GPUDatacenterReleased Oct 5, 2020

NVIDIA A40

48 GB GDDR6 ECC · 384-bit · 10,752 cores

A passively cooled, dual-slot Ampere data center GPU for visual computing: virtual workstations, rendering, simulation and media work. It is the server counterpart of the RTX A6000 and keeps display outputs and the full RTX graphics feature set alongside third-generation Tensor Cores, supports the NVIDIA vGPU software stack, and two cards can be joined over NVLink. Announced at GTC in October 2020.

General

Manufacturer: NVIDIA
Model: A40
Segment: Datacenter
GPU chip: GA102
Architecture: Ampere
Release date: Oct 5, 2020
Price: Search on Amazon

Memory

Memory size: 48 GB
Memory type: GDDR6 ECC
Bus width: 384-bit
Bandwidth: 696 GB/s
Memory clock: 1,812 MHz

Clocks & silicon

GPU clock: 1,305 MHz
Boost clock: 1,740 MHz
Shader units: 10,752 CUDA Cores
Tensor / AI cores: 336
RT cores: 84
Transistors: 28.3 billion
Technology (nm): 8 nm
Samsung 8N (8 nm)
TDP / TBP: 300 W

Interface & APIs

Bus interface: PCIe 4.0 x16
SLI / multi-GPU: NVLink (2-way bridge, 112.5 GB/s)
DirectX: 12 Ultimate (12_2)
OpenGL: 4.6
Vulkan: 1.3
Other APIs: CUDA 8.6, OpenCL 3.0

AI support

Precisions: FP8FP16BF16INT8INT4TF32FP64
AI accelerator: 3rd Gen Tensor Cores (336) + 2nd Gen RT Cores (84)
Peak AI TOPS: 598.6 TOPS
Notes: Third-generation Tensor Cores; FP8 is not supported on Ampere. Peak throughput with structured sparsity: INT4 1,197.4 TOPS, INT8 598.6 TOPS, FP16 and BFLOAT16 299.4 TFLOPS, TF32 149.6 TFLOPS; dense rates are half. FP32 is 37.4 TFLOPS and RT Core performance 73.1 TFLOPS. One encoder and two decoders, including AV1 decode.