A800PCLE single card
-
Detail
Architecture: Ampere, TSMC 7nm, third-generation tensor core
A800 40GB PCIe (active heat dissipation)
Graphics memory: 40GB HBM2, 5120bit, bandwidth 1555GB/s
CUDA core: 6912; tensore: 432
Computing power
FP64: 9.7 TFLOPS
FP32: 19.5 TFLOPS
TF32: 156 TFLOPS (dense)/312 TFLOPS (sparse)
BF16/FP16: 312 TFLOPS (dense)/624 TFLOPS (sparse)
INT8: 624 TOPS (dense)/1248 TOPS (sparse)
Interconnection: PCIe4.0 x16 (64GB/s); NVLink two-way 400GB/s (dual SIM interconnection)
Power consumption TDP: 240W
Form: double-slot full-length FHFL, active air cooling; no display output
MIG: Support, up to 7 instances, single instance 5GB
Cache: L2 Cache 40MB
A800 80GB PCIe (passive heat dissipation)
Graphics memory: 80GB HBM2e, bandwidth 1935GB/s
CUDA core: 6912; tensore: 432
The computing power is the same as above.
Power consumption TDP: 300W
Form: double-groove full-length FHFL, passive heat dissipation
MIG: Support, up to 7 instances, single instance 10GB
Remarks: A800 is a domestic compliance version. Compared with A100, NVLink interconnection bandwidth has been reduced from 600GB/s to 400GB/s; PCIe board, can be directly connected to the standard server, no HGX carrier board, compatible with CUDA/NCCL, used for AI training, reasoning, HPC high-performance computing.
-
Customer ReviewsNo comments