A100 buyer intelligence
A100 40GB vs 80GB, PCIe vs SXM: Buy the Right Configuration
An A100 price is only meaningful after you know the memory capacity, form factor, and machine that will run it.
Four configurations, not one bargain
The NVIDIA A100 datasheet separates 40GB and 80GB models in both PCIe and SXM form factors. The 40GB models use HBM2; the 80GB models use HBM2e. The name “A100” alone leaves two buying decisions unanswered.
Compare like with like. A loose module, a PCIe accelerator, and a complete HGX server have different integration costs. A low asking price does not establish that the listing is usable in your machine.
40GB vs 80GB starts with a memory budget
Write down your model weights, precision, working memory, context length, batch size, and other processes that need GPU memory. A model file size is not the whole memory requirement. Test the intended software configuration before deciding that a capacity number is enough.
80GB provides more room on one GPU. It does not guarantee twice the application speed or remove the need to profile your workload. A 40GB card may be sufficient when the measured workload fits; a cheaper card that requires a different workflow may cost more in time.
PCIe and SXM require different hosts
PCIe is an expansion-card lane; SXM4 belongs to a supported GPU platform and baseboard. NVIDIA lists reference maximum TDPs of 250W for 40GB PCIe, 300W for 80GB PCIe, and 400W for the SXM variants. These figures describe the accelerators, not the full server's wall power.
Check the exact OEM card and host documentation for power, cooling, firmware, mechanical support, and approved configurations. A physically available slot is not a compatibility certificate. For SXM, require the platform bill of materials rather than assuming a loose module includes its tray or cooling hardware.
The PCIe vs SXM guide explains the same buying trap in the H100 lane. Use it for the decision process, not as proof that A100 and H100 parts interchange.
The evidence checklist
- Photos of the actual label, part number, capacity, and PCIe or SXM form factor.
- The seller's included-items list: card or module, carrier, cables, heatsink, tray, or complete server.
- Exact host model and documented support for that part number.
- Recent identification and diagnostic output tied to the offered unit, with memory and error status visible.
- Cooling and power requirements, return terms, shipping protection, and warranty transfer conditions.
Compare the cost of a working system
Include the host, airflow, cabling, freight, power, noise, and a return plan. Do not compare a tested PCIe unit with an untested SXM pull as though memory capacity were the only difference.
For a multi-GPU purchase, ask how the application uses each GPU and the platform interconnect. Do not assume that adding cards automatically makes their memory one shared pool. See the networking and interconnect guide before buying bridges or fabric parts.
Verify the configuration, then check the market.
These are eBay affiliate searches, not verified inventory. Check the exact item and seller terms before buying.