NVIDIA A40 48GB Specs, Benchmarks & Pricing
The NVIDIA A40, launched in October 2020 with part number 900-2G133-0000-000, is a professional-grade GPU built on the GA102 processor. It is tailored for high-end machine learning applications, offering exceptional performance in both AI training and inference tasks. The A40's introduction marks a significant milestone in GPU technology, showcasing advancements in memory capacity, bandwidth, and processing power, suitable for the most demanding AI workloads.
- Release Date: October 5, 2020
- MSRP: $27,500 USD†
- GPU Architecture: Ampere
- Hardware-Accelerated GEMM Operations:FP64 FP32 TF32 BF16 FP16 FP8 FP4 INT8 INT4
- CUDA Compute Capability : 8.6
† Estimated MSRP of $27,500 USD for the NVIDIA A40 based on OEM list prices. NVIDIA does not publish official MSRP for data center GPUs. OEM list prices corroborated by 2 independent sources: Cisco UCSC-GPU-A40 has list price of $27,561.22 per Hummingbird Networks, and Lenovo ThinkSystem A40 (4X67A72593) is listed at €26,526.32 (~$28,500 USD) per Redcorp. Note: Current market/resale prices (~$5,500) are significantly lower due to depreciation but do not reflect original launch list pricing.
Strengths
- Excellent FP32 compute performance (top 24% of GPUs)
- Excellent memory capacity (top 14% of GPUs)
Specifications for NVIDIA A40
| Specification | Performance Ranking |
|---|---|
| FP32 TFLOPs | |
| FP16 TFLOPs | |
| Tensor Core Count | |
| Memory Capacity (GB) | |
| Memory Bandwidth (GB/s) | |
| Int8 TOPs | |
| FP8 TFLOPs | Data not available |
| FP4 TFLOPs | Data not available |
Real-time NVIDIA A40 GPU Prices
Compare Price/Performance to other GPUs
Compare NVIDIA A40 to Another GPU
Price History
NVIDIA A40 Price History
This chart tracks NVIDIA A40 prices over the past 6 months based on eBay listings. The yellow line shows the average of the three lowest-priced listings each day—a useful indicator for finding deals.
Trend: Over this period, the lowest average price has been down 26.8%. The current lowest average is $3,683, compared to a typical $4,454 over the period. The lowest average price is calculated as the average of the three lowest-priced listings each day.
Product Identifiers
Manufacturer Part Numbers (4)
- NVIDIA Part Number
- 900-2G133-0000-000
- NVIDIA Part Number
- 900-2G133-0000-100
- NVIDIA Part Number
- 900-2G133-0100-030
- Board ID
- PG133C
Available from 6 Partners (11 products)
- Nvidia A40 48GB x16 PCI-e 300W DW FH/HL GPU
- W6GX6(part number)
- Dell Nvidia A40 GPU Accelerator 48GB
- M9WT5(part number)
- NVIDIA Ampere A40, PCIe, 300W, 48GB Passive, Double Wide, Full Height GPU
- 490-BGKE(part number)
- NVIDIA Ampere A40 GPU Accelerator
- 490-BGSJ(part number)
- Dell NVIDIA A40 GPU Accelerator
- 490-BGKG(part number)
- Dell NVIDIA A40 GPU Accelerator
- 699-2G133-0200-C00(part number)
- NVIDIA A40 48GB PCIe Non-CEC Accelerator for HPE
- R9S37C(part number)
- ThinkSystem NVIDIA A40 48GB PCIe 4.0 Passive GPU w/o CEC
- 4X67A72593(part number)
- PNY NVIDIA A40 48GB GDDR6 ECC Data Centre GPU
- TCSA40M-PB(sku)
- Cisco NVIDIA Tesla A40 Graphic Card 48GB
- UCSC-GPU-A40(part number)
- Tesla A40 48GB PCIe DSFH w/ATX BKT
- SKY-TESL-A40-48P(part number)
References
- https://www.nvidia.com/en-us/data-center/a40/
- https://www.techpowerup.com/gpu-specs/a40-pcie.c3700
- https://www.nvidia.com/en-us/data-center/tensor-cores/
- https://www.nextplatform.com/2021/04/15/nvidia-rounds-out-ampere-lineup-with-two-new-accelerators/
- https://www.hummingbirdnetworks.com/nvidia-tesla-a40-rtx-gpu-computing-processor-tesla-a40-rtx-48-gb-for-p-n-ucsc-c240-m6l-ch-ucsc-c240-m6s-ch-ucsc-c240-m6sn-ucsc-c240-m6sx-rf-ucsc-gpu-a40
- https://www.redcorp.com/en/product/servers-miscellaneous-accessories/lenovo/thinksystem-nvidia-a40-48gb-pci-e-gen4-passive-gpu-4x67a72593/m852qy81
- https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/a40/NVIDIA%20A40%20Product%20Brief.pdf
- https://pcserverandparts.com/nvidia-a40-gpu-accelerator-48gb-gddr6-tesla-ampere-used-pg133c-900-2g133-0000-000-699-2g133-0200-c00/
- https://www.acmemicro.com/Product/17961/NVIDIA-900-2G133-0000-000-Ampere-A40-48GB-GDDR6-w-ECC---Dual-Slot---PCIe-4-0-Graphic-Card----Passive-Cooling
- https://www.dell.com/en-us/shop/nvidia-ampere-a40-pcie-300w-48gb-passive-double-wide-full-height-gpu/apd/490-bgke/graphic-video-cards
- https://www.expresscomputersystems.com/products/nvidia-a40-48gb-gpu-noncec-accelerator-r9s37c
- https://buy.hpe.com/us/en/options/server-accelerators/computational-graphics-accelerators-for-servers/computational-graphics-accelerators-for-proliant-servers/nvidia-a40-48gb-pcie-non-cec-accelerator-for-hpe/p/r9s37c
- https://lenovopress.lenovo.com/lp1773-thinksystem-nvidia-a40-48gb-pcie-gen4-passive-gpu
- https://www.pny.com/file%20library/company/support/product%20brochures/nvidia%20data%20center%20gpus/nvidia-a40-datasheet.pdf
- https://www.scan.co.uk/products/48gb-pny-nvidia-telsa-a40-card-pcie-40-x16-ampere-10752-core-gddr6-w-ecc-passive
- https://www.expresscomputersystems.com/products/cisco-nvidia-tesla-a40-dw-ucsc-gpu-a40
- https://www.advantech.com/en-us/products/8d6a8886-b57e-493f-87c8-6f1e02d0608a/sky-tesl-a40-48p/mod_458ee4c5-1c38-4ed7-80e6-b246c54d373f
- https://www.ebay.com/p/17053130735
- https://forums.developer.nvidia.com/t/what-is-the-fp64-rpeak-value-for-a5000/196056
- https://images.nvidia.com/content/Solutions/data-center/a40/nvidia-a40-datasheet.pdf
Notes
- Estimated MSRP of $27,500 USD for the NVIDIA A40 based on OEM list prices. NVIDIA does not publish official MSRP for data center GPUs. OEM list prices corroborated by 2 independent sources: Cisco UCSC-GPU-A40 has list price of $27,561.22 per Hummingbird Networks, and Lenovo ThinkSystem A40 (4X67A72593) is listed at €26,526.32 (~$28,500 USD) per Redcorp. Note: Current market/resale prices (~$5,500) are significantly lower due to depreciation but do not reflect original launch list pricing.
- NVIDIA part number 900-2G133-0000-000 is the primary NVPN. Sourced from multiple vendor listings including acmemicro.com, serversupply.com, and pcserverandparts.com
- NVIDIA part number 900-2G133-0000-100 sourced from wiredzone.com (Leadtek variant)
- NVIDIA part number 900-2G133-0100-030 sourced from eBay listing for Dell M9WT5 variant
- Board ID PG133C sourced from multiple vendor listings including pcserverandparts.com, ebay.com, and harddiskdirect.com
- Dell part number W6GX6 sourced from expresscomputersystems.com
- Dell part number M9WT5 sourced from expresscomputersystems.com and eBay listings
- Dell part number 490-BGKE sourced from Dell official website (dell.com) and multiple reseller listings
- Dell part number 490-BGSJ sourced from itcreations.com
- Dell part number 490-BGKG sourced from itcreations.com
- Dell OEM part number 699-2G133-0200-C00 sourced from pcserverandparts.com and marigoldsystems.com
- HPE part number R9S37C sourced from HPE official store (buy.hpe.com) and expresscomputersystems.com
- Lenovo part number 4X67A72593 sourced from Lenovo Press Product Guide (lenovopress.lenovo.com/lp1773) and redcorp.com
- PNY SKU TCSA40M-PB sourced from PNY datasheet, scan.co.uk, and novatech.co.uk
- Cisco part number UCSC-GPU-A40 sourced from Cisco official documentation and expresscomputersystems.com
- Advantech part number SKY-TESL-A40-48P sourced from Advantech official website (advantech.com) and westercom.eu
- Phase 5 audit (2026-09-02): gpuArchitecture normalized from lowercase 'ampere' to canonical 'Ampere' capitalization per the canonical table in .claude/agents/gpu-researcher.md.
- Phase 5 audit (2026-09-02): FP64 removed from supportedHardwareOperations. The A40 uses the GA102 die (confirmed in the summary above: "built on the GA102 processor"), not GA100. Per an NVIDIA staff reply on NVIDIA Developer Forums, 'The Tensor core units in the GA102 chip architecture do not contain the FP64 path that is present in e.g. the Tensor core units in the A100 GPU' (https://forums.developer.nvidia.com/t/what-is-the-fp64-rpeak-value-for-a5000/196056). Canonical Ampere baseline (TF32, BF16, FP16, INT8, INT4) applied per gpu-researcher.md; FP64 Tensor Core support is reserved for GA100-based SKUs (A100, A30) only.
- fp8TFLOPS and fp4TFLOPS are left null (field omitted) and not added to unverifiedSpecs — this is a confirmed hardware negative, not a research gap. The A40's Ampere 3rd-generation Tensor Cores predate both FP8 (introduced in Ada Lovelace/Hopper) and FP4 (introduced in Blackwell) tensor formats, per the canonical architecture table in .claude/agents/gpu-researcher.md.
- fp32TFLOPS of 74.8 is TF32 Tensor Core throughput without sparsity, per the NVIDIA A40 datasheet (nvidia-a40-datasheet.pdf). FP32 vector (CUDA core) throughput is 37.4 TFLOPS.
- fp16TFLOPS of 299.4 is FP16 Tensor Core throughput with 2:4 structured sparsity per the NVIDIA A40 datasheet. Dense FP16 Tensor is 149.7 TFLOPS.