When NVIDIA first spoke about chip-based GPUs, we thought it would be Jensen Huang’s going forward 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, the characteristics of which are now known.
First of all, it should once again be specified that to call the AMD Instinct under the name of GPU, Graphics Processor Unit, is at least a contradiction, since despite the fact of deriving from this type of processors in architecture, those by Lisa Su won them all. function units, the possibility of processing lists of graphics commands in the same processor, and also the absence of a display driver. In short, it is not suitable for rendering graphics, but it is useful for performing parallel calculations at high speed.
AMD’s goal with these cards? Competing with the current NVIDIA A100 and Ponte Vecchio in the HPC GPU market, which by the two brands that also make x86 processors are MCM GPUs.
This would be the AMD MI250X
Enough teasing. MI200 has two variants: MI250 and MI250X
MI250X
110 CU, 1.7 GHz boost
128 GB HBM2e
500 W TDP, 7 nm– ExecutableFix (@ExecuFix) 23 October 2021
The information comes to us through the ExecutableFix insider who, through a series of Tweets, has left information about the AMD Intinct MI200 and our first surprise is that the name is not what we are talking about. were expecting, as these will appear in two models and under the AMD Intinct. Name MI250 and MI250X, so the name MI200 is lost.
First of all, we have to keep in mind that this AMD GPU is a MCM module that connects 2 graphics processors under the same intermediary, so we are really talking about 220 computing units, which, considering the Boost speed of 1.7 GHz, gives 13% more frequency compared to the MI100.
383 FP16 / BF16
– ExecutableFix (@ExecuFix) 23 October 2021
Regarding the computing power of the AMD Instinct MI250X, it’s from 47.9 TFLOPS 32-bit floating point
Regarding the power of 384 TFLOPS FP16 / BF16 refers to the Base units of the matrix, the equivalent of 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 NVIDIA GPUs, so we can say that AMD already has a GPU to stand up to Huang when it comes to speeding up AI algorithms and not just for HPC.
You won’t see it on your PC
The fact that the power consumption for the whole GPU is 500W, gives us a clue that this is not a graphics card for a standard PC. It is true that the PCI Express 5.0 interface allows to reach 600 W of consumption, everything indicates that it is the GPU that will integrate one of the supercomputers with which AMD had a contract for its development, so that they can be mounted on an OAM module, very similar to the one in the image above.
We also can’t forget that the GPU when using HBM2E memory and Dual GPU setup will look a lot like the photo above these lines, with its RAM and both GPUs above it. ‘a CoWoS-S interposer from TSMC. All this has a manufacturing cost which 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 logically not its sector.
However, we should not exclude a PCIe version of the AMD Instinct MI250X and MI250, as this 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.