Short answer: choose A40 when a qualified passive server needs 48GB, vGPU/virtual-workstation behavior or a supported two-card NVLink path. Choose A100 80GB when the measured workload crosses 48GB or specifically needs A100’s HBM, MIG or scale-up platform. A100 40GB has more published bandwidth but less capacity than A40, so it is not automatically the better local-AI card.
The decision in one minute
- A40 48GB: Ampere capability 8.6, 48GB GDDR6 ECC, 696GB/s, 300W, passive PCIe card, virtualization-first display behavior, vGPU, no MIG and a supported two-way NVLink bridge.
- A100 40GB PCIe: Ampere capability 8.0, 40GB HBM2, 1,555GB/s, 250W, passive compute accelerator with MIG and no display/NVENC.
- A100 80GB PCIe: 80GB HBM2e, 1,935GB/s, 300W; the only option here that raises the single-device capacity ceiling above 48GB.
- A100 SXM: 40GB/80GB module for a compatible HGX/DGX platform—not an ordinary PCIe card.
A40 vs A100 specifications
| Hardware | Architecture / capability | Memory / bandwidth | Power | Platform features |
|---|---|---|---|---|
| NVIDIA A40 | Ampere / 8.6 | 48GB GDDR6 ECC / 696GB/s | 300W | Passive PCIe; vGPU; display ports disabled by default; no MIG; two-way NVLink support |
| A100 40GB PCIe | Ampere / 8.0 | 40GB HBM2 / 1,555GB/s | 250W | Passive PCIe compute accelerator; MIG; no display/NVENC |
| A100 80GB PCIe | Ampere / 8.0 | 80GB HBM2e / 1,935GB/s | 300W | Passive PCIe compute accelerator; MIG; no display/NVENC |
| A100 40GB/80GB SXM | Ampere / 8.0 | 40GB HBM2 / 80GB HBM2e; 1,555 / 2,039GB/s | 400W | SXM for compatible HGX/DGX; MIG and NVLink/NVSwitch platform |
Both products are Ampere-generation data-center cards, but capability 8.6 and 8.0 are different software targets. A40 emphasizes virtual graphics, vGPU and configurable display behavior; A100 emphasizes compute, MIG and scale-up systems. Confirm the exact driver, CUDA toolkit, framework, kernels, vGPU licensing, hypervisor and qualified host.
Exact Qwen3.5 file-size boundaries
Margins subtract decimal file totals from nominal decimal memory. They exclude KV cache, hybrid-attention state, vision processing, allocator behavior and temporary buffers. Tight positive margins are warnings, not proven fits.
| Artifact | Files | A100 40GB | A40 48GB | A100 80GB |
|---|---|---|---|---|
| Qwen3.5-35B-A3B Q8_0 + F16 projector | 37.802GB | 2.198GB — unsafe | 10.198GB | 42.198GB |
| Qwen3.5-35B-A3B official BF16 | 71.904GB | Does not fit files | Does not fit files | 8.096GB — bounded |
| Qwen3.5-122B-A10B Q4_K_M + projector | 77.446GB | Does not fit files | Does not fit files | 2.554GB — unsafe |
| Qwen3.5-122B-A10B official GPTQ-Int4 | 78.864GB | Does not fit files | Does not fit files | 1.136GB — unsafe |
| Qwen3.5-122B FP8 / Q8 + projector / BF16 | 127.163 / 130.781 / 250.173GB | Does not fit files | Does not fit files | Does not fit files |
A40 holds the listed 35B Q8-plus-projector files with 10.198GB nominal room, while A100 40GB leaves only 2.198GB. A100 80GB changes the 35B BF16 file-fit outcome, yet its two listed 122B 4-bit margins remain unsafe. HBM bandwidth does not override capacity or establish tokens per second.
Cooling, power and form factor are purchase gates
- Every PCIe card here is passive. The host needs documented directed airflow; physical fit in a tower is insufficient.
- A40 power: NVIDIA specifies an 8-pin CPU power connector. Do not substitute a superficially similar PCIe GPU cable.
- A100 SXM: a bare module requires its matching baseboard, cooling, power, firmware and management stack.
- MIG: A100 can be partitioned, so record the visible slice and memory. A40 does not support MIG.
- NVLink: A40 supports a two-way low-profile bridge; A100 support depends on form factor/platform. Neither automatically becomes one transparent memory pool.
Which should you buy?
- Choose A40 when a measured workload fits 48GB and passive-server, vGPU/virtual-workstation or supported two-card behavior matters.
- Choose A100 40GB only when the workload fits below 40GB with real headroom and specifically benefits from A100 HBM, MIG or compute-platform behavior.
- Choose A100 80GB when crossing 48GB prevents offload or partitioning and the passive-PCIe or SXM platform is already justified.
- Benchmark complete systems when speed or efficiency decides: hold artifact, runtime, context, concurrency, placement and timing definition constant.
Create the missing controlled pair
The current fixed database has no accepted A40 or A100 observations. Load an exact starter, then record host, airflow, connector, exposed memory, MIG/NVLink, driver/CUDA, artifact, runtime, context, concurrency, timing definition and peak use. Nothing is saved or submitted automatically.
Used-market checklist
- Verify the exact label and part number: A40, A100 PCIe 40GB, A100 PCIe 80GB or A100 SXM.
- Request current device identity, visible memory, health/error evidence and any MIG state.
- Document the complete host, directed airflow, 8-pin CPU power or accelerator platform and PCIe lanes.
- Confirm driver/runtime support and vGPU/virtualization licensing where relevant.
- Compare delivered-system cost, warranty and return terms via the screened-offers directory and used-hardware checklist.
Continue: read the A40 guide, A100 40GB vs 80GB guide, 48GB vs 80GB capacity guide, VRAM calculator, or save private criteria and follow RSS.
Sources and limits
- NVIDIA A40 datasheet
- NVIDIA A40 product page
- NVIDIA A40 product brief and NVLink requirements
- NVIDIA A100 40GB and 80GB datasheet
- NVIDIA CUDA compute-capability list
- Qwen3.5-35B file provenance and Qwen3.5-122B file provenance
Reviewed October 25, 2026. Specifications are sourced from NVIDIA. HighVRAM has not independently benchmarked an A40–A100 pair. Used configurations, licensing and prices change.
Considering the Hopper step instead? Read the direct A40 versus H100 comparison for 48GB, 80GB and 94GB model-fit boundaries plus PCIe, SXM, MIG and NVLink safeguards.
Newer consumer GPU or larger accelerator? Compare RTX 5090 with A100 40GB/80GB using the same capacity-first method.
The A100 also has a direct compact-system comparison: DGX Spark vs A100.