Alongside a slew of software-related announcements this morning from NVIDIA as part of their fall GTC, the company has also quietly announced a new server GPU product for the accelerator market: the NVIDIA A2. The new low-end member of the Ampere-based A-series accelerator family is designed for entry-level inference tasks, and thanks to its relatively small size and low power consumption, is also being aimed at edge computing scenarios as well.

Along with serving as the low-end entry point into NVIDIA’s GPU accelerator product stack, the A2 seems intended to largely replace what was the last remaining member of NVIDIA’s previous generation cards, the T4. Though a bit of a higher-end card, the T4 was designed for many of the same inference workloads, and came in the same HHHL single-slot form factor. So the release of the A2 finishes the Ampere-ficiation of NVIDIA accelerator lineup, giving NVIDIA’s server customers a fresh entry-level card.

NVIDIA ML Accelerator Specification Comparison
  A100 A30 A2
FP32 CUDA Cores 6912 3584 1280
Tensor Cores 432 224 40
Boost Clock 1.41GHz 1.44GHz 1.77GHz
Memory Clock 3.2Gbps HBM2e 2.4Gbps HBM2 12.5Gbps GDDR6
Memory Bus Width 5120-bit 3072-bit 128-bit
Memory Bandwidth 2.0TB/sec 933GB/sec 200GB/sec
VRAM 80GB 24GB 16GB
Single Precision 19.5 TFLOPS 10.3 TFLOPS 4.5 TFLOPS
Double Precision 9.7 TFLOPS 5.2 TFLOPS 0.14 TFLOPS
INT8 Tensor 624 TOPS 330 TOPS 36 TOPS
FP16 Tensor 312 TFLOPS 165 TFLOPS 18 TFLOPS
TF32 Tensor 156 TFLOPS 82 TFLOPS 9 TFLOPS
Interconnect NVLink 3
12 Links
PCIe 4.0 x16 +
NVLink 3 (4 Links)
PCIe 4.0 x8
GPU GA100 GA100 GA107
Transistor Count 54.2B 54.2B ?
TDP 400W 165W 40W-60W
Manufacturing Process TSMC 7N TSMC 7N Samsung 8nm
Form Factor SXM4 SXM4 HHHL-SS PCIe
Architecture Ampere Ampere Ampere

Going by NVIDIA’s official specifications, the A2 appears to be using a heavily cut-down version of their low-end GA107 GPU. With only 1280 CUDA cores (and 40 tensor cores), the A2 is only using about half of GA107’s capacity. But this is consistent with the size and power-optimized goal of the card. A2 only draws 60W out of the box, and can be configured to drop down even further, to 42W.

Compared to its compute cores, NVIDIA is keeping GA107’s full memory bus for the A2 card. The 128-bit memory bus is paired with 16GB of GDDR6, which is clocked at a slightly unusual 12.5Gbps. This works out to a flat 200GB/second of memory bandwidth, so it would seem someone really wanted to have a nice, round number there.

Otherwise, as previously mentioned, this is a PCIe card in a half height, half-length, single-slot (HHHL-SS) form factor. And like all of NVIDIA’s server cards, A2 is passively cooled, relying on airflow from the host chassis. Speaking of the host, GA107 only offers 8 PCIe lanes, so the card gets a PCIe 4.0 x8 connection back to its host CPU.

Wrapping things up, according to NVIDIA the A2 is available immediately. NVIDIA does not provide public pricing for its server cards, but the new accelerator should be available through NVIDIA’s regular OEM partners.

Source: NVIDIA

Comments Locked

16 Comments

View All Comments

  • sabot00 - Tuesday, November 9, 2021 - link

    Can this play games if one virtualizes the card (to get around no display out).
  • RU482 - Tuesday, November 9, 2021 - link

    I doubt you'd be blown away by the performance if it could
  • mode_13h - Tuesday, November 9, 2021 - link

    Yes, quite likely. Given that it's based on a GA107, performance would be similar to a RTX 3050.
  • Spunjji - Wednesday, November 10, 2021 - link

    About 2/3 of an RTX 3050 when you balance out the clock speed and core count differences.
  • catavalon21 - Tuesday, November 9, 2021 - link

    I'm trying to figure out the purpose of this card. With a 1/32 Double Precision ratio and a core count not seen so small in accelerators since Maxwell - what niche is this filling? As a T4 successor it falls a bit short in comparison. Less memory bandwidth. Fewer cores. Less compute capability. Just...why?
  • mode_13h - Tuesday, November 9, 2021 - link

    As they say, its niche is inference. For that, you mostly need int8 and BF16.

    Compared to their other Tesla products, what it gives you is a GPU that can run with no aux power and has the same CUDA capability as the rest of their Ampere line. That includes tensor cores with BF16, which I think T4 is lacking.
  • mode_13h - Tuesday, November 9, 2021 - link

    > Less compute capability.

    Uh, those words have a very specific meaning, in Nvidia/CUDA context. What I think you mean is less compute capacity or performance.

    https://en.wikipedia.org/wiki/CUDA#Version_feature...
  • catavalon21 - Thursday, November 11, 2021 - link

    point taken
  • Gigaplex - Tuesday, November 9, 2021 - link

    Replacement for the low end Quadro line, often used in workstation grade office boxes that don't need much GPU power, but need something with certified drivers (Intel support is garbage).
  • Gigaplex - Tuesday, November 9, 2021 - link

    Scratch that, I didn't see that this line doesn't have the display outputs. Stupid NVIDIA claiming that Quadro will be replaced by the A model line and then mixing it up with stuff like this.

Log in

Don't have an account? Sign up now