Today we are going to take a quick look at the NVIDIA A40 GPU. The A40 is a 48GB GPU that is passively cooled, yet also has video outputs making it a very interesting card. Furthermore, it has a 300W TDP making it a PCIe Gen4 GPU with a relatively high TDP that can be used in professional workstations and servers. We have already seen the A40 in a number of STH system reviews, but today, we decided it was time to give the unit its own piece.
NVIDIA A40 48GB GPU Hardware Overview
The A40 card itself is a double-width full-height full-length PCIe Gen4 GPU. This uses NVIDIA’s gold theme and looks similar to many of the other NVIDIA PCIe GPUs of this generation that we have used.
As we expect, this is a newer PCIe Gen4 generation GPU, so it is likely to be paired with 3rd Generation Intel Xeon Scalable “Ice Lake” or AMD EPYC 7002 “Rome” or EPYC 7003 “Milan” CPUs. We still get a standard PCIe Gen4 x16 edge connector on the card.
Cooling on the GPU is passive, but NVIDIA says that the card is able to handle airflow in either direction so air can be pushed or pulled through the heatsink. As some frame of reference, the GPU with its PCIe I/O plate is around 1kg/ 2.2lbs so the heatsink has significant heft. We also get a standard 8-pin data center GPU power connector on the rear of the unit. That section also includes holes for vendor-specific mounting support brackets.
On the top of the unit, we get something that is fairly interesting in that we get a NVLINK connector. We did not have the bridge for this one when we had multiple A40’s but the feature is present.
On the I/O faceplate of the unit, we get three display outputs. These outputs are actually disabled by default. NVIDIA does this so SR-IOV support works out-of-the-box for NVIDIA Virtual GPU software, such as the editions used for VDI. One can also enable the display outputs for other software configurations and can also span multiple A40 cards for display using Quadro Sync II.
This is a big deal since we can get SR-IOV on this GPU, but we do not get NVIDIA MIG, or Multi-Instance GPU on this card as we would on something like the A100. We have looked at MIG a number of times in NVIDIA A100 system reviews. Here is what a dense configuration of 8x NVIDIA A40’s looks like:
The Ampere-generation GPU itself is a GA102-895 with a 1305MHz base and a 1740MHz boost clock and 10,752 CUDA cores. The 48GB of memory is GDDR6 at 7250MHz and with a 384-bit wide bus width for a total of 696GB/s of peak memory bandwidth. We also get ECC memory in this class of GPU. Here are the full specs:
Next, we are going to take a quick look at performance before we move on to power consumption and topology.