Computer

AMD’s multi-chip GPU – 4x faster with 128GB VRAM

When NVIDIA first talked about chiplet-based GPUs, we thought it was going to be Jensen Huang’s that would get ahead with the first design on the market, but while Hopper is nothing more than a promise on paper, the red team already has its first GPU of its kind, the AMD Instinct MI250X, of which we now know its specifications.

First of all, we must clarify again that calling the AMD Instinct under the name of GPU, Graphics Processor Unit, is at least a contradiction, since despite deriving from this type of processors in architecture those of Lisa Su have taken them away all fixed function units, the ability to process graphical command lists on the ditto processor and also lack a display driver. In short, it is not useful for rendering graphics, but it is for performing parallel calculations at high speed.

AMD’s goal with these cards? To compete against the current NVIDIA A100 and against Ponte Vecchio in the HPC GPU market, which by the two brands that also makes x86 CPUs is in MCM GPUs.

This would be the AMD MI250X

The information comes to us through the ExecutableFix insider who through a series of Tweets has been leaving information about the AMD Intinct MI200 and our first surprise is that the name is not what we expected, since these will appear in two models and under the AMD Intinct MI250 and MI250X name, so the MI200 name is lost.

First of all, we have to bear in mind that this AMD GPU is a MCM module linking 2 graphics processors under the same interposer, so we are really talking about 220 Compute Units, which taking into account the Boost 1.7 GHz speed results in 13% more frequency when compared to the MI100.

Regarding the computing power of the AMD Instinct MI250X, this is from 47.9 TFLOPS in 32-bit floating point, but what really matters in this type of GPU is FP64 computing where we should expect the same (or at least similar) performance. So AMD would have stopped using the SIMD on registration. Which, combined with the use of a Dual GPU, gives it four times the power compared to the AMD Instinct MI100.

Regarding the power of 384 TFLOPS FP16 / BF16 refers to the Matrix Core Units, the equivalent to NVIDIA Tensor Cores, the power ratio between the FP32 ALUs in the SIMD units and the Tensor units of the Compute Unit is the same as in the NVIDIA GPUs, so it can be said that AMD already has a GPU to stand up to Huang’s when it comes to speeding up AI algorithms and not just for HPC.

You will not see it on your PC

Intel-Nervana-NNP-L-1000-OAM-Cover

The fact that the power consumption for the entire GPU is 500W, gives us a hint that it is not a graphics card for a standard PC. It is true that the PCI Express 5.0 interface allows reaching 600 W of consumption, everything indicates that this is the GPU that will integrate one of the supercomputers with which AMD had a contract for its development, so they could be mounted on a OAM module, very similar to the one in the image above.

CoWoS Roadmap

We can also not forget that the GPU when using HBM2E memory and a Dual GPU configuration will look very similar to the photo above these lines, with its RAM memory and the two GPUs on top of a CoWoS-S interposer of TSMC. All this has a manufacturing cost that makes these integrated circuits prohibitive for the domestic market. In any case, the AMD Instinct MI250X cannot generate graphics of any kind, so home PCs are not its sector logically.

However, we must not rule out a PCIe version of the AMD Instinct MI250X and MI250, since it is the form factor of the current MI100, but everything indicates that the version for the El Capitan supercomputer will be OAM and surely there they can be seen for the first time.

Related Articles

Leave a Reply

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