DJ12
Member
They didn't take a bigger RDNA2 design at all, they took the same one and added 4CUs to each shader array.
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
PS5 Die Shot has been revealed
- Thread starter LivingD3AD
- Start date
SenjutsuSage
Banned
Another thing people aren't factoring in is that neither Nvidia or AMD share ALL the details of what changes they've made and how for competitive purposes when they release new hardware.
But there's a reason RDNA 1 doesn't support Mesh Shaders, VRS, Sampler Feedback and Ray Tracing, or Machine Learning Acceleration. The hardware change that AMD has been most forthcoming about is how they added ray tracing. They showed us that.
They've been less clear on how RDNA 2 now supports Mesh Shaders, less clear on how it now supports Sampler Feedback, and they told us how it supports VRS in hardware, the ROPs. The ROP units are clearly different and have been upgraded.
My assumption for RDNA 2 and mesh shaders is the geometry engine is upgraded and possibly command processor. Hell, there could be a change to compute units we aren't aware of for Sampler Feedback, it could be the TMUs, could also be related to the ROPs, I really and truly have no idea. Point I'm making is this: Physical appearance may not have changed, but a crap ton clearly changed internally, and basically had to in order for RDNA 2 to have all these new features advantages it has over RDNA 1.
Nevaeh
Member
and MLID is now exposing RedGamingTech but he really is just exposing himself too lol
You really love to see it
Last edited:
This. Move on arguing from a place of unconfirmed -gotchas-, Speculation is one thing, outright warring over things you have no definite answer to is another- that is served just to annoy and leading to thread bans.
S0ULZB0URNE
Member
Read through this thread RDNA 2 clocks have been posted.
prinz_valium
Member
Some YT drama and beef. You love to see it
But those (while also ridiculous) don't take to Twitter and demand clarification from Sony employees.
ethomaz
Banned
I think you get it wrong... he was inserted in the discussion by a Xbox Fan that demanded his clarification.
Last edited:
S0ULZB0URNE
Member
He is taking out of both sides of his mouth.But he says it's 100% RDNA2.
You can't say something is 100% and have part of it be something else.
Garani
Member
Another fanboy. Nothing to see here, move along.
But we still get the stupid FUD about PS5 being RDNA1. Oh, and Sony employees keep themselves away from social media, and the only poor soul who answered a DM got treated like shit and almost lost his job.
You know why I am here to bullshit on NeoGAF? Because it's not in my line of work. You can bet that I am keeping myself miles aways from any conversation that has to do with my employer or the industry that employer is in: if I were to speak I would get totally destroyed by some keyboard warrior and I would loose my job to boot. So yeah, there are no Sony engineers that will interact with you for that exact reason.
SenjutsuSage
Banned
The silicon includes the ENTIRE Xbox Series X system on a chip, GPU, CPU, Decompression hardware, audio hardware etc. The GPU portion is 100% RDNA 2. The CPU portion is Zen 2 and other parts are custom. What is so difficult for you to understand. A CPU can't be RDNA 2.
DJ12
Member
If you were talking to me... I am no expert, but pictures are worth a thousand words.
This is an RDNA2 slide, 80CUs in total. 20 CUs (10 per shader array), or 10 DCUs (5 per shader array) per shader engine if you prefer.
Series X, 28 CUs (14 per shader array), or 14 DCUs (7 per shader array) per shader engine.
Here's a link to a slide from MS's hot chips briefing stating these facts: xboxseries-xhot-chips-gpu-overview-architecture-RDNA2CU.jpg (2970×1700) (redgamingtech.com)
PS5, 20 CUs (10 per shader array), or 10 DCUs (5 per shader array) per shader engine.
S0ULZB0URNE
Member
The silicon includes the ENTIRE Xbox Series X system on a chip, GPU, CPU, Decompression hardware, audio hardware etc. The GPU portion is 100% RDNA 2. The CPU portion is Zen 2 and other parts are custom. What is so difficult for you to understand. A CPU can't be RDNA 2.
That's not my take from this but ok.
Elog
Member
Can we just stop the stupidity? Both consoles are RDNA2 based with customisations. The End.
Personally, I really like the die shot that this thread is supposed to be about. Interesting customisations to the CU/TMUs that is different between the two consoles and curious about what logic customisations that have been made to the GE in the PS5. The lack of infinity cache was at least for me expected (even though one could dream about some off-die cache pool!).
Personally, I really like the die shot that this thread is supposed to be about. Interesting customisations to the CU/TMUs that is different between the two consoles and curious about what logic customisations that have been made to the GE in the PS5. The lack of infinity cache was at least for me expected (even though one could dream about some off-die cache pool!).
Tickle My Tendies
Member
You deserve GAF Gold for this.
Thank you.
In the end, sooner or later, it really doesn't matter. Both consoles kick major ass and we are all in for a great gen.
Last edited:
HoofHearted
Member
I know that clocks have been "posted" - I'm trying to understand relevance in RDNA 1 vs 2.... You mentioned a "spec" attached - where is it?
There seems to be quite a bit of chatter around the clock speed as a direct indicator which is just.... weird.
onesvenus
Member
You haven't read the thread obviously. When he talks about the silicon he is talking about all the SoC, that's why he references Zen2.
bender
What time is it?
If only you had a youtube channel and knew how to rap, then we could heed your advice.
DJ12
Member
And yet we have twitter experts with knowledge of RDNA 1 and 2 stating that neither console is full RDNA2.
I would wager when they/he mention "full RDNA2" they are referring to the feature set, rather than the technical layout.
Last edited:
Well it is stupid and ridiculous either way. If the system designer said it is RDNA2 then it is RDNA2. I also stay away from discussion involving my line of work, is just too nerve wrecking to respond to idiots go on and on about shit they don't understand because they read an article about some building 400 meters tall or saw an episode of some structure on Discory channel.
Loxus
Member
No, if you look at both, you'll see Navi22 has more things on die and the ROPs are more squared than rectangular.
geordiemp
Member
You did not post my calculation that it was likely zero or at most 8 MB if we wre lucky, but you knew that as you really are Riky.
I was quoting someone elses comment. Try reading, I know its hard.
Last edited:
John Wick
Member
He means other things beside the CPU and GPU are custom like the Hardware Audio processor.
That's not my take from this but ok.
Last edited:
S0ULZB0URNE
Member
So just the like PS5 minus the DX 12+ silicon being cut out.
geordiemp
Member
RDNA2 is a marketing term used by whoever wants to with AMD blessing - who cares ? Could be called blollocks2 it does not matter.
What matters is how the chip runs games, thats all. The rest is just PR.
The interesting part of the ps5 is how the CU is arranged and stuff we cant see from die shots anyway.
Locuza
Member
That's just a slightly different graphical representation.
The package numbers are from Navi10:
https://www.techpowerup.com/gpu-specs/amd-navi-10.g861
You can also see 8 memory chips on the board (256-Bit) however Navi22 only supports 6/12 since the G6-Interface got reduced to 192-Bit.
Last edited:
geordiemp
Member
You should read what I said the maximum possible IC could be and most likely was zero, You quoting me quoting someone else.
Loxus
Member
Good point.
SenjutsuSage
Banned
If you were talking to me... I am no expert, but pictures are worth a thousand words.
This is an RDNA2 slide, 80CUs in total. 20 CUs (10 per shader array), or 10 DCUs (5 per shader array) per shader engine if you prefer.
Series X, 28 CUs (14 per shader array), or 14 DCUs (7 per shader array) per shader engine.
Here's a link to a slide from MS's hot chips briefing stating these facts: xboxseries-xhot-chips-gpu-overview-architecture-RDNA2CU.jpg (2970×1700) (redgamingtech.com)
PS5, 20 CUs (10 per shader array), or 10 DCUs (5 per shader array) per shader engine.
What are you implying? That because the compute unit layout is different compared to a single RDNA 2 GPU, one is more or less in line with the RDNA 2 spec or design? You would be wrong. Not all RDNA 2 GPU are perfect multiples of 10 in terms of their number of Compute Units or even perfectly divisible across 4 shader engines.Did you forget about the 6800 XT and 6800? 72 Compute Units and 60 Respectively.
RX 6800 would thus have 30 DCUs total. Wait a minute, this doesn't perfect divide by 4... how does this work??
1st shader engine on RX 6800 has 8 DCU
2nd shader engine on RX 6800 has 7 DCU
3rd shader engine on RX 6800 has 8 DCU
4th shader engine on RX 6800 has 7 DCU
Totaling 30 DCU for 60 Compute Units
For the 6800 XT's 72 Compute Units, that would be 36 Dual CUs with 9 in each of the 4 shader engines. This one maps out perfectly.
Let's now have a look at the Series X layout. 56 CUs total, meaning 28 DCU. But remember 2 must be disabled since Series X only has 52 active Compute Units, leaving 26 DCU.
This translates to 7 DCU per shader engine, but remember that 2 of these must be disabled.
Leaving Series X with
7 active DCU in 1st shader engine
6 active DCU in 2nd shader engine (one disabled Dual CU)
7 active DCU in 3rd shader engine
6 active DCU in 4th shader engine (one disabled Dual CU)
Series X is pretty much scary close to the RX 6800 in its design and Dual Compute Unit Layout. Remove a single Dual Compute Unit from each shader engine belong to RX 6800, and you're left with the Xbox Series X design.
See how that game works? You seem to want to believe they took the RDNA 1 5700XT layout and added 4 CUs (2 Dual Compute Units) per shader engine, but why not RX 6800 and they simply removed 2 CUs (1 Dual Compute Unit) per shader engine?
No matter how one wishes to view the matter, Series X is closer to RX 6800 than it is to RX 5700 XT.
Now what about the PS5 layout? 40 total compute units aka 20 Dual Compute Units, but only 36 active compute units, leaving you with only 18 active Dual Compute Units on PS5. Similar type of layout to RX 6800 and Series X.
PS5 has
5 active Dual CUs in 1st shader engine
4 active Dual CUs in 2nd shader engine (one disabled Dual CU)
5 active Dual CUs in 3rd shader engine
4 active Dual CUs in 4th shader engine (one disabled Dual CU)
PS5 = 10 CUs, 8 CUs, 10 CUs, 8 CUs design = 36 CUs
Series X = 14 CUs, 12 CUs, 14 CUs, 14 CUs design = 52 CUs
Series X is more or less RX 6800 minus 4 Dual CUs and 32 less ROPS and of course no Infinity Cache.
S0ULZB0URNE
Member
And a lower clocked GPU with different power requirements.
SenjutsuSage
Banned
Last edited:
kyliethicc
Member
The consoles are not like the Big Navi 21 die. Navi 21 is twice as large in layout.
What are you implying? That because the compute unit layout is different compared to a single RDNA 2 GPU, one is more or less in line with the RDNA 2 spec or design? You would be wrong. Not all RDNA 2 GPU are perfect multiples of 10 in terms of their number of Compute Units or even perfectly divisible across 4 shader engines.Did you forget about the 6800 XT and 6800? 72 Compute Units and 60 Respectively.
RX 6800 would thus have 30 DCUs total. Wait a minute, this doesn't perfect divide by 4... how does this work??
1st shader engine on RX 6800 has 8 DCU
2nd shader engine on RX 6800 has 7 DCU
3rd shader engine on RX 6800 has 8 DCU
4th shader engine on RX 6800 has 7 DCU
Totaling 30 DCU for 60 Compute Units
For the 6800 XT's 72 Compute Units, that would be 36 Dual CUs with 9 in each of the 4 shader engines. This one maps out perfectly.
Let's now have a look at the Series X layout. 56 CUs total, meaning 28 DCU. But remember 2 must be disabled since Series X only has 52 active Compute Units, leaving 26 DCU.
This translates to 7 DCU per shader engine, but remember that 2 of these must be disabled.
Leaving Series X with
7 active DCU in 1st shader engine
6 active DCU in 2nd shader engine (one disabled Dual CU)
7 active DCU in 3rd shader engine
6 active DCU in 4th shader engine (one disabled Dual CU)
Series X is pretty much scary close to the RX 6800 in its design and Dual Compute Unit Layout. Remove a single Dual Compute Unit from each shader engine belong to RX 6800, and you're left with the Xbox Series X design.
See how that game works? You seem to want to believe they took the RDNA 1 5700XT layout and added 4 CUs (2 Dual Compute Units) per shader engine, but why not RX 6800 and they simply removed 2 CUs (1 Dual Compute Unit) per shader engine?
No matter how one wishes to view the matter, Series X is closer to RX 6800 than it is to RX 5700 XT.
Now what about the PS5 layout? 40 total compute units aka 20 Dual Compute Units, but only 36 active compute units, leaving you with only 18 active Dual Compute Units on PS5. Similar type of layout to RX 6800 and Series X.
PS5 has
5 active Dual CUs in 1st shader engine
4 active Dual CUs in 2nd shader engine (one disabled Dual CU)
5 active Dual CUs in 3rd shader engine
4 active Dual CUs in 4th shader engine (one disabled Dual CU)
PS5 = 10 CUs, 8 CUs, 10 CUs, 8 CUs design = 36 CUs
Series X = 14 CUs, 12 CUs, 14 CUs, 14 CUs design = 52 CUs
Series X is more or less RX 6800 minus 4 Dual CUs and 32 less ROPS and of course no Infinity Cache.
The XSX GPU has 4 shader arrays with 7 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The PS5 GPU has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The XSS GPU has 2 shader arrays with 6 DCUs each. 1 shader engine. 2 DCUs are disabled for yields.
The 6800 has 8 shader arrays with 5 DCUs each. 4 shader engines. 10 DCUs are disabled. Yields, segmentation.
The 6800 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. 4 DCUs are disabled. Yields, segmentation.
The 6900 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. No DCUs are disabled. Full Navi21 die.
The upcoming Navi22 die will be used for the 6700 XT and 6700. This is much closer to the consoles.
The 6700 has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled. Yields, segmentation.
The 6700 XT has 4 shader arrays with 5 DCUs each. 2 shader engines. No DCUs are disabled. Full Navi22 die.
The PS5 will be basically same as Navi 22 6700, while XSX is again the same, but with 7 DCUs per SA instead of 5.
Neither console has 4 shader engines like Navi 21. Only 2. And XSS only has 1 SE.
Navi 21 is Navi10 x2.
XSX is just 2 extra DCUs per SA. Same 2 SE as PS5, Navi22 and Navi 10.
Last edited:
S0ULZB0URNE
Member
Anyone who games on PC knows the GPU'S don't stay at the fixed clock.Series X's fixed clock is higher than the RX 6800's game frequency.
My 1080ti for example reaches well above it's fixed clock to around 2030mhz.
Last edited:
DJ12
Member
Way to follow the convo.
What are you implying? That because the compute unit layout is different compared to a single RDNA 2 GPU, one is more or less in line with the RDNA 2 spec or design? You would be wrong. Not all RDNA 2 GPU are perfect multiples of 10 in terms of their number of Compute Units or even perfectly divisible across 4 shader engines.Did you forget about the 6800 XT and 6800? 72 Compute Units and 60 Respectively.
RX 6800 would thus have 30 DCUs total. Wait a minute, this doesn't perfect divide by 4... how does this work??
1st shader engine on RX 6800 has 8 DCU
2nd shader engine on RX 6800 has 7 DCU
3rd shader engine on RX 6800 has 8 DCU
4th shader engine on RX 6800 has 7 DCU
Totaling 30 DCU for 60 Compute Units
For the 6800 XT's 72 Compute Units, that would be 36 Dual CUs with 9 in each of the 4 shader engines. This one maps out perfectly.
Let's now have a look at the Series X layout. 56 CUs total, meaning 28 DCU. But remember 2 must be disabled since Series X only has 52 active Compute Units, leaving 26 DCU.
This translates to 7 DCU per shader engine, but remember that 2 of these must be disabled.
Leaving Series X with
7 active DCU in 1st shader engine
6 active DCU in 2nd shader engine (one disabled Dual CU)
7 active DCU in 3rd shader engine
6 active DCU in 4th shader engine (one disabled Dual CU)
Series X is pretty much scary close to the RX 6800 in its design and Dual Compute Unit Layout. Remove a single Dual Compute Unit from each shader engine belong to RX 6800, and you're left with the Xbox Series X design.
See how that game works? You seem to want to believe they took the RDNA 1 5700XT layout and added 4 CUs (2 Dual Compute Units) per shader engine, but why not RX 6800 and they simply removed 2 CUs (1 Dual Compute Unit) per shader engine?
No matter how one wishes to view the matter, Series X is closer to RX 6800 than it is to RX 5700 XT.
Now what about the PS5 layout? 40 total compute units aka 20 Dual Compute Units, but only 36 active compute units, leaving you with only 18 active Dual Compute Units on PS5. Similar type of layout to RX 6800 and Series X.
PS5 has
5 active Dual CUs in 1st shader engine
4 active Dual CUs in 2nd shader engine (one disabled Dual CU)
5 active Dual CUs in 3rd shader engine
4 active Dual CUs in 4th shader engine (one disabled Dual CU)
PS5 = 10 CUs, 8 CUs, 10 CUs, 8 CUs design = 36 CUs
Series X = 14 CUs, 12 CUs, 14 CUs, 14 CUs design = 52 CUs
Series X is more or less RX 6800 minus 4 Dual CUs and 32 less ROPS and of course no Infinity Cache.
Someone said MS paid for the "Big navi" design to get more CUs.
Which of course is total rubbish.
RDNA 2 is 10 CUs per shader engine (as per the PS5 regardless of any that are disabled)
As I said MS had AMD fit more CU's in the shader array, they did not buy a "big navi" design.
PS5 and Series X both have 2 shader engines.
Series X is only "scarily" close to the 6800 if you add some fanboyism and squint your eyes.
"The RX 6800 is further cut down by enabling 60 out of 80 CUs. It's unclear if a full Shader Engine or 10 random Dual CUs have been disabled to accomplish this." (this is from tech powerup's write up of the 6800 series)
So it has 3 or 4 shader engines (most likely 4 as requiring 3 fully functional shader engines probably misses the point of disabling some for yeilds anyway) enabled with 10 DCUs as per PS5 and all other RDNA2 cards with 10 CUs disabled across the board.
There is no 14 CU shader engine design for RDNA 2 besides what MS have asked for.
Last edited:
ethomaz
Banned
RX 6800 runs at avg. 2200Mhz in gaming.Series X's fixed clock is higher than the RX 6800's game frequency.
Even the mininum is around 200-300Mhz higher than that game clock.
Last edited:
Physiognomonics
Member
All those AMD GPUs have 5 DCUs by shader array because it's the sweatspot between compute and rasterization, ideal for gaming. PS5 has also 5 DCUs by shader array because it's the best balance for gaming on RDNA 1 and RDNA 2 GPUs.The consoles are not like the Big Navi 21 die. Navi 21 is twice as large in layout.
The XSX GPU has 4 shader arrays with 7 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The PS5 GPU has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The XSS GPU has 2 shader arrays with 6 DCUs each. 1 shader engine. 2 DCUs are disabled for yields.
The 6800 has 8 shader arrays with 5 DCUs each. 4 shader engines. 10 DCUs are disabled. Yields, segmentation.
The 6800 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. 4 DCUs are disabled. Yields, segmentation.
The 6900 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. No DCUs are disabled. Full Navi21 die.
The upcoming Navi22 die will be used for the 6700 XT and 6700. This is much closer to the consoles.
The 6700 has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled. Yields, segmentation.
The 6700 XT has 4 shader arrays with 5 DCUs each. 2 shader engines. No DCUs are disabled. Full Navi22 die.
The PS5 will be basically same as Navi 22 6700, while XSX is again the same, but with 7 DCUs per SA instead of 5.
Neither console has 4 shader engines like Navi 21. Only 2. And XSS only has 1 SE.
Navi 21 is Navi10 x2.
XSX has 7 DCUs by shader array because it has being designed for a dual purpose: gaming console and cloud computing. This is also why the console is mostly performing under it's theoretical TF number, usually like a 10tf AMD GPU like PS5.
But it should be performing pretty well at compute tasks so in specific modes like Cloud computing or even compute bound photo mode in a game.
Last edited:
SenjutsuSage
Banned
The consoles are not like the Big Navi 21 die. Navi 21 is twice as large in layout.
The XSX GPU has 4 shader arrays with 7 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The PS5 GPU has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled for yields.
The XSS GPU has 2 shader arrays with 6 DCUs each. 1 shader engine. 2 DCUs are disabled for yields.
The 6800 has 8 shader arrays with 5 DCUs each. 4 shader engines. 10 DCUs are disabled. Yields, segmentation.
The 6800 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. 4 DCUs are disabled. Yields, segmentation.
The 6900 XT has 8 shader arrays with 5 DCUs each. 4 shader engines. No DCUs are disabled. Full Navi21 die.
The upcoming Navi22 die will be used for the 6700 XT and 6700. This is much closer to the consoles.
The 6700 has 4 shader arrays with 5 DCUs each. 2 shader engines. 2 DCUs are disabled. Yields, segmentation.
The 6700 XT has 4 shader arrays with 5 DCUs each. 2 shader engines. No DCUs are disabled. Full Navi22 die.
The PS5 will be basically same as Navi 22 6700, while XSX is again the same, but with 7 DCUs per SA instead of 5.
Neither console has 4 shader engines like Navi 21. Only 2. And XSS only has 1 SE.
Navi 21 is Navi10 x2.
XSX is just 2 extra DCUs per SA. Same 2 SE as PS5, Navi22 and Navi 10.
According to the official whitepaper, a shader engine consists of a rasterizer unit and a primitive unit. Series X clearly has a prim unit and raster unit in each of those 4 engines. Are you certain that isn't 4 distinct shader engines?
Okay, see my mistake. They include 4 raster and prim units no matter what the design. And so this mean, i need to adjust what I said about the RX 6800. That only has 3 shader engines I learned. Did they cut a raster and prim unit for that card? Otherwise I don't see how they did it. After all, they disabled the last shader engine altogether.
Last edited:
Knitted Knight
Banned
So when are we gonna talk about the games?
SenjutsuSage
Banned
Oh god, that nonsense from Moore's law again. The console has access to all 12 TFLOPS for gaming. There isn't any performance dedicated solely to cloud computing.
S0ULZB0URNE
Member
Which games for what platform?
ethomaz
Banned
Probably in threads about these games.
Wait it is already happening.
Last edited:
thicc_girls_are_teh_best
Member
Pretty much; like the Series X, PS5 has a mix of RDNA 1, RDNA 1.1 and RDNA 2 silicon, among other things not currently documented as known quantities in AMD GPUs (in PS5's case, the cache scrubbers, and in Series X's case, the mip-blending hardware for SFS in the GPU).
Neither's pure RDNA 2, I highly doubt either have any RDNA 3 features, and they're both powerful pieces of kit regardless. It's disingenuous calling either RDNA 1.5 just because the FPUs might be cut down on one, or the GPU clocks are relatively mild on the other.
Watching the fallout with MLID/RGT etc. tho, I can't lie, it's been quite hilarious.
SirTerry-T
Member
I stumble into these threads knowing full well that I will not understand 99.999% of the technical conversations taking place in here but I do recognise arrogance and ignorance when I see it.
The sheer nerve of certain individuals in threads like these when they are engaged in conversations with the people who have actually worked on the hardware in question is as shocking as it is pathetic.
Makes me wonder why they don't apply to design these machines themselves.
They're a joke.
The sheer nerve of certain individuals in threads like these when they are engaged in conversations with the people who have actually worked on the hardware in question is as shocking as it is pathetic.
Makes me wonder why they don't apply to design these machines themselves.
They're a joke.
kyliethicc
Member
Could it be 4 SE with 1 SA per SE? Perhaps. Doesn't look that way tho. Idk.
If you look at the slide with their GPU layout, it has 2 things labeled shader input. Those would seem to indicate 2 shader engines, idk.
The 2 groups of 28 CUs are also spaced further apart in 2 groups, which also seems clear that its 4 SA but 2 SE.
And btw, none of this is a criticism of the XSX. Just my thoughts and guesses.
Edit - another thought is ROPs:
Navi 21 die has 80 CUs in 4 Shader Engines, with 128 ROPs. 32 ROPS per SE.
Navi 22 die will have 40 CUs, 2 SEs, with 64 ROPs. 32 ROPs per SE.
PS5 GPU has 40 CUs, 2 SEs, with 64 ROPs. 32 ROPs per SE.
XSX GPU also has 64 ROPs. So .. it likely has 2 SEs and same 32 ROPs per SE.
Otherwise, it would have just 16 ROPs per SE if it had 4 SEs. Personally, I doubt that.
Thats not what that poster meant. I don't know or care what that youtuber guy said but Phil Spencer himself has said the XSX SoC was designed with the Azure silicon engineer team as well. That's why Azure engineers did the hot chips presentation. The chip was built for multiple purposes, console chip and xcloud server chip.
Thats why Xbox have said the XSX when used in xcloud servers can run 4 Xbox One games simultaneously from 1 SoC. Because it was built for that as well. Thats why in their recent presentation they discuss the flexibility of the CU design to serve both console & cloud server roles.
"Multiple uses" refers to use in console (XSX) and cloud (xcloud server.)
Xbox Cloud Gaming refers to when used in cloud servers. They only need 24 DCUs, 48 CUs total. Because Xbox One games use 12 CUs each. Each Xbox One has 7 DCUs on chip, 6 active. Thats why they built the XSX GPU to be 4 arrays of 7 DCUs. To be able to run 4 Xbox One games at once from 1 chip. Xbox One games use 5 GB RAM. For 4 games, the server chip would need 20 GB RAM. For the typical setup of 20 GB RAM, one would use either 10x 2 GB RAM chips or 20x 1 GB RAM chips, connected to a 320 bit interface. Thats most likely why the XSX console has the 320 bit bus but split bandwidth setup with 16 GB RAM. They wanted more bandwidth than 448 GB/s, and wanted the 320 bit bus on die for its secondary use in xcloud, but didn't want to spend the extra cash per console to put 20 GB RAM in each XSX. And using faster 16 Gb/s G6 RAM would have also cost more and or increased heat generated inside the console.
Plus, normally, any XSX chip that doesn't yield 52 CUs would end being wasted. But for using in xcloud, they only need 48, so they can actually save extra money and use more of the chips per wafer that they're paying TSMC to make for them. Its a clever way of building 1 chip that can serve 2 roles.
This is why once they determined they had 52 active CUs in the XSX, they just picked the clock speed to achieve 12 TFLOPS which Phil said was their goal all along. They did the same thing with XOneX, where they arrived at 40 active CUs and set the oddly specific clock of 1.172 GHz needed to get exactly 6.0 TFLOPs. And the same thing with XSS. Its a 20 CU GPU, and their goal was 4 TFLOPs, so they set clock to exactly 1.565 GHz to get exactly 4.0 TFLOPs.
Last edited:
There is the GPU and there are other parts of silicon on a die. He's speaking English, that's all. May be that's the problem.