Digital Foundry Article Technical Discussion [2022]
- Thread starter BRiT
- Start date
- Status
I'm not so about that =P It can probably hurt Series X pretty badly if I'm being honest, something like cache scrubbers would greatly improve performance for Series X as well.Thanks again iroI was about to ask if this "cache scrubbing" process was automated or something that had to be done by the developer. Cause it's pretty obvious if it needs dedicated resources to maintain it's not ever being touched for third party games. If Sony could automate the process it could really help as a background process.
It probably needs this type of thing. Series X i think is definitely not bandwidth starved with that insane amount of bandwidth. and I thought even at launch that 448 may not be enough for PS5. But developers seem to make due with what they've got
There's a lot of things at play here, it's very hard to quantify. But programmers who really excel at optimization and fully understand how to leverage a GPU, the cache scrubbers are probably not that impactful overall. But developers who don't have the time or energy to deal with it, cache scrubbers is probably doing a big lift. With PS5 as the lead console, this puts Series console ports at a pretty bad disadvantage at times.
I do hope to see cache scrubbers show up on AMD hardware at a later time. But if it never shows up, then it's performance is probably less substantial than we thought.
Series S could definitely use the help of some dedicated silicon to help maximize the limited number of resources available to it.
Series S could definitely use the help of some dedicated silicon to help maximize the limited number of resources available to it.
Mark Cerny told specifically about this too saying AMD can reuse anything if they think it will be useful and said specifically AMD did not find something useful on PC side for the cach scrubbers. The cache scrubbers can't work without the coherency engine part the SSD solution of Sony. It means probably work in coordination with SSD maker on PC to add some coherency engine just for AMD, it seems not really useful and not something SSD maker would lost their time for only AMD. This is probably not useful out of a console where you want a fully integrated solution.
They aren't interested by Id buffer and maybe flexible rasterization and they probably have good reason.
They aren't interested by Id buffer and maybe flexible rasterization and they probably have good reason.
Last edited:
It depends on a couple of things. If it's in general, used all the time whether the SSD calls for info or not, then it's a pretty big impact. Cause you flush the cache on all sorts of operations, like state changes etc.
If the cache scrubbers only work in conjunction with the SSD, ehhhh, then I see why it's never coming to GPUs, that vertical integration between SSD I/O and GPU is too tight for the PC space. And then the amount it gets used is probably not as massive if it's tied to the SSD.
okay that settles the discussion pretty quickly then.
davis.anthony
Veteran
XSX has the lowest amount if bandwidth per CU on ANY RDNA2 powered GPU.
davis.anthony
Veteran
The cache scrubbers have never been spoken about by Sony in detail and they have never explained how they work
Last edited by a moderator:
jesus christ
Pretty normal for last gen, its just that people never thought about it. Not before FG came into play. 35ms doesnt seem so bad now.
About cache scrubbs, the performance boost isnt really visible compared to equal class gpu’s (6600xt). Infinity cache is a more substantional performance future. If the cache scrubbs are there to make the gpu perform as it should then they created them to allivate performance.
Probably way cheaper than IC too.
Inuhanyou
Veteran
Really?! But that's a huge amount of bandwidth they have. More than half of 1tb...
Inuhanyou
Veteran
It makes sense. If it's there to "break even" and retain more of the performance the gpu should be getting as opposed to gaining anything extra.
I don't know how GPU and CPU retention works but it seems like you can lose a lot of performance in APUs and such if things are not integrated just right. And cache scrubbers may be a way to help ease this issue.
Also about input lag, I remember it being a huge conversation around the time of killzone 2 on ps3
but that generation was always weird
Last edited:
davis.anthony
Veteran
Yep
XSX
560GB/s / 52 CU's = 10.76GB/s per CU
XSS
224GB/s / 20 CU's = 11.20GB/s per CU
PS5
448GB/s / 36 CU's = 12.44GB/s per CU
The problem XSX has is Microsoft have built a wider GPU than Sony but are trying to feed it with a slower front end, less cache and less bandwidth.
He spoke about it in the road to PS5. This is just above and in the @BRiT quote of the same video. An I can add Mark Cerny told too we have a custom RDNA 2 include part of the feature set we think useful for PS5 and we have our own feature only useful in PS5 like cache scrubber. They talk about this the first time they gave technical details about the PS5.
EDIT: From 25:40 he explains all the stuff about the difference of the PS5 custom GPU and at 25:55, he said directly than GPU cache scrubber are too specific to what Sony want to accomplish and AMD don't need this on PC.
The cache is inline with the number of CUs, which is inline with the larger memory bus. L2 is 5MB vs PS5s 4MB. 25% larger to accommodate 25% more CUs approximately.
It shouldn't be that off, they would have found major issues during simulation here if the CUs sat around starved for work to do.
The split memory pool is likely having a much larger effect here than everything else.
If we see the way PS5 exclusives are being built, with VRAM requirements much higher than 10GB. Welp, XSX is SOL then, because now it needs to do weird shuffling operations.
Any 3rd party that goes down this route to take full advantage of the 12-13GB on PS5 will need to backtrack on series consoles as they can't really do the same here with XSX.
I very much doubt cache optimization is the killer here as any work done on PS5 would carry over to XSX in this regard.
Inuhanyou
Veteran
Slower "front end"? What does this mean exactly? And what is the cache disparity exactly? I just kind of assumed series x had an out and out advantage in most areas so this is new to me
- Status
Similar threads
