Short answer: NVIDIA A40 provides 48GB GDDR6 ECC on an Ampere data-center card, but it is not an RTX A6000 with a different label. A40 uses passive cooling, an 8-pin CPU power connector and virtualization-first display behavior. It can be a useful CUDA memory tier only when the host chassis, airflow, power and software are deliberately designed around it.
Before buying an A40
- Verify directed server airflow; the A40 heatsink is passive.
- Verify the auxiliary connector is the documented 8-pin CPU connector, not an assumed PCIe-GPU pinout.
- Expect display connectors to be disabled in the default virtualization configuration unless enabled with management tools.
- Confirm the exact runtime supports compute capability 8.6 and the intended multi-GPU behavior.
NVIDIA A40 specifications relevant to local AI
| Architecture | NVIDIA Ampere |
|---|---|
| GPU memory | 48GB GDDR6 with ECC |
| Published memory bandwidth | 696GB/s in NVIDIA’s A40 datasheet |
| Compute capability | 8.6 |
| Maximum power | 300W |
| Power connector | 8-pin CPU |
| Form factor | 10.5-inch, dual-slot PCIe card |
| Cooling | Passive |
| NVLink | Two-way low-profile bridge; 112.5GB/s bidirectional |
| MIG | No |
What can fit in 48GB?
A40 has the same nominal capacity as RTX A6000, RTX 6000 Ada, L40S and Radeon PRO W7900. Fit depends on the exact model artifact or tensors, KV cache, runtime workspaces, context, batch size and concurrency. A 70B-class 4-bit artifact can fit in some configurations but remain tight as cache or runtime allocations grow.
Use the 70B VRAM requirements guide and LLM VRAM calculator. If 48GB is the chosen tier, compare A40 with the broader 48GB GPU decision guide.
Passive cooling is a system requirement
The card has no onboard blower. The server must force sufficient air through the heatsink across the expected ambient temperature and workload. A desktop case, open bench or improvised fan is not proof of a validated thermal design. Confirm inlet temperature, airflow direction, pressure, adjacent-card spacing and sustained temperatures in the final chassis.
Power and display traps
- 8-pin CPU power: use the documented cable and server power-delivery path; connector shape alone does not establish safe pin compatibility.
- Display ports: NVIDIA documents three DisplayPort 1.4 connectors, but the A40 is configured for virtualization by default with physical outputs disabled. Enabling them requires the appropriate management procedure and should not be assumed in a used listing.
- 300W is card power: size the PSU, cabling, circuit and cooling for the whole system, not only the accelerator.
CUDA, vGPU and compute capability
NVIDIA lists A40 at compute capability 8.6. Confirm the driver, CUDA toolkit, framework build, runtime, quantization kernels and compiled extensions. NVIDIA also lists vGPU software support, but licenses and supported profiles are separate from owning the card. Do not include paid virtualization features in a plan until entitlement and platform support are confirmed.
Can two A40 cards make 96GB?
NVIDIA supports a two-way NVLink bridge and explicitly conditions 96GB scaling on application support. NVLink does not automatically present two A40 cards as one transparent 96GB device. The runtime must partition the model or workload, and per-device headroom and duplicated allocations still matter. Read the multi-GPU VRAM guide.
A40 vs nearby 48GB alternatives
| Alternative | Main reason to compare | Different constraint |
|---|---|---|
| RTX A6000 | Same Ampere generation, 48GB and 300W | Active workstation cooling and different power/display integration |
| RTX 6000 Ada | Newer active-cooled 48GB CUDA workstation card | No NVLink; CEM5 16-pin power |
| L40S | Newer passive 48GB Ada server accelerator | 350W, compute capability 8.9 and no NVLink |
| Radeon PRO W7900 | Active-cooled 48GB workstation alternative | ROCm software lane rather than CUDA |
Used A40 checklist
- Request label and part-number photos; distinguish A40 from A100, A30 and RTX A6000.
- Ask for current device identity, health and memory-error evidence from the exact card.
- Document the chassis airflow and temperature plan before purchase.
- Verify the 8-pin CPU power cable and pinout from the system or board documentation.
- Confirm slot clearance, firmware, PCIe lanes and a separate management/display plan.
- Check whether the planned vGPU feature requires a license.
- Review return terms and test promptly under a representative sustained workload.
Use the full used-GPU buying checklist for seller evidence and arrival testing.
Official NVIDIA sources
- NVIDIA A40 datasheet
- NVIDIA A40 product page
- NVIDIA A40 product brief and NVLink requirements
- NVIDIA CUDA compute-capability table
Reviewed October 25, 2026. Specifications are sourced from NVIDIA. HighVRAM has not independently benchmarked A40. Used-card condition, licensing, availability and prices change.
Same 48GB Ampere tier, different cooling: compare NVIDIA A40 with RTX A6000 before choosing a passive server or active workstation installation.
Older or newer passive 48GB server card? Compare NVIDIA A40 with L40S across architecture, power connector, bandwidth, NVLink and media engines.
Would A100 40GB or 80GB better fit the server? Compare A40 with A100 across exact Qwen3.5 margins, HBM bandwidth, MIG, NVLink and PCIe/SXM requirements.
Would Hopper capacity change the deployment? Compare NVIDIA A40 with H100 across exact 48GB/80GB/94GB boundaries, cooling, power, MIG, NVLink and PCIe/SXM identity.
Passive server card or active Ada workstation? Compare A40 with RTX 6000 Ada while holding 48GB and 300W constant and isolating cooling, connector, NVLink, software and display differences.
Considering a newer consumer workstation instead? Compare A40 48GB with RTX 5090 32GB across exact model fit, Blackwell versus Ampere, bandwidth, power and host requirements.