through.

Log in

Don't have an account? Sign up now