The AI boom has meant that China does not want to depend on American companies and develop their own GPUs. The last one to join is called Metax and just introducedr your first graphics card with an AI-optimized GPU of the generative type. which they have baptized as MXN100. What are its specifications and features? Let’s see
Until recently, graphics cards in servers were seen only in scientific institutions and large render farms for movies with computer-generated graphics. However, for some time now there are two tasks that justify its existence: generative AI and video transcoding, and obviously China does not want to be left behind.
Metax MXN100, a graphics card for generative AI in the cloud made in China
The Metax graphics card MXN100 It is not designed for the home PC market, but for data centers looking for a GPU that is compact in size and low in power. And it is that this hardware has the particularity of taking up very little space, since it occupies a single slot and uses completely passive cooling. Although the manufacturer has not officially revealed it, these two points indicate that we are dealing with a low-end GPU with low energy consumption. As for its video memory, it is of the type HBM2Ebut we do not know at the moment the size of the bus, at least officially.
However, by visiting the net we have been able to find a promotional render of the MXN100 without its encapsulation and we have been able to see how it is uses a single HBM2E memory stack, which in terms of bandwidth is equivalent to a 128-bit GDDR6 bus. In any case, it does not seem that this chip is optimized for games, but rather for the AI and if we look at its specifications, we have that its capacity is 80 TFLOPS in FP16 and 160 TOPS in Int8a figure that may seem very high if we think that they come from SIMD units, however, it is the unit performance for AI and you have to put things in their real context.
If we make a comparison with the NVIDIA RTX we will find that the RTX 3060 is rated at 102 TFLOPS at FP16 from its Tensor Coresso Metax’s proposal in raw power would be a little below, but above the 52 Tensor TFLOPS of the RTX 2060.
Is its performance disappointing?
We have to assume that this graphics card is not going to be sold to the public, but to be mounted on servers where performance per watt is important and they will be running permanently. This makes its low consumption very attractive, despite the fact that its gross power is lower than in other solutions on the market. But let’s not forget that each additional MHz after a certain point has an increasingly higher cost.
We say this due to the fact that we have already seen many attempts by China to gain a foothold in the graphics card market, with attempts that often turn out to be very poor. However, this is not the case at hand. We do not know what the TDP of the Metax MXN100 isbut we don’t think it’s more than 75 W it gives the PCI Express port on its own without external connectors. Although it does not have the power of the RTX 40, we do not doubt that its performance per watt is competitive.
In computing and artificial intelligence, having several graphics cards working in parallel as long as the energy budget is met is not a bad option. The key is that if the performance you get is much better under certain cooling and consumption conditions than a powerful NVIDIA RTX.