P100 SXM2 16GB
NVIDIAdatacenter| Overview | |
|---|---|
| Architecture | Pascal (2016) |
| Launch date | Apr 2016 (announced at GTC); datasheet dated Oct 2016 |
| Compute capability | 6.0 |
| Memory | |
| VRAM | 16 GB HBM2 |
| Memory bandwidth | 732 GB/s |
| ECC | Yes |
| Compute — vector | |
| FP64 | 5.3 TFLOPS |
| FP32 | 10.6 TFLOPS |
| FP16 | 21.2 TFLOPS |
| Cores & clocks | |
| Streaming Multiprocessors | 56 |
| Shader cores | 3,584 |
| Board & system | |
| TDP | 300 W |
| Form factor | SXM2 |
| Interconnect | NVLink — 160 GB/s |
Compatible CUDA Toolkit versions
Added 2026-09 alongside V100/T4/RTX 3090 to extend NVIDIA coverage further back — P100 predates this site's previous oldest NVIDIA entry (2020's A100) by four years and is still encountered on older training clusters. Core figures (CUDA core count, FP64/FP32/FP16 TFLOPS, memory size/bandwidth, TDP, ECC, form factor) sourced directly from NVIDIA's official Tesla P100 datasheet (nvidia.com/content/dam/en-zz/Solutions/ Data-Center/tesla-p100/pdf/nvidia-tesla-p100-datasheet.pdf, SXM2 column, footer dated Oct16), read via pdftotext. interconnectBandwidthGBs (160, NVLink) is from the companion Technical Overview PDF on the same page (nvidia-teslap100-techoverview.pdf), which explicitly states "four NVLink connections per GPU — each delivering 40 GB/sec bi-directional interconnect bandwidth, Tesla P100 delivers 160 GB/s bidirectional bandwidth in total."
fp16TFLOPSVector (21.2) is Pascal's native packed-FP16 throughput via CUDA cores, listed in the datasheet as "Half-Precision Performance" — correctly placed in this schema's vector (non-Tensor) field, not fp16TFLOPS/matrix, because Pascal has no Tensor Cores at all (introduced with Volta the following generation); fp16TFLOPS, bf16TFLOPS, tf32TFLOPS, fp8TFLOPS, and int8TOPS are all correctly omitted rather than left as zero. eccSupport is "Yes" per the datasheet's own note: "Native support with no capacity or performance overhead" (HBM2's on-package ECC, unlike GDDR5-based Kepler-generation cards which took a bandwidth/capacity hit for ECC). No transistor count, process node, PCIe generation, or MSRP in either sourced document — P100 shipped through OEM/server partners, not direct retail; a 12GB HBM2 SKU also exists but isn't modeled here.
2026-09-15 — added computeUnitCount (56 SMs) from Table 1 of NVIDIA's Volta architecture whitepaper (WP-08608-001_v1.1), whose GPU-comparison table covers Tesla K40/M40/P100/V100 side by side and lists GP100 (Pascal) at 56 SMs, 64 FP32 cores per SM, 3,584 FP32 cores per GPU — that last figure matching the shaderCoreCount already recorded here from the P100 datasheet itself. matrixCoreCount stays correctly empty: Pascal has no Tensor Cores (introduced with Volta the following year).