id
accelerators/nvidia-a10
Memory bandwidth (GB/s)
600
Bandwidth basis
GDDR6 12.5 Gbps effective x 384-bit / 8 = 600 GB/s. NVIDIA's A10 datasheet prints only the 600 GB/s aggregate; the 12.5 Gbps per-pin rate is derived from that aggregate (600 x 8 / 384). The commonly-cited 12 Gbps figure gives 576 GB/s, which does not match the datasheet.
Peak FLOPS
precisionfp32tflops31.2denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for NVIDIA A10 (source of record: 'NVIDIA A10 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisiontf32tflops62.5denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own tf32 for NVIDIA A10 (source of record: 'NVIDIA A10 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionbf16tflops125denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own bf16 for NVIDIA A10 (source of record: 'NVIDIA A10 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionfp16tflops125denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp16 for NVIDIA A10 (source of record: 'NVIDIA A10 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionint8tflops250denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int8 for NVIDIA A10 (source of record: 'NVIDIA A10 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionint4tflops500denseyesvendor_claimyessource_idnv-a10-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int4 for NVIDIA A10 (source of record: 'NVIDIA A10 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 125 / INT8 250 TFLOPS are the DENSE values (sparse = 2x). These rows were previously flagged dense:false, which wrongly told a reader they were sparse. The stored numbers are correct dense values. Tensor figures are WITH SPARITY (sheet marks '*'): TF32 125 sparse / 62.5 dense, BF16+FP16 250 sparse / 125 dense, INT8 500 sparse / 250 dense, INT4 1000 sparse / 500 dense. FP32 31.2 is dense. A10 is GDDR6, NOT GDDR6X - do not assume the A10G/ workstation parts' memory here; A10G is a separate cut. 150W single-slot FHFL, no NVLink, no MIG, no FP64 Tensor Core path listed on the sheet (GA102 has no FP64 tensor path). Bandwidth is under 1/3 of H100 SXM at a quarter of the watts, which is why this generation's inference cards are memory-starved. No 'other' row: GA102 (A10) has no FP64 hardware path of significance, and the datasheet lists no FP64 figure.