NVIDIA A40 48GB GDDR6 PCIe GPU – 900-2G133-0000-000 – Refurbished

The NVIDIA A40 48GB (NVIDIA part number 900-2G133-0000-000) is a passively cooled, dual-slot PCIe Gen4 data-center GPU with 48GB of GDDR6 memory. It is one of the most cost-effective ways to run 70B-parameter language models at 4-bit on a single card.

Condition: Refurbished – tested before shipping.

Specifications

• NVIDIA part number: 900-2G133-0000-000
• Architecture: NVIDIA Ampere
• Memory: 48GB GDDR6 with ECC, 384-bit, 696 GB/s
• CUDA cores: 10,752
• Interface: PCIe 4.0 x16
• Max power: 300W
• Power input: 1x 8-pin (EPS/CPU-style, not a standard PCIe graphics connector)
• Cooling: Passive – requires server chassis airflow
• Form factor: Dual-slot, full height, full length
• NVLink: 2-way bridge support

Also sold as

The same GPU appears under OEM part numbers, including Dell 490-BGKE / M9WT5, HPE R7E31C and Lenovo 4X67A72593. Board number 699-2G133-0200-C00 (PG133C).

Power cable by server

The A40 takes an EPS-style 8-pin input, so the server-side GPU cable matters:

• HPE ProLiant DL380 / DL385 Gen10: the 871829-B21 keyed kit, or its 869820-001 cable on its own ($49)
• Dell PowerEdge R750 / R7525: Dell's 12-pin to 8-pin GPU cable (T34N0)
• Dell PowerEdge R740 / R740xd: Dell's 4VPD3 EPS-style GPU cable ($47)

Every server also needs its GPU enablement parts (riser, shroud, high-performance fans). Not sure? Send us your server model and we'll confirm what you need.

Want it already installed?

Our Dell R740xd A40 build ships with the card, cable and enablement kit installed and tested. See what a 48GB A40 can run for measured model sizes and tokens per second, and why a GPU won't work in a server for the eight install problems to check first.

FAQ

What power cable does the NVIDIA A40 use?
One 8-pin EPS-style (CPU-type) input. A standard PCIe 8-pin graphics cable will not fit. Use the server vendor's GPU cable: HPE 869820-001 for DL380 Gen10, Dell 4VPD3 for R740/R740xd ($47 at Resilient Tec), Dell T34N0 for R750/R7525.

Can the A40 run in a desktop or workstation?
Not without dedicated airflow. It is passively cooled and depends on server fans pushing air through the heatsink.

What is the largest model it can run?
A 70B-parameter model at 4-bit quantization fits in its 48GB. The 27–34B class at 4-bit is the sweet spot for speed.

What condition is it in?
Refurbished and tested before shipping, backed by the Resilient Tec lifetime warranty.