Power and low consumption: NVIDIA RTX A4500 and A2000 server GPUs

From the launch of its Turing architecture to the market, known as GeForce RTX 20 in the gaming market, the Quadro brand was replaced by a new one, that of RTX Axxx. Well, without prior notice NVIDIA has launched new professional GPUs, A4500 and A2000, this time based on its Ampere architecture, the same as the RTX 30. What specifications and how do they differ from the GeForce for gaming?

Graphics cards are used in various markets and although the best known is PC gaming, they are used in other markets because they are processing units that are not only used to generate graphics. As is the case with scientific computing and lately artificial intelligence. Although it is not the case of the RTX A4500 and RTX A2000, since they are designed for the professional graphics market.

RTX A4500 Specifications

Architecture Ampere
Lithographic process Samsung 8 nm
Die size 628 mm2
Transistors 23.8 billion
SM / CU 56
CUDA cores (FP32) 7,168
TMU 224
ROPs 112
RT Cores 56
Tensor Cores 224
L2 cache 6 MB
Base Clock 1,656 MHz
VRAM Clock 16 Gbps
Bus 320 bits
VRAM bandwidth 640 GB / s
FP32 performance 23.7 TFLOPS
PCIe version 4.0 x16
TDP 200 W
Feeding 8 pins
NVLink / SLI / Crossfire Yes
Video outputs 4 x DisplayPort 1.4

The RTX A4500 uses the same GPU as the RTX 3080 (Ti) and RTX 3090, that is, the GA102, but because it is designed to work in data centers to provide cloud computing services its specifications are different, since As with server CPUs that need to work 24 hours a day and 7 days a week, that is why its specifications are somewhat lower than graphics cards for gaming.

This professional graphics card consumes 200 W, a very low figure for one based on the GA102 GPU, although it really is not exactly the same chip as in the hardware that is sold for gaming, since it has some differences that are key for the market of workstations and centers of data and that we tell you some sections below.

Your exact specifications? Have 7168 ALU on FP32 known in NVIDIA hardware as CUDA cores, since the Ampere architecture is based this are 56 SM that inside include the same number of RT Cores and 224 Tensor Cores. Its clock speed and power? 1,656 MHz speed giving a rate of 23, 7 TFLOPS single precision floating point, 32 bit, and 189.2 TFLOPS 16-bit floating point across the Tensor Cores. Refering to VRAM have 20 GB on x8 mode configured on a bus 320 bit GDDR6 at 16 Gbps for a bandwidth of 640 GB / s.

Each of the RTX A4500s connects to the workstation to the data center through a PCI Express 4.0 interface, but we can interconnect two through the NVLink interface and a bridge to work together.

RTX A2000 Specifications

Architecture Ampere
Lithographic process Samsung 8 nm
Die size 276 mm2
Transistors 13,250 million
SM / CU 26
CUDA cores (FP32) 3,328
TMU 104
ROPs 48
RT Cores 26
Tensor Cores 104
L2 cache 3 MB
Base Clock 1,200 MHz
VRAM Clock 12 Gbps
Bus 192 bits
VRAM bandwidth 288 GB / s
FP32 performance 8 TFLOPS
PCIe version 4.0 x16
TDP 70 W
Feeding No
NVLink / SLI / Crossfire No
Video outputs 4 x DisplayPort 1.4

The second of the professional GPUs just released by NVIDIA is the RTX A2000, which is more modest than his la RTX A4500as it is based on the GA106 GPU, the same as the RTX 3060 gaming graphics cards, however, it has been configured to operate in a data center or workstation like the rest of the RTX A family.

We find a graphics card with a very low consumption, of only 70 W, so it works with the power provided by the port itself and does not require an external connector. This is offered in two different versions, one with 6 GB of GDDR6 VRAM and another with 12 GB; in both versions the VRAM runs at 12 Gbps, which gives a bandwidth of 288 GB / s. Regarding the changes made to the professional variant of the GA106 are the same as the GA102 of the RTX A4500.

Although what really interests us are its specifications and in that case we have 3328 CUDA cores configured in 26 SM that together contain 26 RT Cores and 104 Tensor Cores that give it a power of 8 TFLOPS in FP32 and 63.9 TFLOPS in FP16 a through the Tensor Cores. Unlike the RTX A4500, this model lacks an NVLink interface and therefore we cannot tandem two graphics cards.

Differences with the RTX 30

Although NVIDIA announces these graphics cards with the same GPUs that we can find in the NVIDIA GeForce RTX 30, this is not really the case, since at the hardware level there are important changes that require a completely new circuitry and cannot be achieved only through the drivers.

  • The first change is in the command processor found in the central part of the GPU, in these cases they have been designed to handle virtualized multiple screen lists that are independent of each other. This is key for remote or cloud computing, as it allows you to divide your power among multiple clients. The RTX 30 for gaming can only handle two display lists, but no more.
  • The second element is in the local memory of the card, that is, the VRAM. In the case of the RTX A4500 GDDR6X is not used, but GDDR6 in order to save consumption, however, the GDDR6X does not support bug fix or ECC while the GDDR6 yes, something key in server environment. Such a feature is also found on the RTX A200.
  • Third, we have its video output and it is that although both support HDCP 2.2, however there are four DisplayPort 1.4 connectors with audio and do not use any HDMI output.
  • Nor can we forget the fact that they have a Most advanced NVENC which enables encode multiple streams video at the same time and is therefore not limited like the hardware video codec version for GeForce.
  • The last change has to do with the fact that they have to function in a data center at all hours without problems, that is why lack clock speed spike mechanisms, what we know as Boost speed.

Although the biggest difference is in the price, which in this range of graphics cards is much higher than a conventional GeForce, so much so that it is not even profitable even for cryptocurrency mining, even taking into account its power per watt for its low consumption, although at the moment we do not know it.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *