I have twice typed a post on this, but also deleted it twice prior to posting. Since I want to avoid technicalities, and not all AMD GPUOpen documents are public.
For now, we know that FSR 4 is dependent on FP8 WMMA. It is not known whether it needs other instructions or libraries. Documentation supports that the RDNA 3 architecture can do FP16, BF16, IU8 (8-bit integer) and IU4 (4-bit integer) datatypes. Not going into the technicalities, but WMMA is available in the rocWMMA library.
This library is portable with nvcuda::wwma and it support MFMA and WMMA instructions. This is applicable to RDNA 3 and CDNA 1 and 2.
Everything can be read here:
https://gpuopen.com/learn/wmma_on_rdna3/
RDNA 3, unlike RDNA 2 and 1, has AI Accelerators. It is not like Nvidia, which has Tensor cores that are dedicated, and CUDA cores that are dedicated. The AI Accelerators, however, are not dedicated and have to do other computational tasks.
This is something many here should know already. CDNA, unlike RDNA, has Matrix Cores, that is dedicated. Trying to keep this short. FP8 instructions were introduced to MI300, well, GFX940. RDNA 3 is GFX11xx and RDNA 4 is GFX12xx. Way too many numbers, but codes, compilers and so on, are written. All that remains is optimization. Some think that AMD is not doing its best at this, but that is what it is.
I typed all this to get back to FSR 4. The word is that FSR 4 will not be available on RDNA 3 because RDN A3 does not have native FP8 WMMA. It is too early to say, but RDNA4 does support new libraries that have been introduced in CDNA3. FSR4 was likely developed on CDNA 3. Backporting is the wrong word, though some would say that FSR 4 could be backported to RDNA3, but with limited capabilities. Well FP8 -> FP16. Later FSR iterations could move to other instructions or back to FP16 depending on how well utilization is optimised. Could FP8 WMMA be emulated on RDNA 3, possible, but emulation is a resource hog, again, optmisation.
Why do I type all this, well, it is a bore to read documentation, but FSR 4 is theoretically possible on RDNA3, and RDNA3 at the top-end has the RAM. Will AMD dedicate resources to this... dunno. FSR 4 on the 7900 series cards could put the 9070, results pending, in bad light. It is without a doubt that FP8 WMMA, its 'magic', that essentially is a lower precision data type than FP16 will be called AI. AI and its revolutionary tricks.
This begs the question, can "AI" bring FSR 4 to RDNA3? Though this depends on FSR 4 dependencies that are not all yet known. There is a reason why AMD doesn't have a new driver out other than a preview, which usually showcases its new beta things. The answer is: It's possible. RocWMMA and other hardware-accelerated applications/instructions have shown improvements. Slow, but steady strides.
The problem that buyers might have is that RDNA 4 is the last bridge to cross to UDNA. Back to the GCN days. UDNA will be as dedicated as CDNA, since well, it is CDNA and RDNA in one.
Anyway, it is an old LLVM examination:
As 2024 continues on, because time never stops, AMD has been working on their upcoming RDNA 4 architecture.
chipsandcheese.com
and we don't actually know what RDNA 4 looks like, architecturally, until it is released with its documentation.
All that can be said now is that RDNA 3 and older... eh. Support will slow down. Nvidia is beating AMD at long-term support. Though AMD, is understandably making radical architectural changes. Not all hope is lost. AMD is pursuing to querying the market to bring ROCm, updated, to older GPUs to enable them "AI" capabilities, and this could include GCN, well, Vega, not older, and some think that HBM2 can still do some new tricks. Frontier cards had 16GB RAM and so did VII.
So RAM. By now gamers should know why 12-16GB is a 'consumer' limit, unless you go halo class. It is because they don't consumer products to be bought by their commercial customers.
Anyhow, long post. TLDR, FSR 4 is technically possible on RDNA 3 (at the high end). Who knows. Marketing and technical teams... You know, AMD wants to sell new GPUs.
www.titan-ice.co.za
