TBH, I can already see what the PS5 can really do thanks to Bad Robo's Dreams content on YT.
Got a link for this? I'm intrigued...
Can the Tempest Engine be fed directly from the SSD... Bypassing RAM use altogether.
How effective would this be say for ML, its speculation atm but, it has been suggested Insomniacgames recent ML ZivaRT was running off the TE.
I don't know if the Tempest Engine can get data DMA'd directly from the SSD, but I'm sure that in most cases it will take data from RAM because it's audio input data HAS to originate from the game simulation and thus the CPU.
A mere cursory consideration of the suggestion that the TE is both responsible for IG's implementation of ML or any ML at all, and you can quickly dismiss such claims out of hand as ridiculous nonsense.
The TE is a single RDNA CU customized. ML computation nominally runs on ultra-wide low precision vector arrays like Tensor Cores, and/or on the entire GPU graphics and compute array, i.e. ALL the GPU CUs. How could one single custom CU somehow become more proficient at ML computation than the entire GPU vector array?... It's pretty much common sense that the TE won't be anywhere close to be powerful enough.
It's like the early suggestions that the XSX was using its CPU for ML computation. It's a bit daft. Even with the CPU AVX256 vector extensions, you won't get anywhere close to comparable performance with ML inferencing than you would running in GPU compute on even a low-end GPU.
” probably doesn’t meet the content and quality standards for posting on the forum...
As the company’s slogan says, “Play Has No limits”. 
www.neogaf.com
