The problem with belittling someone who is willing to work hard and suffer is that you can end up eating your words despite the stress and bullying you put on them. The US has always thought that China would be good at manufacturing in bulk and cheap, and that high-level design in hardware and software was for another more competent country. Well, China today presented its new GPUs from the hand of of twowho have been called BR100 y BR104, with big surprises.
It’s too late to stop China. The US knows it, they are going to dominate the world and we will be the ones who will work for them in not many decades, so they are trying to slow them down as much as possible, but the push of the Americans will not last long. The obvious proof is that despite all the restrictions, China has created graphics cards that are on par with the current NVIDIA Ampere, and they even say that they are faster.
Biren BR100, a monster that competes with AMD
What we are going to see now is not just a presentation of a company towards its products, it is a declaration of intent showing muscle, literally speaking. Biren has set out to take on the Western world just like Innosilicon and for that, to really make noise, you have to launch something very competent.
Well, the first of the GPUs that we will see is the BR100, a technological wonder that is not totally cutting edge, but almost. And we are talking about a graphics card manufactured in 7nm and that has been built with TSMC’s leading process called 2.5D CoWos which we have spoken on occasion. As for transistors, it integrates nothing less than 77 billionwhich gives to offer a quite impressive power of 1.000 TFLOPS en FP16 y 2.000 TFLOPS en FP8.
But why has Biren used the process of TSMC con CoWos? Well, because they have included in the package 64 GB de HBM2Ewhich offers them nothing less than 2,3 TFLOP of bandwidth. We do not know the exact frequencies because they have not been revealed, but they have shown that it integrates a last level cache of 300 MB, really impressive. It should be noted that it integrates PCIe 5.0 and the latest generation interconnection protocol CXL.
A graphics card for PC, what can compete with the RTX 30?
So that we understand a little the design and complexity, the NVIDIA H100, the company’s current top of the range, has only 3 billion more transistors, but it has a much more advanced node (4N). In addition, as they have stated, with such a large number of transistors they can replicate and imitate the nerve cells of the human brain. Logically, it is a GPU dedicated to HPC and AI to unseat NVIDIA from its supercomputers and replace its products with those of Biren, to the point that according to the data provided
Finally, there is a second model called BR104 which is just half as fast as its big brother, where curiously it comes in PCIe format with a TDP of up to 300 watts. It is rumored that this version could compete with a RTX 30 mid-range in performance, so the potential is really interesting. In any case and knowing that the BR100 is not faster than the GPU Hopper H100 and for a long time, China begins to surround the US, to be independent in more or less cutting-edge technology despite the restrictions, it is a matter of time before it equals the power of Western hardware and the optimization of our software, and being cheaper.