The NVIDIA A100 is an Ampere-generation data-center accelerator for AI, analytics and high-performance computing. For a procurement decision, start with the exact module, server compatibility and workload memory requirement. A listing that says only “A100” is not enough to establish that the hardware will fit, operate or meet your acceptance criteria.
Technical review: 16 September 2026. This guide separates manufacturer specifications from buyer checks. It does not state current inventory, a fixed market price or a guaranteed performance improvement.
What should you confirm before requesting an A100 quote?
Request the full manufacturer and OEM part numbers, memory capacity, PCIe or SXM form factor, quantity, condition and destination. Include the server model and the intended workload. Ask the supplier to identify anything that remains unverified in writing rather than accepting a generic compatibility statement.
| Decision | Evidence to request | Why it affects the purchase |
|---|---|---|
| 40GB or 80GB | Exact product label and OEM configuration | The job must fit within the usable memory of the chosen configuration. |
| PCIe or SXM | Server bill of materials and supported accelerator list | These are different hardware formats; an SXM module is not a PCIe add-in card. |
| Power and cooling | Server documentation for the exact accelerator configuration | A mechanically plausible installation may still lack the required power delivery or airflow. |
| Condition and provenance | Written condition, serial-number records, available traceability and agreed inspection report | Photographs alone do not establish service history or authenticity. |
| Acceptance | Agreed diagnostic procedure, workload test, result format and rejection process | Both parties need the same definition of a conforming delivery. |
A100 specifications: distinguish memory, cache and bandwidth
NVIDIA lists A100 products with 40GB and 80GB GPU memory and both PCIe and SXM configurations. The architecture whitepaper specifies a 40MB L2 cache. Memory capacity, cache capacity and memory bandwidth are different measurements: a bandwidth figure expressed in TB/s does not mean the GPU has a terabyte-sized cache.
Use the NVIDIA A100 datasheet for the selected model and the A100 architecture whitepaper for architecture details. Record the document revision alongside the proposed part number so an RFQ does not combine specifications from different variants.
How should you interpret performance claims?
The A100 includes third-generation Tensor Cores. NVIDIA describes structured sparsity support that can increase supported Tensor Core operation throughput relative to the corresponding dense operations. That capability is not a promise to double every application, and headline generation-to-generation multipliers cannot be applied to an unspecified job.
For a comparison, define the model or application, dataset, numerical precision, batch size, software versions and acceptable output quality. Measure the metric the deployment needs: for example, time to a completed training result or inference latency at a stated throughput. Keep the test method identical across candidate systems and document any optimization used on only one candidate.
Do not choose A100, H100 or a GeForce GPU using a generic “times faster” statement. A buyer should compare the complete supported system, required features, deployment restrictions and measured workload result. Gaming comparisons do not establish suitability for a data-center procurement decision.
Can one A100 serve multiple workloads with MIG?
NVIDIA Multi-Instance GPU (MIG) allows supported A100 configurations to be partitioned into up to seven GPU instances. The available profiles determine memory and compute allocations. Confirm the profile for the exact memory variant rather than assuming every partition has the same capacity.
Configuration changes require operational planning. NVIDIA's MIG setup guide describes enabling MIG, GPU resets and instance management. Do not promise transparent resizing of a running job. Check driver, hypervisor and orchestration requirements, then test the intended lifecycle before production use.
Ask the platform owner to provide a workload-to-instance allocation table. It should show peak memory, concurrency, restart tolerance and who is responsible for changing profiles. A procurement specification can require evidence that this agreed configuration works, rather than merely requiring a card advertised as “MIG capable.”
What evidence belongs in an A100 sourcing RFQ?
- Hardware identity: exact MPN, OEM reference, memory variant, form factor and required accessories.
- Deployment: server model, current configuration, target quantity and engineering contact for compatibility review.
- Commercial scope: delivery destination, requested date, quoted condition, quote validity and separately stated freight or other charges.
- Quality scope: required traceability records, inspection criteria, diagnostic outputs and acceptance responsibilities.
- Remedy: written warranty scope, return procedure and treatment of a failed acceptance test, as agreed in the quotation and order.
If a supplier proposes another variant or accelerator, treat it as a separate engineering decision. Any alternative requires validation against the hardware, software and workload requirements. A similar product name or memory capacity does not establish interchangeability.
How much does an A100 cost?
This page does not publish a current unit price. Request a dated quotation identifying the exact part, quantity, condition, included accessories and applicable terms. Compare offers on the same scope. An old article's dollar figure is not evidence of an executable offer today, and the lowest unit price may omit requirements needed for acceptance.
What should happen before releasing a purchase order?
Reconcile the quotation line by line with the RFQ. Resolve differences in model, condition, test scope and required records before release. Record who approves technical substitutions and who accepts delivered hardware. For a pilot purchase, define the evidence required before expanding to the remaining quantity; do not treat the pilot result as proof of every later lot.
For a sourcing review, send your BOM or exact part-number list with the variant, quantity, destination and acceptance requirements. If the specification is incomplete, send the missing points as an inquiry. Availability, compatibility, lead time and commercial terms must be confirmed for the actual request.
