PS5 is NOT 13.8TF. Only hardcore fanboys may belive in this.....
Of course. PS5 is 14TF, HBM2, ultra SSD, dedicated RT procesor and super cooling solution. Dream more....
Based on what we know, agreed. The 13.8TF claim is probably coming from that Japanese article (that didn't mention sources btw), but it was comparing PS5's Navi capabilities to Radeon VII's GCN TF capabilities. It wasn't
actually saying PS5 is 13.8TF Navi. If you read the wording, it says "comparable to 13.8 TF Radeon VII", or something like that. Here is the direct translated quote:
The approximate absolute performance value is determined by the manufacturing process to be adopted and the size of the chip (the number of transistors), so in light of the above, the GPU of the PS5 "is at best about 13.8 TFLOPS of Radeon VII. "Below" appears dim.
The bolded part is where it leaves room to speculate they are speaking of relative architectural performance here, not actual Moore's Law direct numerical comparisons. Just throwing "of" in there can really change things. Granted, it is a translated article, and something could've been lost in translation, but the rest of the article reads pretty concisely and seems like a fair and accurate translation even knowing the differences in language.
However, for PS5 to actually be 13.8 TF Navi, this would imply (best case/most optimal scenario) a 60CU chip with all CUs turned on and clocked @ 1800MHz (upper limit of Navi's sweetspot). Only conflicting part with that is there's only evidence of Arden with 56-60CUs and that is seemingly the XSX chip (also all Oberon-related benchmarks and leaks point to a 256-bit memory bus which would seem too small/narrow for a chip that size which could easily hold a 384-bit GDDR6 bus; the memory bandwidths reported are in range with a 256-bit bus and 14Gbps or 16Gbps chips)
So basically, if Navi is 20-25% more efficient than GCN, that article was most likely speculating PS5 to be somewhere in the range of 10.35 - 11.04TF Navi...
...which is EXACTLY in line with the speculation that Oberon might have about 48 active CUs (as per some chip die estimates and graphics for the APUs drawn up with measurements, which was posted on Era a while back), as that can provide a range of 10.448 TF Navi - 11.0592 TF Navi, within Navi's sweetspot 1.7GHz - 1.8GHz range. It also lines up with what another poster mentioned a while ago in PS4 Pro's GPU having a block of CUs disabled in regression testing for PS4 compatibility.
Basically, that article lines up very well with:
>Chip die estimates speculating an additional 12CUs on Oberon that were disabled (possibly for reasons either related to regression tests or bugs in the silicon that need to be fixed if they haven't already)
>Consistent GPU benchmark datamines (very likely the chip is still Oberon-related going this late into the cycle, high probability of a future GPU benchmark datamine for final Oberon chip showing 48 active CUs) (also some of the later steppings have been for bug fixes to the memory bandwidth, hinting possibility one reason we haven't seen a benchmark with more CUs active yet could be due to bugs that have to be worked out of the silicon).
>Github leak (the chip is very likely still Oberon or a future Oberon stepping; Github data already had other blocks such as RT disabled since RT would not be needed for regression testing of PS4 and PS4 Pro compatibility)
>Posters such as VFX_Veteran and R600 (who have speculated closer to the 9-10 TF Navi range; the Japanese article's comparison to Vega would line up with VFX, R600 etc's upper end limit if Navi delivers a 25% efficiency over GCN)
>Insiders such as OsirisBlack, KleeGameFan, Heisenberg etc (they have been providing 12/13TF number estimates for both systems and differentials of +/- 10%, but these estimates seem more in line with having used GCN as the basis of reference, i.e 12-13 GCN TF could be delivered via a range of 9TF Navi - 10.4TF Navi, and 48 CUs @ 1700MHz gives around 10.448TF Navi)(Keep in mind some insiders have at times also "accepted" a 9-10TF Navi range even after going 12/13 shortly after the XSX reveal at TGAs, when other leaks were coming out)
So essentially, ATM the
most probable scenario seems to be a 48CU PS5 providing anywhere between 10.4 TF Navi - 11 TF Navi, and a 56CU XSX providing around at least 12.1 TF Navi (but potentially up to 12.9 TF), all estimates based on Navi's sweetspot range. The
more realistic range would be PS5 pushing near 11 TF Navi and XSX pushing closer to 12.1 TF Navi (a differential of ~ 10%) ,
but it's still possible you could get a PS5 with 10.4 TF Navi and an XSX with 12.9 TF Navi (which would be a differential of ~ 24%). However there are likely going to be several factors that prevent that scenario from actually happening.
EDIT: Now to be fair, it's also possible XSX's speculated claims could have been conflating direct numerical comparisons with architectural references too, i.e rather than being 12TF Navi, XSX could actually be closer to 9 TF Navi (architecture improvements over GCN would give around 12 TF GCN performance). However, estimates of the APU size would suggest that there is more than enough room for around 52-56 active CUs, if not slightly more, and to reach 9TF Navi with that number the GPU would have be severely underclocked.
Perhaps at some point it was underclocked due to silicon bugs and early phase testing, but with that active number of CUs it would not have any problem reaching 12TF Navi at the low end of Navi's sweetspot. At the very least, XSX would be able to provide PS5's speculative low-end for TF performance (10.448), which in Navi, is still quite more than 9 TF Navi and therefore easily "more" than 2x the X.
Yet Phil very clearly said "2x" seemingly as an exact measurement, and an APU the size XSX's is estimated to be at, would not need to underclock itself so low to only deliver 9 TF Navi. So that is main reason why many feel the "2x of X" claim is in reference to 12TF Navi.