first architecture news for GPU

In an interview with 4Gamers.net, both David Wang and Rick Bergman have launched interesting arguments about what their company will do with new and future GPUs. These will include architecture RDNA 4 and about its future news They have wanted to comment on some interesting details, which, without revealing too much, have made it clear that their path will distance themselves from NVIDIA quite a bit in the war for gaming and GPUs.

The RX 7000 with RDNA 3 has just come out and we are already thinking about RDNA 4. The truth is that NVIDIA has put a considerable margin of distance with its rival, especially considering that Ada Lovelace is a “rehash” of Ampere with few optimizations (RT and AI on RT Cores and Tensor Cores plus 4N). Therefore, AMD’s approach to its next architecture must be much more aggressive.

AI in gaming and the future that AMD sees

NVIDIA is taking advantage of its best AI system, both in hardware and software. DLSS 3 and the new Tensor Cores are hurting AMD… Or so we thought. But according to Wang, AMD is not going down that path, nor will it be in RDNA 4. In his opinion, AI is being misused in the desktop and gaming market, all despite the fact that AMD has incorporated it into RDNA 3. the calls AI Acceleratorwhich are nothing more than units designed to solve and accelerate the inference of Artificial Intelligence.

It is the equivalent of NVIDIA’s Tensor Cores or Intel’s MXMs. Compared these with the AI ​​Accelerator it turns out that the AMD option for the same range of GPUs is ostensibly inferior:

• Tensor Cores on RTX 4080 -> 195 TFLOPS
• MXM on Arc A770 -> 138 TFLOPS
• AI Accelerator on RX 7900 XTX -> 123 >TFLOPS

For this reason, AMD is not going to surpass its competitors in this aspect at any time, but it is not trying as such either. According to Wang, the current inference model “should not be limited to use focused on image processing.” The best example we have been enjoying for a long time is nothing more than FSR and its versions compared to DLSS. The former does not use an inference accelerator to increase performance and quality, the latter proprietary to NVIDIA does.

So why is AMD including AI Accelerators then? According to Wang, to get players to benefit from more advanced and fun games, but… How? The answer comes from the movement and behavior of the enemies and NPCs, but above all, a new thrust within the video game sector such as the so-called “neural graphics”:

Applying AI technology to make enemies and NPCs smarter or to make them look smart has a lot to do with the desEsports Extrasof the game mechanics, so depending on the GPU performance it seems to be a problem if behaves stupidly. By the way, neural graphics is a wide range of technologies, such as accelerating graphics pipelines with the help of artificial intelligence technology, improving quality, or generating new CG content.

AMD RDNA 4: first news and focus

There was not too much spoiler about the new architecture, although there was talk of techniques such as Primitive Shaders, Mesh Shaders and other techniques that are being implemented in their respective advances. The most important thing with what is to come were the declarations of bergmann about RDNA 4:

GPUs on PC these days are constantly evolving: from RDNA 1 to RDNA 2 and RDNA 3, and we promise to evolve to RDNA 4 with even higher performance in a Next future.

One of the parts that will bring news in view of what Wang said is the new Multi Draw Indirect or MDI that already integrates RDNA 3 and that in RDNA 4 it will have multiple improvements. It is being studied and investigated how to make it faster, because it will generate the graphics processing tasks directly on the GPU without the help of the CPU, which will reduce the bottleneck on it.

What will be tried is to completely avoid the transmission of data between the memory and the processor, so that the use of PCIe is also limited. Not so much for bandwidth, which we have more than enough with PCIe 5.0, but from the point of view of latency, the main problem in any PC and even more so with DDR5 at current speeds.

To give us an idea of ​​the support it is receiving, in RDNA 3 MDI is 2.3 times faster than in RDNA 2. Therefore, it is expected that in the new RX 8000 together with the news of RDNA 4 achieve even better performance and the dependence on the CPU is less and less, allowing it to scale more in FPS thanks to the fact that the work will be done by the graphics card. According to the statements we have seen, that “Next future” it could be late next year, or even sooner, as the gap with NVIDIA has widened.