NVIDIA V100 SXM2 32GB
nvidia volta 2017
Fields
id
accelerators/nvidia-v100-sxm2-32gbVendor
nvidia
Architecture
volta
Release year
2017
Process (nm)
12
Form factors
sxm
VRAM (GB)
32
Memory type
HBM2
Memory bus (bit)
4,096
Memory bandwidth (GB/s)
900
Bandwidth basis
900 GB/s as published for V100 SXM2 32GB (nv-v100-datasheet). HBM2 on a 4096-bit interface: the per-pin rate is 1.76 Gbps (900 x 8 / 4096). The commonly-cited 1.95 Gbps figure gives 998.4 GB/s, which does not match the datasheet. Treat the datasheet number as authoritative and the bus width as the 4096-bit Volta HBM2 interface.
Peak FLOPS
precisionfp64tflops7.8denseyesvendor_claimyessource_idnv-v100-datasheetnotes—basisvendor-spec-sheetbasis_detailV100 datasheet, FP64 row '7.8 TFLOPS'. Volta publishes no separate FP64 tensor-core path, so one row is the complete published set.precisionfp32tflops15.7denseyesvendor_claimyessource_idnv-v100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for NVIDIA V100 SXM2 32GB (source of record: 'NVIDIA V100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionfp16tflops125denseyesvendor_claimyessource_idnv-v100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp16 for NVIDIA V100 SXM2 32GB (source of record: 'NVIDIA V100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.
TDP (W)
300
Unified memory
no
Consumer part
no
notes
Tesla-generation part. NVIDIA quotes 125 TFLOPS 'Tensor Performance' for SXM2 (FP16, dense - the datasheet has no sparsity notation, and Volta Tensor Cores have no 2:4 sparse path). PCIe V100 32GB is the same silicon at 14 TFLOPS FP32 / 112 TFLOPS FP16 tensor / 900 GB/s / 250W but is a different form factor and power envelope. Bus-width figure of 4096-bit is the standard Volta HBM2 width and is NOT stated in this datasheet (the sheet omits it) - treat it as lower-confidence than the bandwidth number. The one 'other' row here is FP64 (7.8 TFLOPS, dense) - the schema's precision enum has no fp64 member, so it is filed under 'other'.
Notes
Tesla-generation part. NVIDIA quotes 125 TFLOPS 'Tensor Performance' for SXM2 (FP16, dense - the datasheet has no sparsity notation, and Volta Tensor Cores have no 2:4 sparse path). PCIe V100 32GB is the same silicon at 14 TFLOPS FP32 / 112 TFLOPS FP16 tensor / 900 GB/s / 250W but is a different form factor and power envelope. Bus-width figure of 4096-bit is the standard Volta HBM2 width and is NOT stated in this datasheet (the sheet omits it) - treat it as lower-confidence than the bandwidth number. The one 'other' row here is FP64 (7.8 TFLOPS, dense) - the schema's precision enum has no fp64 member, so it is filed under 'other'.