inference/research
verifiedhigh · 0.85updated 2026-10-14

NVIDIA A100 40GB PCIe

nvidia ampere 2020

Fields

id
accelerators/nvidia-a100-40gb-pcie
Vendor
nvidia
Architecture
ampere
Release year
2020
Process (nm)
7
Form factors
pcie
VRAM (GB)
40
Memory type
HBM2
Memory bus (bit)
5,120
Memory bandwidth (GB/s)
1,555
Bandwidth basis
A100 40GB PCIe Product Brief (PB-10137-001_v03) specifies memory identically to the SXM4 part: 'Memory clock 1215 MHz', 'Memory type HBM2', 'Memory size 40 GB', 'Memory bus width 5120 bits', 'Peak memory bandwidth Up to 1555 GB/s'. 1215 MHz x 2 (DDR) x 5120 bit / 8 = 1555 GB/s. The bandwidth is the same as the SXM4 module; the difference is the 250 W power envelope and lower boost clocks.
Peak FLOPS
precisionfp16tflops312denseyesvendor_claimyessource_idacc-fill-nvidia-ampere-whitepaper-a100-40gbbasisvendor-whitepaperbasis_detailAmpere GA100 whitepaper Table 4 FP16 tensor dense 312 TFLOPS for the A100 40GB die. The PCIe card carries the same GA100-883AA-A1 silicon; the 250 W cap vs 400 W SXM4 means sustained clocks are lower in practice.unit—precisionint8tflops624denseyesvendor_claimyessource_idacc-fill-nvidia-ampere-whitepaper-a100-40gbbasisvendor-whitepaperbasis_detailAmpere GA100 whitepaper Table 4 INT8 dense 624 TOPS for the A100 40GB die; same silicon on the PCIe card.unit—
TDP (W)
250
Unified memory
no
Consumer part
no
notes
The zero-drama A100 route: the same GA100 die as nvidia-a100-40gb-sxm4 in a passive FHFL PCIe card that drops into any x16 slot - no HGX baseboard, no SXM4-to-PCIe adapter. 40 GB HBM2, 5120-bit, 1555 GB/s identical to SXM4; the differences are the 250 W cap (vs 400 W SXM4), no NVSwitch, and lower boost clocks. THE PRICE IS A FORM-FACTOR STORY, NOT A SILICON STORY: the used 40GB PCIe card asks ~$6,995-8,074 (2026-10-03 eBay sweep) - roughly 2.8x the $2,499 bare SXM4 module. That gap is the HGX baseboard problem priced in: SXM4-to-PCIe adapters at $648-1,200 put total-landed SXM+adapter at ~$3,150-3,700, still half the PCIe ask, but the adapter route is grey-market. Full current CUDA stack (CC8.0). Measured sibling: A100 80GB SXM does 133 tok/s Llama-3-8B Q4_K_M decode and 22-24 tok/s on 70B Q4_K_M (XiongjieDai, RunPod); the 40GB part has ~24% less bandwidth -> expect ~100-110 tok/s 8B, ~17-19 tok/s 70B.

Notes

The zero-drama A100 route: the same GA100 die as nvidia-a100-40gb-sxm4 in a passive FHFL PCIe card that drops into any x16 slot - no HGX baseboard, no SXM4-to-PCIe adapter. 40 GB HBM2, 5120-bit, 1555 GB/s identical to SXM4; the differences are the 250 W cap (vs 400 W SXM4), no NVSwitch, and lower boost clocks. THE PRICE IS A FORM-FACTOR STORY, NOT A SILICON STORY: the used 40GB PCIe card asks ~$6,995-8,074 (2026-10-03 eBay sweep) - roughly 2.8x the $2,499 bare SXM4 module. That gap is the HGX baseboard problem priced in: SXM4-to-PCIe adapters at $648-1,200 put total-landed SXM+adapter at ~$3,150-3,700, still half the PCIe ask, but the adapter route is grey-market. Full current CUDA stack (CC8.0). Measured sibling: A100 80GB SXM does 133 tok/s Llama-3-8B Q4_K_M decode and 22-24 tok/s on 70B Q4_K_M (XiongjieDai, RunPod); the 40GB part has ~24% less bandwidth -> expect ~100-110 tok/s 8B, ~17-19 tok/s 70B.

Cross-references

Interconnect (2)