id
accelerators/nvidia-a100-80gb-pcie
Memory bandwidth (GB/s)
1,935
Bandwidth basis
HBM2e 1,935 GB/s as published (nv-a100-datasheet), 5120-bit bus.
Peak FLOPS
precisionfp64tflops9.7denseyesvendor_claimyessource_idnv-a100-datasheetnotes—basisvendor-spec-sheetbasis_detailA100 80GB PCIe datasheet, FP64 row printed as '9.7 TFLOPS' (vector). The PCIe part publishes no FP64 Tensor Core row, so there is no companion 19.5 figure here - that is why this record has one 'other' row and nvidia-a100-40gb-sxm4 has two.unitfp64-vectorprecisionfp32tflops19.5denseyesvendor_claimyessource_idnv-a100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for NVIDIA A100 80GB PCIe (source of record: 'NVIDIA A100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisiontf32tflops156denseyesvendor_claimyessource_idnv-a100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own tf32 for NVIDIA A100 80GB PCIe (source of record: 'NVIDIA A100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionbf16tflops312denseyesvendor_claimyessource_idnv-a100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own bf16 for NVIDIA A100 80GB PCIe (source of record: 'NVIDIA A100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionfp16tflops312denseyesvendor_claimyessource_idnv-a100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp16 for NVIDIA A100 80GB PCIe (source of record: 'NVIDIA A100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionint8tflops624denseyesvendor_claimyessource_idnv-a100-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int8 for NVIDIA A100 80GB PCIe (source of record: 'NVIDIA A100 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.
notes
AUDIT CORRECTION: the dense/sparse flag on the tensor rows was inverted. FP16 312 / INT8 624 TFLOPS are DENSE. NVIDIA A100 page prints '312 | 624*' where * = With sparsity. These rows were previously flagged dense:false, which wrongly told a reader they were sparse. The stored numbers are correct dense values. Same Tensor Core throughput as A100 80GB SXM4 - the PCIe cut is memory-bandwidth and power limited, not compute limited. Tensor figures are WITH SPARSITY; dense is half (TF32 156, BF16/FP16 312, INT8 624 TOPS dense). NVLink on the PCIe card is bridge-only, up to 2 GPUs at 600 GB/s, versus SXM's full 600 GB/s with NVSwitch at up to 16 GPUs. 300W vs 400W. The 'other' row is FP64 9.7 TFLOPS (dense) - this schema's precision enum has no fp64 member.