I don't know... I always find posts like these somewhat hard to believe. Or hell even borderline disingenuous.
You can think what you will. If I took a picture of the 600+ games and of the numerous consoles I own you would find it to be overwhelmingly Sony centric.
How is that OT though? What does that have to do with PS5 Pro specs, considering that it was always meant to be a home console and not portable. If we are just using this thread to shit on various platforms and post our personal buying preferences, here is my take:
And therein lies my issue with comments like those. First off, your excitement for a PS5pro vs Switch 2... has nothing to do with a thread like this. Secondly, of course, a whole new console "generation" like the Switch 2 should be more exciting than what is literally just a souped-up PS5 (PS5 pro) and lastly, I don't know what more substance you expect a PS5pro to add to the PS5... anything outside better looking and performing PS5 games is quite literally an unrealistic expectation. Thats literally what the PS5pro is supposed to be and do, something that takes your PS5 games, and makes them look a little better while performing better too.
As does Nvidia. So technically, that's an apples-to-apples comparison. Technically.... but how the AMD and Nvidia GPUs perform, are just vastly different. Even when the specs suggest they should be very similar.
Kepler is not always accurate with his info, you can go and check for yourself, the tweets are still there.
With that being said, so far these are the SE count of RDNA4.
Navi44 - 16WGP (2SE, 32CU, 64ROPs)
Navi48 - 32WGP (4SE, 64CU, 128ROPs)
So the WGP per SE is still the same as RDNA3.
8WGP or 16CU per SE.
PS5 having 4SE is more probable than 2SE.
2SE with that many CU seems to be inefficient as well.
With higher end RDNA4 going by Navi44/48, it appears to be still 8WGP or 16CU per SE.
And it seems each SED houses 2 Shader Engines.
Navi41
1SED = 2SE/16WGP(32CU)
6SED = 12SE/96WGP(192CU)
1AID = 2×64bit or (4×32bit GDDR PHY)
2AID = 4×64bit GDDR PHY = 256bit
Navi40
1SED = 2SE/16WGP(32CU)
9SED = 18SE/144WGP(288CU)
1AID = 2×64bit or (4×32bit GDDR PHY)
3AID = 6×64bit GDDR PHY = 384bit
The concept of 1SED having 2SE is the same as Zen4C CCD having 2CCX.
Here we can see the size of 2SE.
Compared to XSX we have 8 more active CUs (running at a considerably higher frequency).
.
It was always just a question of clock speeds which I'll remind folks isn't 100% confirmed until Sony releases the official specs themselves. I've said many times that the clocks weren't going to be lower than the base system. Absolutely min was going to be 60 CUs with dual issue (RDNA3+) with no clock increase (2.23 Ghz) = 34.3 TFLOPS. I was hoping we'd get at least a 10% clock bump which would put it at 2.45 Ghz (~38TFLOPS). I'm still really curious why there isn't a fixed clock increase for either the CPU or GPU assuming it's using RDNA 4 as rumored
I'd just remind folks that 34-36 TFLOPs with 60CUs is ~6800XT level. That is in the 80-90% range for a real world FPS increase in games (assuming GPU bound). Nearly the 2x raster target that had been rumored for a while now:
Now if someone could explain the +45% "rendering performance" metric....?
Let's make something clear - caches are designed for >90% hitrates. Making them larger doesn't dramatically change the hitrates (unless the balance was really broken somewhere) - it just smooths out the 'bumps'. And where and how often such 'bumps' typically happen is discovered through profiling and statistics, which as you say - Sony and other companies have tons of data on to base such decisions from.
Yea but that's just stating the obvious. Caches themselves are a trade-off - if high performance memory was plausible/not cost prohibitive, hw makers would gladly skip the caches alltogether.
36TF is the new rumour.But 6800 level of performance in raster makes perfect sense.
33.5TF is slightly slower than 7700XT and 7700XT = 6800. And 6800 is...
This is compared to 6700, 44% - pretty much what Sony said.
Numbers checks out, expecting magical performance of 6800XT or even 3090 (lol) is setting yourself to disappointment. I bet it will be much faster than 6800 with RT but 90% of games are still raster only.
That's the reality (based on rumors).
The more clients you have for the L1 that might execute different pieces of code loading and writing data from different memory regions the more cache risks falling over (different CUs might work on different things… but sure that is the point of developers profiling and optimising their code
). 
I wouldn't rely on RDNA3 PC parts to scale RDNA 3.5/4 console parts. I just don't see how increasing the CU count by 66.67%, perhaps increasing the clocks (at least the max boost is 5% higher), and increasing the bandwidth by 29% lead to just a ~45% overall performance. Of course, there are as-of-yet-unknown details such as the ROPs, L2 cache, or memory chip speeds but, that just sounds like too meager of an improvement for something that comes out 4 years later after two die shrinks.But 6800 level of performance in raster makes perfect sense.
33.5TF is slightly slower than 7700XT and 7700XT = 6800. And 6800 is...
This is compared to 6700, 44% - pretty much what Sony said.
Numbers checks out, expecting magical performance of 6800XT or even 3090 (lol) is setting yourself to disappointment. I bet it will be much faster than 6800 with RT but 90% of games are still raster only.
That's the reality (based on rumors).
onQ123 mode on:
Sure, but it’s the exact same technology. Obviously, compute won’t scale linearly. The Pro not only has all those advantages but is two generations ahead.
That would mean that AMD has RT significantly faster than even Lovelace pound-for-pound which I highly doubt.
I will add when I was told 4070 levels of real world performance this was long before anyone heard a peep about PSSR, I don't know how that changes the equation and haven't asked
Yea but this case is precisely where L1 being larger probably helps the most. Shader workloads are still generally coherent (compared to what we do on CPUs) which would make L2 cache not the primary bottleneck. But yea it's all guesswork on our end of course - without knowing the stats for the software that Sony/AMD have.
If you have writes that go straight to memory you're bandwidth bound anyway? Caches are primarily a latency mitigation, saving on bandwidth is - not a common purpose (that's what scratchpads are for).
Depends on cache hierarchy we're talking about. OOOE can't absorb memory latency either - that's orders of magnitude too slow. But it can mitigate for L1 misses where In-Order wouldn't, yes.
We'll see but as I said, 45% sounds insanely low for two generations and a 4-year gap. I know PSSR and better RT are the big selling points but still, you need a strong baseline to work with.
I have little doubt this will be the case. 3080/4070 performance is pretty much exactly like the PS5 was positioned relative to the 2070S/1080 Ti.
I am wondering what the typical temporal coherence is across an entire shader engine (across many CUs).
I wonder if the local data share (LDS) and other on-chip scratchpads are increased per CU too.
You’re not alone bud. Been a PS main for years but their current direction is boring and sub-par.
basically a boost mode for unpatched games , makes sense
100%, they have a very shaky history on this, especially on the Great Dictator's side but also Richard's. The whole DF team always tend to make worst/most pessimistic assumptions when it comes to playstation hardware.
I feel the same way though. I like the ps5 but when I look at the last few games I've played most are last gen or not pushing the hardware that hard: marvel knights, dredge, nier automata, rdr2, bioshock....
Those two modifications (inherited from RDNA3) yet again show that there is more to RT performance than just the number of intersection engines/CUs. Cache amount/latency/bandwidth are all important factors.
Yep.
This is standard RDNA3 setup in this case, a rather imperative increase with that much WGP per SA/SE.
Those cards shouldn't be compared directly. Look at the post above yours. PS5 Pro should (and will) be compared with RDNA4, not RDNA3.
Not all is raster. My theory is that as we progress with the Emperor’s New Clothes that is raytracing in todays technology you end up with inefficiencies, hence the 45% increase.
Wait, isn't that the same number you began wit?
i think mostly yes unless their AI upscaling is actually really good.
Is that similar to subpixel rendering of text in Windows?
It uses a 200mm^2 GCD and 4x MCDs from Navi 31.
