AMD’s FSR 2.0 technology was recently introduced, but it seems that the Sunnyvale company is already working on FSR 3.0, and that it could mark a significant technological leap. As many of our readers will know, what differentiates the second generation DLSS and FSR 2.0 is that the first uses artificial intelligence, and that it is accelerated through the tensor cores that are included in the GeForce RTX 20 and RTX 30 graphics cards, while than the second it does not use AI or rely on dedicated hardware.
This could change with the arrival of FSR 3.0 and the Radeon RX 7000, a new generation of graphics cards that will debut later this year and that, according to an interesting leak, will be ready to work at the hardware level with the Wave Matrix Multiply-Accumulate instructions. These instructions work with arrays of tables containing numbers, and they allow you to speed up the work required to use artificial intelligence and deep learning algorithms to multiply large sets of numbers.
Although it is interesting, the truth is that it is not something new, since AMD’s CDNA architecture already offered support for Matrix-Fused-Multiply-Add, which were accelerated in the Matrix Core Engines. It is clear that to work optimally with these new instructions, the Radeon RX 7000 should have specialized hardwareand therefore the FSR 3.0 would end up being exclusive to that graphic generationalthough I still do not know what kind of implementation AMD would carry out in this regard, and the truth is that it is a very interesting topic.
AMD FSR 3.0 and artificial intelligence: challenges at the hardware level
And also software. I do not want to go into complex issues because in the end what is important are the keys behind this news. If FSR 3.0 uses specialized instructions for AI, its quality should improve markedly, and so could its performance. This is important because would put it squarely in the same league as NVIDIA’s DLSS and Intel’s XeSS playbut at the same time raises many unresolved questions.
First of all, there is what we have already anticipated, and that is that AMD would have to integrate specialized hardware in the Radeon RX 7000 to work with the new workload that FSR 3.0 supported by artificial intelligence would entail. It may seem easy, but this requires very valuable space at the silicon level, and involves the addition of a new component that consumes resources and generates heat. AMD is already quite limited on space at the silicon level in their Radeon RX 6000s, as they use L3 cache as “infinite cache”, but I think moving to the 5nm node and using an MCM design will allow them to overcome this problem. .
Personally, I think it is likely that AMD will maintain the approach used in CDNA-based graphics solutions, and that the Radeon RX 7000 will come equipped with one Matrix Core Engine (this may not be its final name) for each compute unitwhich will be specialized in AI.
If all this is confirmed, the FSR 3.0 will mark an important turning point, but also dwill make the Radeon RX 6000 “obsolete” and earlier models, which will be limited to FSR 2.0 technology. To all this we must add another important question, and that is that if FSR 3.0 ends up requiring specialized hardware could become a closed standardas is NVIDIA’s DLSS.
