inference/research
verifiedhigh · 0.90updated 2026-10-03

NVIDIA A30

nvidia ampere 2021

Fields

id
accelerators/nvidia-a30
Vendor
nvidia
Architecture
ampere
Release year
2021
Process (nm)
7
Form factors
pcie
VRAM (GB)
24
Memory type
HBM2
Memory bus (bit)
4,096
Memory bandwidth (GB/s)
933
Bandwidth basis
HBM2 933 GB/s as published (nv-a30-datasheet). GA100 die with only 4 of 6 HBM2 stacks enabled gives the 4096-bit bus and the odd 933 GB/s figure - this is why A30 bandwidth is not a round number.
Peak FLOPS
precisionfp64tflops5.2denseyesvendor_claimyessource_idnv-a30-datasheetnotes—basisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp64 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionfp32tflops10.3denseyesvendor_claimyessource_idnv-a30-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp32 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisiontf32tflops82.5denseyesvendor_claimyessource_idnv-a30-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own tf32 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionbf16tflops165denseyesvendor_claimyessource_idnv-a30-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own bf16 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionfp16tflops165denseyesvendor_claimyessource_idnv-a30-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own fp16 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.precisionint8tflops330.5denseyesvendor_claimyessource_idnv-a30-datasheetbasisvendor-spec-sheetbasis_detailTranscribed from the vendor's own int8 for NVIDIA A30 (source of record: 'NVIDIA A30 Tensor Core GPU datasheet', kind: spec-sheet). Recorded dense:true as printed.
TDP (W)
165
Unified memory
no
Consumer part
no
notes
AUDIT CORRECTION: the dense/sparse flag on the tensor rows was inverted. FP16 165 / INT8 330.5 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 Core figures are WITH SPARITY (sheet marks '*'): TF32 165 sparse / 82.5 dense, BF16+FP16 330 sparse / 165 dense, INT8 661 sparse / 330.5 dense. FP64 5.2 and FP32 10.3 are dense. A30 is the memory-rich inference part: 24GB HBM2 at only 933 GB/s and half-rate FP64 vs A100, so it is a bandwidth-per-watt and capacity part, not a FLOPS part. MIG profiles are asymmetric (4x6GB / 2x12GB / 1x24GB), which differs from A100's 7x10GB. NVLink is third-gen bridge-only, 2 GPUs at 200 GB/s. The 'other' row is FP64 5.2 TFLOPS (dense) - this schema's precision enum has no fp64 member.

Notes

AUDIT CORRECTION: the dense/sparse flag on the tensor rows was inverted. FP16 165 / INT8 330.5 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 Core figures are WITH SPARITY (sheet marks '*'): TF32 165 sparse / 82.5 dense, BF16+FP16 330 sparse / 165 dense, INT8 661 sparse / 330.5 dense. FP64 5.2 and FP32 10.3 are dense. A30 is the memory-rich inference part: 24GB HBM2 at only 933 GB/s and half-rate FP64 vs A100, so it is a bandwidth-per-watt and capacity part, not a FLOPS part. MIG profiles are asymmetric (4x6GB / 2x12GB / 1x24GB), which differs from A100's 7x10GB. NVLink is third-gen bridge-only, 2 GPUs at 200 GB/s. The 'other' row is FP64 5.2 TFLOPS (dense) - this schema's precision enum has no fp64 member.

Cross-references

Interconnect (2)