When we saw the release of the NVIDIA A100 and the subsequent PCIe version, we were expecting NVIDIA to fill in its lineup earlier rather than later with its Ampere generation. We were slightly surprised that NVIDIA chose to launch the consumer NVIDIA GeForce RTX 3000 Series Launch before Quadro and other data center GPUs. Now, NVIDIA is starting to fill out the higher-value portions of its portfolio. The first entry is the NVIDIA A40 GPU.
NVIDIA A40 GPU
The NVIDIA A40 GPU is a rung lower than the NVIDIA A100, but it offers quite a few interesting features. For example, it uses GDDR6 ECC memory instead of HBM2. It supports NVLINK but is a PCIe Gen4 device. We will also note that we see a 300W passively cooled GPU instead of active cooling like we see on the NVIDIA RTX A6000 that is also being launched as part of today’s announcements.
|Die Size||628.4 mm2|
|CUDA Parallel Processing Cores||10,752|
|NVIDIA Tensor Cores (3rd Gen)||336|
|NVIDIA RT Cores (2nd Gen)||84|
|GPU Memory||48 GB GDDR6 with ECC|
|Memory Bandwidth||696 GB/s|
|NVLink||2-way low profile (2-slot)|
|NVLink Interconnect||NVIDIA NVLink 112.5 GB/s (bidirectional)
PCIe Gen4 16 GB/s
|Max Power Consumption||300W|
|Graphics Bus||PCI Express 4.0 x16|
|Display Connectors||DP 1.4 (3)|
|Form Factor||4.4” H x 10.5” L Dual Slot|
|vGPU Software Support||NVIDIA GRID®, NVIDIA Quadro® Virtual Data Center Workstation, NVIDIA Virtual Compute Server*|
|vGPU Profiles Supported||1 GB, 2 GB, 3 GB, 4 GB, 6 GB, 8 GB, 12 GB, 16
GB, 24 GB, 48 GB
|NVIDIA® 3D Vision® and 3D Vision Pro||Support via 3 pin mini DIN|
|Frame lock||Compatible (with Quadro Sync II)|
|Power Connector||1x 8-pin CPU|
|NVENC | NVDEC||1x ENC | 2x DEC (includes AV1 decode)|
|NEBS Ready||Level 3|
|Secure and Measured Boot with HW Root of
Something else that is interesting in this card is that it has three DisplayPort ports. Since this is a data center focused GPU, the vGPU feature cannot be used if those DP ports are active. As a result, they are disabled by default.
Another interesting note is that these cards can use vGPU to split the GPU resources for applications such as VDI. They do not support the MIG feature found on the NVIDIA A100 to split the GPU.
Seeing more Ampere is a great sign. Something else we did notice is that this is a 300W GPU with 696GB/s of memory bandwidth. While the compute units on the NVIDIA RTX A6000 are similar, as is the 48GB of memory, the RTX A6000 has 768GB/s of memory bandwidth. If you are looking for what is effectively the actively cooled version of this card, take a look at that article.