While waiting for Intel to confirm the launch of its ARC Alchemist for the Spanish market, we see that we could see a fourth competitor in the market and it would not come from the United States as is the case with NVIDIA, AMD and Intel, but the other superpower is China and no, we’re not talking about the Fantasy One, but about of twoa Chinese manufacturer that is currently focusing on artificial intelligence and data centers.
The evolution in recent years in graphics chips, if we exclude AMD from the equation, is to include units in their cores to accelerate the parallel execution of artificial intelligence algorithms. Said elements, called according to each brand with the name of Tensor Cores, Matrix Units or any other synonym. This means that, as happened with scientific computing, they reached new markets. However, these technologies find their application in games and other applications, both home and professional.
Why is Biren a potential threat to NVIDIA, AMD and Intel?
A few weeks ago, Chinese hardware designer Biren showed his BR100, a brown beast of 77,000 million transistors and, therefore, with the same complexity as the RTX 4090, however, it is not a card designed for graphics, but rather it is focused on the AI market and does not have the fixed function units for graphs. So in that aspect it’s similar to AMD’s CDNAs, except using systolic charts, so it’s a waterline torpedo from NVIDIA with its H100 and Intel with Ponte Vecchio.
Biren’s most advanced configuration is made up of two GPUs and their respective HBM2 memories in their interposer, a very common configuration in server graphics cards. It is not a disaggregated chip, in fact there is a configuration with a single graphics chip called BR104. Of course, what is striking are its performance. The total power of the systolic tables of the BR100 allows it to reach 2 PetaFLOPS of power, in other words, 2000 TeraFLOPS operating in 8-bit floating point. The power of the NVIDIA H100 chip? 3 petaFLOPS.
However, it’s the 32-bit floating-point power that grabs the Chinese-made GPU’s attention, clocking in at 256 TFLOPS overall, 128 TFLOPs per chip. Which means that China has its own architecture which has all the capabilities to have a derivative for games. All this on a chip manufactured under TSMC’s 7nm node and using the same CoWoS technology as AMD’s Instinct MI200.
And what about the graphics card for games?
It would not cost the Chinese company anything to modify its architecture to launch a gaming version of its GPUs, however, it remains to be seen whether the United States is not against the use of more advanced nodes than TSMC’s 7 nm, because with these densities the size of the chips and the cost is prohibitive for the domestic market. However, the changes they need to make are simple:
- Fixed function units added:
- Rasterizer.
- Textures.
- ROPS.
- Laser trace.
- Advanced hardware video and audio codecs.
- Using GDDR6 memory instead of HBM2 to reduce cost.
In conclusion, China already has the tools to launch its own graphics cards and do without American companies in this field, for the moment in terms of artificial intelligence in servers, but at the same time they have everything to launch their own line of gaming graphics cards in the near future. Will we see a fourth competitor in the market? At the moment, the only Chinese product is the Fantasy One based on PowerVR technology, but they do not reach the level of power necessary to spit NVIDIA and AMD, which Biren’s architecture does.
However, the drivers will not make it easy for you
And it is that the biggest problem when making a graphics card for games is not just the hardware, but the drivers. But they’re asking one of the three builders who’ve sweated hours of work to get decent drivers and even in some cases there’s still a hell of a lot of work to do. It is therefore not only limited to the hardware part, but also to the software, which over time has become equally important to ensure the good performance and success of a gaming graphics card.
