NVIDIA A100 PCIe 40GB
Ampere PCIe card at 250 W, and the clearest documented depreciation curve in this category, May 2020 to Q3 2026.
NVIDIA A100 PCIe 40GB is one of 24 data center GPUs tracked in the SecondWatt catalogue, each with specifications, lead times and indicative secondary-market pricing.
NVIDIA A100 PCIe 40GB dossier: 250 W passive, 1,555 GB/s HBM2, rack kW, export status, depreciation since 2020 and a $4,000 to $8,000 used range.
The NVIDIA A100 PCIe 40GB is an Ampere accelerator on a full-height full-length dual-slot PCIe card 10.5 inches long, rated 250 W with a passive thermal solution. NVIDIA's own product brief, document PB-10137-001_v03 of September 2020, gives 40 GB of HBM2 at up to 1,555 GB/s, PCI Express 4.0 x16, an NVLink bridge at 600 GB/s, GPU SKU GA100-883AA-A1, and clocks of 765 MHz base and 1,410 MHz boost. At 250 W this is the lowest-power HBM part in this category and the easiest A100 to place in an existing server. Compute is identical to every other A100 variant on NVIDIA's tables, so what a buyer gives up against the 80 GB cards is memory capacity, bandwidth and 5 GB rather than 10 GB multi-instance slices, not throughput. It is also the part with the most complete price history in this category. A single card was offered by a reseller at $12,500 in May 2020; Compute Exchange's Q3 2026 band is $4,000 to $6,000. That is 52 to 68 percent over 6.33 years, or 10.9 to 16.5 percent a year compounded, and it is the strongest evidence in this category that accelerator residuals behave predictably. Note that this variant no longer appears on NVIDIA's live A100 product page, which is the earliest end-of-life signal in this category.
Why it matters
This card carries the documented depreciation curve that makes residual-value estimates defensible in a category where almost nothing else has one. Three independent channels support a $4,000 to $8,000 range. At 250 W it also shows how wide the A100 power spread is on identical silicon, from 250 W here to 400 W on the SXM4 module.
Who buys it
Operators wanting genuine HBM bandwidth and double precision at the lowest power and price in the Ampere line. The 250 W passive card is the easiest A100 to fit into a conventional server fleet. Academic, teaching and scientific buyers come for 9.7 TFLOPS of FP64 at a fraction of current-generation cost, and multi-instance partitioning suits shared research clusters.
Role in the data center
Inference, fine-tuning and shared multi-tenant serving, plus academic and scientific double precision at 9.7 TFLOPS FP64. Seven 5 GB multi-instance slices suit research clusters running many small jobs. Lands well inside a 26 to 40 kW rack density band on air.
Power envelope
Specifications verified 2026-09-10.
- Per accelerator
- 250 W — Verified: Maximum configurable board power, manufacturer specification. source
- Cooling class
- Air-coolable — Derived: Per-accelerator power below 400 W — standard air-cooled halls.
No published node-level power rating exists for this part, so node, rack and per-MW figures are not stated. The baseboard sibling does carry one — A100 SXM4 80GB.
Node figures are published OEM system ratings; rack counts are arithmetic against a stated rack budget and move with your own cooling design. Assumption set version 2026-09-11a.
Key specifications
| Architecture | Ampere |
|---|---|
| GPU SKU | GA100-883AA-A1 |
| Memory capacity | 40 GB HBM2 |
| Memory bandwidth | up to 1,555 GB/s |
| Maximum thermal design power | 250 W |
| Thermal solution | Passive, server airflow |
| Form factor | Full-height full-length 10.5 in, dual slot |
| Host interconnect | PCI Express 4.0 x16 |
| NVLink bridge | 600 GB/s |
| Base clock | 765 MHz |
| Boost clock | 1,410 MHz |
| Multi-instance partitioning | Up to 7 at 5 GB each |
| FP64 | 9.7 TFLOPS |
| Accelerator-only load, 8 cards | 2.0 kW |
| Used range | 4,000 to 8,000 USD per card, three channels |
| Named in NVIDIA's 17 October 2023 export filing | Yes |
Technical summary
Architecture: Ampere, GPU SKU GA100-883AA-A1; 250 W maximum thermal design power. Memory: 40 GB HBM2, up to 1,555 GB/s bandwidth. Form factor: full-height full-length 10.5 inch dual-slot card, passive cooling. Interconnect: PCI Express 4.0 x16; NVLink bridge at 600 GB/s. Clocks: 765 MHz base, 1,410 MHz boost. Multi-instance partitioning: up to 7 instances of 5 GB each. Compute: FP64 9.7 TFLOPS, FP64 tensor core 19.5, FP32 19.5 TFLOPS. TF32 156 TFLOPS dense, 312 with sparsity; BF16 and FP16 312 dense, 624 with sparsity. Compute is identical to every other A100 variant; no FP8 support. Accelerator-only load, 8 cards: 2.0 kW; proxy-estimated node draw, not an OEM figure, 4.1 kW. Depreciation: $12,500 in May 2020 to $4,000 to $6,000 in Q3 2026. Export: named in NVIDIA's Form 8-K of 17 October 2023 as licence-controlled.
Major variations
A100 SXM4 40GB is the same memory configuration at 400 W on an HGX baseboard, with full NVLink at 600 GB/s across a module domain rather than a bridged pair. It is 1.6 times the power of this card for the same 40 GB and 1,555 GB/s. The 80 GB variants are separate parts: A100 PCIe 80GB at 300 W with 1,935 GB/s and 10 GB multi-instance slices, and A100 SXM4 80GB at 400 W with 2,039 GB/s. Compute is identical across all four on NVIDIA's own tables, so the family differs on memory, bandwidth, power and form factor but never on throughput.
Configurations and options
Deployed in conventional PCIe Gen4 servers, typically four to eight cards a node, optionally with NVLink bridges joining pairs at 600 GB/s. Eight cards is 2.0 kW of card load, and NVIDIA's DGX A100 node ratio of 2.03x, used here only as a proxy because the manufacturer publishes no node figure, suggests roughly 4.1 kW, the lowest of any HBM configuration in this category. Multi-instance partitioning divides the card into up to seven 5 GB instances, which is well suited to shared academic and research clusters where many small jobs matter more than one large one. NVIDIA publishes no configurable-TDP range, so 250 W should be treated as fixed.
Compatibility and dependencies
Fits any PCI Express 4.0 x16 slot with a full-height full-length dual-slot space 10.5 inches long and airflow rated for a 250 W passive card. NVLink bridging joins two cards at 600 GB/s, so the interconnect domain is a pair. There is no liquid-cooled variant of this card, unlike the 80 GB PCIe part. At 250 W it is the least demanding HBM part here and the easiest to place. Eight cards draw 2.0 kW and an estimated node lands near 4.1 kW, so a rack of these sits well inside any 26 to 40 kW air envelope. Plan alongside racks, PDUs and CRAC or chiller capacity. The binding constraint in practice is 40 GB of capacity rather than anything in the facility.
Pricing and availability
Published price range: $4,000 to $8,000 per card used or refurbished, three independent channels, Q3 2026. Evidence: channel observation · used or refurbished condition · Single card. Trend: falling A reseller offered one card at $12,500 in May 2020, the month NVIDIA began shipping the A100. Compute Exchange's Q3 2026 band is $4,000 to $6,000 with a High supply signal. Over 6.33 years that is 52 percent to the $6,000 top, 60 percent to a $5,000 midpoint and 68 percent to the $4,000 floor. Compounded: 10.9, 13.5 and 16.5 percent a year. NVIDIA's launch anchor is the DGX A100 at $199,000 on 14 May 2020 for eight 40 GB modules, $24,875 a module on a system basis. Hashrate Index projected on 8 April 2026 a further 10 to 15 percent through 2026, inside the observed rate, which is what makes an A100 residual estimate defensible where a current-generation one is not. Lead time, new: Past its selling life and dropped from NVIDIA's current A100 product page. New cards are not a normal order route; the market is used and refurbished stock.. Lead time, used or refurbished: In stock to a few days from refurbishers, with IT Creations quoting 3 to 4 days on comparable A100 PCIe stock. Supply is rated High by Compute Exchange.. Warranty, new: Not applicable in the current market; this card is bought used or refurbished.. Warranty, used: 3-year reseller warranty on refurbished stock per Network Outlet. No manufacturer warranty transfer was documented anywhere in this category..
Lifecycle and maintenance
This variant carries the earliest end-of-life signal in this category. As of 10 September 2026 it no longer appears on NVIDIA's live A100 product page, which shows only the 80 GB PCIe and 80 GB SXM columns. Its datasheet and product brief PB-10137-001_v03 remain published, so the part is still documented even though it has been dropped from the current page. As the oldest A100 variant, launched September 2020, a fleet card may carry well over 35,000 hours. Operator depreciation practice runs 4 to 6 years against an observed 7.5-year service life on an earlier generation, so this card is at or past its accounting life but inside its usable one.
Common failure points
- HBM2 memory stack — Rising correctable ECC counts, then retired rows, then uncorrectable errors and job crashes under sustained memory pressure
- Passive cooling airflow dependency — Throttling only under full load in a chassis whose fan curve was not qualified for a 250 W passive card
- Thermal interface material — Boost clock no longer reaches 1,410 MHz at constant load as die temperature climbs over months
- PCIe link degradation — Link trains below PCI Express 4.0 x16 or reports corrected-error storms on the host link
- NVLink bridge — A bridged pair reports below 600 GB/s or fails to enumerate as a pair
- Driver and CUDA support horizon — Card functions but a framework release drops Ampere compute capability, and no FP8 path exists for quantised models
Inspection checklist
Driver-reported memory must read 40 GB HBM2: 80 GB indicates the 300 W HBM2e card sold as this part. Stream benchmark bandwidth within 10 percent of 1,555 GB/s: 1,935 GB/s identifies the 80 GB variant instead. Sustained power draw within 5 percent of 250 W under load: 300 W or 400 W identifies a different A100 variant. Clocks reach 765 MHz base and 1,410 MHz boost under load: sustained shortfall indicates thermal or power limiting. GPU SKU reads GA100-883AA-A1: part number must match 900-21001-0000-000 or RH1X7 as observed in channel. HBM2 correctable and uncorrectable ECC counters read before purchase: repeat after a 24-hour soak at full load. Retired-page and row-remap logs pulled before purchase: HBM retirement is the dominant end-of-life signal here. Host PCIe link trains at PCI Express 4.0 x16: negotiating Gen3 or x8 halves host bandwidth.
Procurement channels
Secondary is the whole market. Compute Exchange acts as a broker with a published band and High supply, and refurbishers such as Network Outlet hold stock at $7,800 with a 3-year warranty and an export-restriction notice. Quote-only channels carry inventory without displaying prices, including IT Creations, which quotes 3 to 4 days on comparable A100 PCIe stock, and PC Server and Parts, which prices accelerators through a quote desk. Live eBay listings exist but every figure captured there was an asking price rather than a completed sale.
Regional notes
Export-controlled under NVIDIA's Form 8-K of 17 October 2023, which extends the licence to any incorporating system, for China and Country Groups D1, D4 and D5. Resale is itself an export: a $4,000 card carries the same obligation as a $27,000 one. BIS guidance of 31 May 2026 turns on the counterparty: a licence is required for any entity headquartered in Country Group D:5 or Macau, or with an ultimate parent there, from every destination outside the US, regardless of location. 91 FR 1684 of 15 January 2026 gates case-by-case China and Macau review at both total processing performance under 21,000 and DRAM bandwidth under 6,500 GB/s. Its 1,555 GB/s is far under the bandwidth gate.