Parts/Accelerators/NVIDIA Tesla P100
Accelerators · Pascal · 2016
NVIDIA Tesla P100
First NVIDIA data center GPU with HBM2 and NVLink.
Why this part
The first NVIDIA part with HBM and the first with NVLink, and both changes were about feeding the arithmetic rather than increasing it. Stacked memory on the same package tripled bandwidth over GDDR5, and NVLink gave a path off the board that was not PCIe. Piz Daint's upgrade to P100 doubled its performance without changing its footprint, which is the argument that persuaded a generation of European centres to accelerate.
Systems using this part
- Systems
- 28
- Rmax underneath
- 70.56 PFlop/s
- Units recorded
- 7,846
15 of 28 edges disclose a quantity
| System | Role | Country | Units | System Rmax | Supplied by | Rank |
|---|---|---|---|---|---|---|
| Saga 2019-10 | Accelerator | Norway | - | - | NVIDIA | - |
| Mustang (AFRL) 2018-12 | Accelerator | USA | - | 3.22 PFlop/s | NVIDIA | #48 → |
| Taiwania 2018-05 | Accelerator | Taiwan | 256 | 1.33 PFlop/s | NVIDIA | #148 → |
| HPC4 2018-01 | Accelerator | Italy | 3,200 | 12.21 PFlop/s | NVIDIA | #13 → |
| CMA PAI-B 2018 | Accelerator | China | - | 2.55 PFlop/s | NVIDIA | #68 → |
| OzSTAR (2018) 2018 | Accelerator | Australia | - | - | NVIDIA | - |
| Genius (KU Leuven) 2018 | Accelerator | Belgium | 24 | - | NVIDIA | - |
| OCTOPUS 2017-12 | Accelerator | Japan | 148 | - | NVIDIA | - |
| MN-1 2017-09 | Accelerator | Japan | 1,024 | 1.39 PFlop/s | NVIDIA | #92 → |
| TSUBAME3.0 2017-08 | Accelerator | Japan | - | 8.13 PFlop/s | NVIDIA | #59 → |
| Piz Daint 2017-06 | Accelerator | Switzerland | - | 21.23 PFlop/s | NVIDIA | #68 → |
| Graham 2017-05 | Accelerator | Canada | - | 1.23 PFlop/s | NVIDIA | #166 → |
| Cedar (SFU) 2017-04 | Accelerator | Canada | 584 | 1.34 PFlop/s | NVIDIA | #410 → |
| Wilkes-2 2017 | Accelerator | United Kingdom | - | 1.19 PFlop/s | NVIDIA | #101 → |
| RAIDEN 2017 | Accelerator | Japan | 192 | 1.21 PFlop/s | NVIDIA | #171 → |
| Reedbush-H 2017 | Accelerator | Japan | 240 | 800.00 TFlop/s | NVIDIA | #204 → |
| Reedbush-L 2017 | Accelerator | Japan | 256 | 810.00 TFlop/s | NVIDIA | #292 → |
| Kukai 2017 | Accelerator | Japan | - | 460.70 TFlop/s | NVIDIA | #2 → |
| AIST AI Cloud 2017 | Accelerator | Japan | 400 | 961.00 TFlop/s | NVIDIA | #3 → |
| CNIC Era-AI 2017 | Accelerator | China | - | 1.24 PFlop/s | NVIDIA | #110 → |
| Facebook DGX-1 Cluster 2017 | Accelerator | USA | - | 3.31 PFlop/s | NVIDIA | #32 → |
| DGX SaturnV 2016-11 | Accelerator | USA | 992 | 4.00 PFlop/s | NVIDIA | #29 → |
| GOSAT-2 RCF2 2016-03 | Accelerator | Japan | - | 770.40 TFlop/s | NVIDIA | #8 → |
| Owens 2016 | Accelerator | USA | 160 | 705.97 TFlop/s | NVIDIA | #273 → |
| Comet 2015-04 | Accelerator | USA | 144 | - | NVIDIA | - |
| Lomonosov-2 2014 | Accelerator | Russia | - | 2.48 PFlop/s | NVIDIA | - |
| Resnick HPC Cluster | Accelerator | USA | 200 | - | NVIDIA | - |
| Hawk (Cardiff) | Accelerator | United Kingdom | 26 | - | NVIDIA | - |
“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.