Support NeoGAF

[ethomaz]

We need to get this clarified because it would be pretty interesting if so (although I am not sure how big the difference would be). Need to double check the HotChips presentation... that seems like an odd implementation of RDNA2 RT.

RT implementation at the die seems different but really need a RDNA2 die shot to comparison.
I believe nobody did a die shot from RX 6800 chips yet.

BTW where are you getting PS5 has more ROPs? They have the same... PS5 using the RDNA1 version with higher clock (more fill rate) and Series the RDNA 2 version with lower clock.

 

[Locuza]

Does this card not have Infinity Cache?
Also, it has a similar die layout to PS5.
Even the ROPs.
So PS5 ROPs RDNA 2?
I can't keep up with this thread.

Navi22 has the Infinity Cache/L3$ however the image is from Navi10.

The shader cores of the Xbox are also more suitable to machine learning, which could be an advantage if Microsoft succeeds in implementing an equivalent to Nvidia’s DLSS (an advanced neural network solution for AI).

Well they occur in Shader Cores. So the XSX has 97TOPs of INT4 and in theory the PS5 might have 82TOPs of INT4. The XSX has a compute advantage, so naturally it will have an advantage at this.

3.) As far as I know Sony did never state anything in relation to machine learning, though from my perspective the feature is too important to simply not mention it
97 vs 82. You can see the marketing issue with this? Especially if they are behind DirectML in SW, which is likely the case. To the average consumer 12 is not that much more than 10. They would likely shoot themselves in the foot.
Also notice how they did not reveal the Series S numbers either? That is because it is 32 TOPs of INT4.

Having more shader cores doesn't make the cores themselves more suitable for machine learning and as you pointed out the difference in theoretical throughput wouldn't be large anyway if the PS5 would support mixed precision dot-product instructions.
I don't think David Cage would have mentioned it if the PS5 would support those instructions.

In fact, for me infinity cache is not RDNA2 related, but it's navy 21 related. I'm pretty sure that with smaller GPU using RDNA 2, you'll have less IC quantity or no IC.

You may see it as an optional configuration detail.
The IC appears to be included on every upcoming discrete RDNA2 GPU but RDNA2 APUs will likely not have it.

Again:

I would be more conservative with the analyses being made on 9 Mpix photographs to interpret the lower level customizations that Sony or Microsoft applied in >12 Billion transistor chips.
Some things will be noticeable if there photos of previous chips with similar architectures to compare against, like approximate cache sizes, how many PHYs, number of WGPs and their redundancy, etc. Others will not be noticeable, because one pixel equates to over 1K transistors and the photo isn't perfectly focused.

That's not to say that Locuza and others aren't doing a spectacular job, but many things will be guesses.

I know it's a long video but it goes through multiple differences between RDNA1, 2 and the Xbox Series X:


The information is fully based on AMD's firmware files where parts of it can also be visually confirmed on the die shot.

In the video, also on this forum and on twitter I mentioned how I would classify the differences and technology.
In short, some differences can be simply seen as a configuration details or adjustments for better backwards compability or even performance, since the RDNA2 frontend setup has less primitive units than previously.
Neither console aligns to 100% to RDNA1 or RDNA2 GPUs from AMD.
Does that mean it's not "full RDNA2"?
This depends what someone sees as "full RDNA2", it's at least not identical to the way AMD builds their hardware and there are some differences but ultimately does that matter?

I stated the following:



PS: The RT implementation is the same between PS5, Xbox Series and RDNA2 PC hardware.

 

[Panajev2001a]

RT implementation at the die seems different but really need a RDNA2 die shot to comparison.
I believe nobody did a die shot from RX 6800 chips yet.

BTW where are you getting PS5 has more ROPs? They have the same... PS5 using the RDNA1 version with higher clock (more fill rate) and Series the RDNA 2 version with lower clock.

Yeah, same ROPS count, but not sure they are from different generations: XSX has the standard 64 pixels per clock output that is expected out of PS5 while AMD quoted higher higher throughput for RDNA2, but I might be wrong again.

 

[assurdum]

Such details are unfortunately not visible on die shots, at least not without a massive zoom factor.

I base it of a belly feeling because of some factors.

1.) Rosario Leonardi from Sony made the (now deleted) claim on twitter that there is "no ML stuff" on it:
https://web.archive.org/web/20200715221634/https://twitter.com/ilMal3/status/1283524232933187593

2.) In an Interview David Cage from Quantic Dream claimed that the Xbox Series shader cores are more suitable for machine learning:
https://wccftech.com/david-cage-on-quantic-dream-future-next-gen-gaming-and-not-ruling-out-sequels/

3.) As far as I know Sony did never state anything in relation to machine learning, though from my perspective the feature is too important to simply not mention it.
The same lack of information was present on the VRS side of things which is now fairly certainly not supported on the PS5 hardware because of the visibly older Render Backend design on the die shot.

4.) The graphics core IP version is 10.0 according to older test leaks on GitHub from an AMD employee.
The first Navi GPU (RDNA1) was Navi10 which has the GC version 10.1.0 and it's not supporting dot-product instructions for INT8/4.
While ray tracing acceleration was included on the PS5, I'm bearish on the perspective that Sony also touched the ALUs to include ML instructions.

It's my personal perspective in that regard and not pure evidence.


I missed the interesting GitHub leak in full glory before it was taken down but all the information from that time is legit as far as I saw.

To be fair Mark Cerny mentioned machine learning as part of the ps5 hardware in the road of the ps5...

 

[ethomaz]

Yeah, same ROPS count, but not sure they are from different generations: XSX has the standard 64 pixels per clock output that is expected out of PS5 while AMD quoted higher higher throughput for RDNA2, but I might be wrong again.

I believe the differences AMD listed are these ones:

They increased the number of colors per channel.
They can now process 8x 32bits pixels or 4x 64bits pixels at same time... RDNA can do only 4x 32bits pixels or 2x 64 bits pixels at same time.

16, 32 and 64bits are how many colors the pixels can have... think like the difference when you choose in your monitor 16bits colors or 32bits colors.

Edit - AMD specs page (https://www.amd.com/en/products/specifications/graphics) put RX 6800 at "Up to 202.10 GP/s" that means 96ROPs * Up to 2105Mhz... so I don't thing the color deep matter for fill rate.

 

[onesvenus]

That is not correct from what the GPU die shot and drivers showed.... that is the main point why I asked.

We know the GPU is 100% RDNA2 functional... in simple terms it can to everything RDNA2 card can... imo Infinity Cache doesn't affect that because it is only a performance feature and not a functional.

Said that the we are discussing the GPU silicon design that are not 100% RDNA2... that is what we were asking to him to understand why they choose old parts instead the new RDNA2 ones.

So he answer your question saying it's 100% RDNA2 and you choose not to believe him. Ok.

 

[SirTerry-T]

Got to respect him for having some mighty balls to confront the console maker heads on.
Should’ve done it 37 pages ago though and used the time to play some games instead. I’m playing Amiga right now, zero CUs, 7MHz CPU, still higher and more stable fps on some games than what we have on XSX and PS5

Motorola 68000, Paula, Denise and Agnus>RDNA 2

 

[HoofHearted]

His answers.



Another answer to another guy.



He never touched what was asked.
He keep saying the SoC has Zen2 and other things lol

Had more coffee and read up a bit - A couple of notes:

• I think you're a bit confused and you're trying to correlate a "standard" AMD RDNA2 reference GPU and it's components to the "custom" XSX SoC containing the "specialized [RDNA2] configuration"
  • There's a significant difference here with respect to an SoC versus a dedicated GPU design/architecture
• You're talking to very technical-minded engineer:
  • He's being very direct and clear in his responses to you
  • He clearly is in a position to fully know and understand the specifications of the XSX GPU
  • Yet:
    • You're asking him to go and watch some random guy's video that draws and makes key assumptions based and sourced on random captured technical documents and hardware version details
    • Then you're asking him to explain further on your statement "even the GPU parts that are not RNDA2" when he clearly responded to you with "Series S|X GPUs are 100% RDNA 2. But of course "parts of the silicon" are Zen 2, and other parts are entirely custom."
    • Caveat: (no offense here to Locuza - as the video is certainly a decent and well thought out attempt to gain additional insight)

Your question was asked and clearly answered directly - yet your stance now is that he's doing "PR" (which isn't in his job description).

 

[Elog]

PS: The RT implementation is the same between PS5, Xbox Series and RDNA2 PC hardware.

Thanks for your thoughts on these matters Locuza. Short question: If I add CU mm2 plus TMU mm2 and divide by number of CUs from the die shots (PS5 and XSX), I get a larger mm2 number for PS5 than for XSX that confuse me. Once you break it down it seems to stem from the TMU/RT side rather than the CU side. Have you looked into this? I have seen three estimates of mm2 from the die shots and they all (within the margin of error) come to the same conclusion regarding mm2.

Edit: The same conclusion is true for the entire ROP/Primitive/GE part as well - seemingly significantly larger on the PS5 die shots. Not sure what logic that drives that difference in size either.

 

[Panajev2001a]

I believe the differences AMD listed are these ones:

They increased the number of colors per channel.
They can now process 8x 32bits pixels or 4x 64bits pixels at same time... RDNA can do only 4x 32bits pixels or 2x 64 bits pixels at same time.

16, 32 and 64bits are how many colors the pixels can have... think like the difference when you choose in your monitor 16bits colors or 32bits colors.

Edit - AMD specs page (https://www.amd.com/en/products/specifications/graphics) put RX 6800 at "Up to 202.10 GP/s" that means 96ROPs * Up to 2105Mhz... so I don't thing the color deep matter for fill rate.

XSX processes 64 pixels per clock: ~116 GPixels/s / 1.825 GHz = ~64 pixels/clock so either they have less RB’s and a lower clockspeed or have the same number of RB’s and a lower throughput per clock and a lower clockspeed

 

[ethomaz]

XSX processes 64 pixels per clock: ~116 GPixels/s / 1.825 GHz = ~64 pixels/clock so either they have less RB’s and a lower clockspeed or have the same number of RB’s and a lower throughput per clock and a lower clockspeed

I mean 64 ROPS * 1825Mhz = 116.8 Gpix/s.

 

[Panajev2001a]

PS: The RT implementation is the same between PS5, Xbox Series and RDNA2 PC hardware.

I am trying to find detailed references on this aspect for RDNA 2 desktop cards by AMD vs XSX and PS5: can the desktop cards process ray intersections in parallel to Texture ops?

I am looking again at in rete data for PS5, but it looks like XSX cannot (likely PS5 cannot either, but I do not know for sure).

 

[Panajev2001a]

I mean 64 ROPS * 1825Mhz = 116.8 Gpix/s.

I get that, XSX and PS5 seem to have the same throughput though. 64 ROPS for PS5, 64 pixels/cycle for XSX.

 

[Greggy]

Had more coffee and read up a bit - A couple of notes:

• I think you're a bit confused and you're trying to correlate a "standard" AMD RDNA2 reference GPU and it's components to the "custom" XSX SoC containing the "specialized [RDNA2] configuration"
  • There's a significant difference here with respect to an SoC versus a dedicated GPU design/architecture
• You're talking to very technical-minded engineer:
  • He's being very direct and clear in his responses to you
  • He clearly is in a position to fully know and understand the specifications of the XSX GPU
  • Yet:
    • You're asking him to go and watch some random guy's video that draws and makes key assumptions based and sourced on random captured technical documents and hardware version details
    • Then you're asking him to explain further on your statement "even the GPU parts that are not RNDA2" when he clearly responded to you with "Series S|X GPUs are 100% RDNA 2. But of course "parts of the silicon" are Zen 2, and other parts are entirely custom."
    • Caveat: (no offense here to Locuza - as the video is certainly a decent and well thought out attempt to gain additional insight)

Your question was asked and clearly answered directly - yet your stance now is that he's doing "PR" (which isn't in his job description).

If relentlessly pushing falsehoods officially labelled as falsehoods by the utmost authority on the subject isn't ban worthy console warring, what is?
I'm not even gonna dignify his posts by reporting them. Someone alert the mods of his shameless assaults on the truth. I thought we saw the danger of mechanically repeated lies not too long ago in DC.
And dude (I don't want to qrite his name), have the decency to stop harassing Xbox engineers on twitter. The time they waste pushing back against your propaganda is time stolen away from making a better plaform for all us.

 

[MasterCornholio]

I get that, XSX and PS5 seem to have the same throughput though. 64 ROPS for PS5, 64 pixels/cycle for XSX.

Weird I thought the PS5 had a throughput advantage due to the higher clocks.

 

[ethomaz]

I get that, XSX and PS5 seem to have the same throughput though. 64 ROPS for PS5, 64 pixels/cycle for XSX.

No. PS5 has 64 * up to 2230Mhz = 142.72 Gpix/s.

I understood now.

The RDNA 2 pic is a full chip with 8RB.... 8RB can process 8 32bits pixels per cycle... while the RDNA it is 4RB that can process 4 32bits pixels per cycle.
The increase the number of units from RDNA to RDNA 2.

Each RB is a combination of 16ROPs that combined can run one 32bits Pixels per cycle.
I belive fill rate is probably based on Pixels that are 8bits? Can somebody confirm? Each ROP can do 1x 8bits pixels per cycle and you need 16x ROPs (1x RB) to run a 32bits pixel.

Series X has 4 RBs like RDNA but these RBs are RB+ that have hardware VRS support.

Weird I thought the PS5 had a throughput advantage due to the higher clocks.

It has.

 

[Die Namek Ability]

Had more coffee and read up a bit - A couple of notes:

• I think you're a bit confused and you're trying to correlate a "standard" AMD RDNA2 reference GPU and it's components to the "custom" XSX SoC containing the "specialized [RDNA2] configuration"
  • There's a significant difference here with respect to an SoC versus a dedicated GPU design/architecture
• You're talking to very technical-minded engineer:
  • He's being very direct and clear in his responses to you
  • He clearly is in a position to fully know and understand the specifications of the XSX GPU
  • Yet:
    • You're asking him to go and watch some random guy's video that draws and makes key assumptions based and sourced on random captured technical documents and hardware version details
    • Then you're asking him to explain further on your statement "even the GPU parts that are not RNDA2" when he clearly responded to you with "Series S|X GPUs are 100% RDNA 2. But of course "parts of the silicon" are Zen 2, and other parts are entirely custom."
    • Caveat: (no offense here to Locuza - as the video is certainly a decent and well thought out attempt to gain additional insight)

Your question was asked and clearly answered directly - yet your stance now is that he's doing "PR" (which isn't in his job description).

Seems like the main point of contention is, why are these generalizations ok for MS but not when Cerny states the same for PS? This question isnt directed at you specifically just the audience as a whole. To me it seems like its really just the console warriors that want to label these custom systems as RDNA 1.xxxx or what ever so they can have some imaginary advantage over the other as long as thier number is higher its all good. Bottom line both systems are custom RDNA2 with tweaks that suit thier own specific needs, anything else is speculation at best.

 

[Panajev2001a]

Weird I thought the PS5 had a throughput advantage due to the higher clocks.

It does... the throughput per clock is identical, but PS5 has higher clocks (you may hit bandwidth bottlenecks first perhaps).

 

[DForce]

So let me get this straight.
He is neither a GPU Engineer or the Engineer of XBSX. He just work at Microsoft.
Mark Cerny the Engineer of the PS5, provides break downs of the PS5, you don't believe him and says he's lying.
While when that guy is asked to provide information, he says he's an expert witness.
Lmao, how does he prove anything. He didn't even provide anything meaningful, just regurgitate the same PR from Microsoft.

It's pretty funny, huh?







So if the old leaked information shows it doesn't support HW RT, then Mark Cerny is not credible.

 

[Panajev2001a]

No. PS5 has 64 * up to 2230Mhz = 142.72 Gpix/s.

I understood now.

The RDNA 2 pic is a full chip with 8RB.... 8RB can process 8 32bits pixels per cycle... while the RDNA it is 4RB that can process 4 32bits pixels per cycle.
The increase the number of units from RDNA to RDNA 2.

Each RB is a combination of 16ROPs that combined can run one 32bits Pixels per cycle.
I belive fill rate is probably based on Pixels that are 8bits? Can somebody confirm? Each ROP can do 1x 8bits pixels per cycle and you need 16x ROPs (1x RB) to run a 32bits pixel.

Series X has 4 RBs like RDNA but these RBs are RB+ that have hardware VRS support.


It has.

It is possible that PS5 has more ROPS (1 ROP = 1 Pixel per cycle) and XSX has less ROPS at double throughput per cycle... the throughput per cycle would still be the same across consoles. One just runs at higher clocks.

 

[AudioSlave92210]

Motorola 68000, Paula, Denise and Agnus>RDNA 2

Smooth scrolling with the the blitter and the copper.
Good memories of assembling code with k-seka and later with the devpac. I was a teenager.
And the atari st/ amiga war...

 

[Loope]

It's pretty funny, huh?







So if the old leaked information shows it doesn't support HW RT, then Mark Cerny is not credible.

You're using tweets from 2019 and 2020 to basically discredit someone that already admitted in this very thread he was completely wrong to somehow prove that a person that works directly with a system is regurgitating PR as opposed to the sony fanboy that jumps on the 1st oportunity to shit on MS.

Good show.

By the way if neither of people here worked on a system and someone that did says something about the system, then it is automatically true. Just shut the fuck up, you don't have the knowledge or capability to do something like this. It doesn't matter if is Sony or MS, if they come out and say that something is done that way, then it is a fact. No point in console warriors trying to deny factual information.

 

[DonkeyKong]

I never said the GPU ARE RDNA1.

I said and it is confirmed the GPUs have RDNA1 parts and are not full based in RDNA 2.

Please, just stop now!!

 

[DForce]

You're using tweets from 2019 and 2020 to basically discredit someone that already admitted in this very thread he was completely wrong to somehow prove that a person that works directly with a system is regurgitating PR as opposed to the sony fanboy that jumps on the 1st oportunity to shit on MS.

Good show.

By the way if neither of people here worked on a system and someone that did says something about the system, then it is automatically true. Just shut the fuck up, you don't have the knowledge or capability to do something like this. It doesn't matter if is Sony or MS, if they come out and say that something is done that way, then it is a fact. No point in console warriors trying to deny factual information.

No.

I just find it funny that someone using the, "Microsoft said, so it's confirmed" doesn't apply to Sony and he has a history of calling Mark Cerny a liar over and over again. There's nothing wrong to express doubt if something gives you a reason to believe so, but that belief can also come back to bite you in the a$$ later on.

 

[Lysandros]

Weird I thought the PS5 had a throughput advantage due to the higher clocks.

It is. PS5 has %22 higher pixel fillrate compared to XSX.

 

[Loope]

No.

I just find it funny that someone using the, "Microsoft said, so it's confirmed" doesn't apply to Sony and he has a history of calling Mark Cerny a liar over and over again. There's nothing wrong to express doubt if something gives you a reason to believe so, but that belief can also come back to bite you in the a$$ later on.

Agree, but he did admit he was wrong in this very thread. It's ridiculous that people on the top of the game (working for Sony, MS or Nintendo) are called out on technical stuff. There is no reason whatsoever to doubt Mark Cerny or any other engineer working for these companies, plus insinuating PR talk from people that have way more knowledge that they dream of.

 

[MasterCornholio]

It is. PS5 has %22, higher pixel fillrate compared to XSX.

That's what I've read.

 

[ethomaz]

It is possible that PS5 has more ROPS (1 ROP = 1 Pixel per cycle) and XSX has less ROPS at double throughput per cycle... the throughput per cycle would still be the same across consoles. One just runs at higher clocks.

It is the same.

That double is because RDNA 2 has double RBs but Series X not.

64 ROPS * 1825Mhz = 116.8 Gpix/s... but these are 8bits pixels.... 64 ROPs means 64 8bits pixels per cycle in both.

For 32bits color pixels you need 1 RB (16ROPs).

 

[Locuza]

To be fair Mark Cerny mentioned machine learning as part of the ps5 hardware in the road of the ps5...

I did not catch that in relation to the GPU, if time permits I may check it out later.

Thanks for your thoughts on these matters Locuza. Short question: If I add CU mm2 plus TMU mm2 and divide by number of CUs from the die shots (PS5 and XSX), I get a larger mm2 number for PS5 than for XSX that confuse me. Once you break it down it seems to stem from the TMU/RT side rather than the CU side. Have you looked into this? I have seen three estimates of mm2 from the die shots and they all (within the margin of error) come to the same conclusion regarding mm2.

Edit: The same conclusion is true for the entire ROP/Primitive/GE part as well - seemingly significantly larger on the PS5 die shots. Not sure what logic that drives that difference in size either.

I didn't looked at those comparisons yet, outside of the central block which has the command frontend, geometry processor, etc.
It's quite a bit larger than on Navi10 but I don't know if simply some blocks where moved or if some blocks grew in size and function.

I am trying to find detailed references on this aspect for RDNA 2 desktop cards by AMD vs XSX and PS5: can the desktop cards process ray intersections in parallel to Texture ops?

I am looking again at in rete data for PS5, but it looks like XSX cannot (likely PS5 cannot either, but I do not know for sure).

From my understanding no, since texture or ray ops are going through the same hardware units.
One may also call both simply texture instructions.
The texture adress unit is either doing work for texture sampling or getting BHV data for further intersection tests, that's why it's worded as 4 Texture ops or 4 Ray ops.

The hw implementation is the same between AMD's RDNA2 cards, the PS5 or Xbox Series.

-----

On a side note.

The Xbox Series has in total 8 Render Backends (4RBs per Shader Engine).
Each Render Backend has 8 Color ROPs -> 64 Color ROPs in total.

The PS5 has in total 16 active Render Backends, however each RB has only 4 Color ROPs --> also 64 Color ROPs in total.

As a visual example, the PS5 Render Backend configuration should look like on RDNA1.
(Some may notice that the new Render Backend design leads to less Depth ROPs per Shader Engine).

Nemez did a more precise die shot annotation based on a higher resolution image:

 

[Superayate]

On a side note.

The Xbox Series has in total 4 Render Backends.
Each Render Backend has 8 Color ROPs -> 64 Color ROPs in total.

The PS5 has in total active 8 Render Backends, however each RB has only 4 Color ROPs --> also 64 Color ROPs in total.

As a visual example, the PS5 Render Backend configuration should look like on RDNA1.
(Some may notice that the new Render Backend design leads to less Depth ROPs per Shader Engine).

That could explain why when I was watching the XsX die, I found and thought they were 2 times less Render backend than it is the case on PS5. I have never seen the same detailed floor plan for the XsX and XsS as it is done for the PS5
A new time, thanks for sharing your advice and for these explanations.

 

[BeardGawd]

lol no you didn't!

He added one more piece.

Is that our very own ethomaz being rude to a principle architect like that? Dude should be ashamed of himself constantly spreading disinformation.

 

[Retinoid]

ITT: Dunning-Kruger

 

[ethomaz]

I did not catch that in relation to the GPU, if time permits I may check it out later.

I didn't looked at those comparisons yet, outside of the central block which has the command frontend, geometry processor, etc.
It's quite a bit larger than on Navi10 but I don't know if simply some blocks where moved or if some blocks grew in size and function.

From my understanding no, since texture or ray ops are going through the same hardware units.
One may also call both simply texture instructions.
The texture adress unit is either doing work for texture sampling or getting BHV data for further intersection tests, that's why it's worded as 4 Texture ops or 4 Ray ops.

The hw implementation is the same between AMD's RDNA2 cards, the PS5 or Xbox Series.

-----

On a side note.

The Xbox Series has in total 8 Render Backends (4RBs per Shader Engine).
Each Render Backend has 8 Color ROPs -> 64 Color ROPs in total.

The PS5 has in total 16 active Render Backends, however each RB has only 4 Color ROPs --> also 64 Color ROPs in total.

As a visual example, the PS5 Render Backend configuration should look like on RDNA1.
(Some may notice that the new Render Backend design leads to less Depth ROPs per Shader Engine).

Nemez did a more precise die shot annotation based on a higher resolution image:

This made me confuse now.

How many RBs there is in PS5? The pic shows 5 per Shader Engine.

Edit - The pic is cut... it is not showing half of the units... so there is 8 per Shader Engine... I understood now.

 

[Epic Sax CEO]

People still insisting on "it's RDNA1" (because they can say it's worse than theirs), but the problem on the other thread was me for losing patience.

 

[Clear]

This is all pointless until we have a consensus definition of what exactly constitutes a RDNA2 GPU.

We don't have this and its a kinda tricky thing to parse observationally as different devices within the RDNA2 branding are going to have different layouts and capabilities.

I'd also caution against over-valuing functions being implemented in silicon as being invariably better or more performant than software-based solutions to the same thing. That's not necessarily the case at all.

 

[SenjutsuSage]

He still keep spinning the PR marketing because the silicon has old non-RDNA2 para jn GPU.

Dodged the question 3 times and trying to talk about SoC lol

He didn't dodge once. The entire silicon can't be RDNA 2, cause the entire silicon isn't entirely made up of GPU. The GPU portion for both X and S, he confirmed, is 100% RDNA 2. He said other parts of the "silicon" or System on a chip are Zen 2, other parts are entirely custom.

He also directly challenged your claim that the compute units in Series X are RDNA 1. He said you were wrong specifically about the compute units, and pretty much emphasized his position at Xbox along with his role as someone who would be considered an expert witness on the matter, and on that subject we know for certain he almost definitely is. And if we're being fair, those compute units couldn't possibly be RDNA 1 yet pack all hardware features present in RDNA 2. The evidence is all there against them being RDNA 1. There are no major new graphics features missing. On top of that, there was no IPC improvement for the compute units from RDNA 1 to RDNA 2. The only IPC gain came entirely from Infinity Cache, but the Compute Units in RDNA 1 and RDNA 2 without Infinity Cache get the same amount of work done in the same amount of time.

Infinity Cache is a further separate hardware addition that is more about power consumption, avoiding the use of more expensive RAM and to keep from requiring a bigger memory bus. James Stanard at Microsoft, someone integral in all things relevant to Series X|S at a SoC and architecture, graphics tech, optimization, lighting, compression tech level, has put it to rest. And he did it outside of a planned disclosure, he told us directly when asked.

The base of Xbox Series X|S is pure RDNA 2. They didn't start with an RDNA 1 base. It was RDNA 2 from the start and then injected more IP and made decisions that fit with their power/space/cost constraints and performance goals.

 

[M1chl]

This is the most pointless discussion on gaf I've seen. Since both consoles are custom design. Based on probably probably same point of roadmap which AMD had... Btw Infinite cache is on desktop, because of thinner bus width, it simply does not make sense for wider bus found on for example Xbox. Even 6900XT has 256-bit bus width...

Makes me think that Xbox One OG had same type of "crutch"...

 

[Rob_27]

I don't know why there are people bombarding the likes of James Stanard etc to get silly answers.

Leave them alone they have jobs.

I have nothing against constructive investigation but when there are loads of people bombarding him and others not just on this site. I suppose he could block people but shouldn't have to.

Anyway back to discussion.

 

[Lysandros]

This is all pointless until we have a consensus definition of what exactly constitutes a RDNA2 GPU.

We don't have this and its a kinda tricky thing to parse observationally as different devices within the RDNA2 branding are going to have different layouts and capabilities.

I'd also caution against over-valuing functions being implemented in silicon as being invariably better or more performant than software-based solutions to the same thing. That's not necessarily the case at all.

Sebastion Aaltonen (well known game developer, Sebbi from Beyond3d) has a new twitter post saying 'software' VRS (using compute shaders) is better with deferred renderer game engines. As far as i know the majority of game engines use deferred rendering. He says that fixed function VRS is generally better with forward rendering. Maybe someone can post the original twitter statement here, i don't have an account.

Edit: Sorry Clear, i mistakenly quoted you.

 

[Trainsmove]

PS5 CUs RDNA 1.x?
For Hardware Backwards compatibility. (Not software.)

 

[FatKingBallman]

He didn't dodge once. The entire silicon can't be RDNA 2, cause the entire silicon isn't entirely made up of GPU. The GPU portion for both X and S, he confirmed, is 100% RDNA 2. He said other parts of the "silicon" or System on a chip are Zen 2, other parts are entirely custom.

He also directly challenged your claim that the compute units in Series X are RDNA 1. He said you were wrong specifically about the compute units, and pretty much emphasized his position at Xbox along with his role as someone who would be considered an expert witness on the matter, and on that subject we know for certain he almost definitely is. And if we're being fair, those compute units couldn't possibly be RDNA 1 yet pack all hardware features present in RDNA 2. The evidence is all there against them being RDNA 1. There are no major new graphics features missing. On top of that, there was no IPC improvement for the compute units from RDNA 1 to RDNA 2. The only IPC gain came entirely from Infinity Cache, but the Compute Units in RDNA 1 and RDNA 2 without Infinity Cache get the same amount of work done in the same amount of time.

Infinity Cache is a further separate hardware addition that is more about power consumption, avoiding the use of more expensive RAM and to keep from requiring a bigger memory bus. James Stanard at Microsoft, someone integral in all things relevant to Series X|S at a SoC and architecture, graphics tech, optimization, lighting, compression tech level, has put it to rest. And he did it outside of a planned disclosure, he told us directly when asked.

The base of Xbox Series X|S is pure RDNA 2. They didn't start with an RDNA 1 base. It was RDNA 2 from the start and then injected more IP and made decisions that fit with their power/space/cost constraints and performance goals.

Since your reference for XSX absolute power over PS5 is Hitman 3, so, i guess, you believe Digital Foundry, right?

"now look at the basics here of the dual compute unit setup, RT support aside, this is essentially RDNA as it is presented in Radeon RX 5700 range..."

at 12:50

 

[AgentP]

ITT: Dunning-Kruger

Seriously, 38 pages of this crap, embarrassing.

 

[longdi]

Sebastion Aaltonen (well known game developer, Sebbi from Beyond3d) has a new twitter post saying 'software' VRS (using compute shaders) is better with deferred renderer game engines. As far as i know the majority of game engines use deferred rendering. He says that fixed function VRS is generally better with forward rendering. Maybe someone can post the original twitter statement here, i don't have an account.

Edit: Sorry Clear, i mistakenly quoted you.

i believe SX can have a choice to do sw or hw vrs, since the updated rops just added more hw stuff.

 

[longdi]

I guess if i were to write a guesstimate, here is how both consoles were created.
It seems MS had a bigger R&D budget, naturally being M$ and Phil desire to win the hw race.

Both consoles had to use whatever Amd have for 2020 launch, and that is rDNA2 base, the next step up from rDNA1. Nvidia is too rich and stuckup for them.
So with the rDNA2 base, both then spend the rest of the budget to add extra sauces to it.

Sony chose to spend all of it on creating a forward looking SSD IO design. They were happy enough with base rDNA2 performance, and whatever cool graphics effects, they decided to do it via software and having a strong direct accessible api.

MS also spent on SSD IO design, but not as far, just good enough for current gen hardware. They do have deeper pockets, and took the bigger rDNA2 die design, and worked further with AMD to add extra forward looking efficiency hw around it. -ML, SFS, VRS, Mesh Shaders. These also got use into Amd rDNA2 gpu, so a welcome collab effort.

I think thats all to it.

SX is probably a more rounded device at the same $499 SRP. However its api is either lacking or more bloated (because of long term BC through VM).

Here is hoping those extra hw accelerators it got, do get used over time, and pad over its heavier api layer. Ps5 is punching above its weight and it is MS job to recover asap, if it can, a must.

I mean Phil got his hw win, but he needs to look into their api and show the design wins in actual games.

 

[Die Namek Ability]

i believe SX can have a choice to do sw or hw vrs, since the updated rops just added more hw stuff.

Do you think there is a possibility it could do both? That would be pretty neat. Not sure if even possible but i would imagine that would be a huge boon for MS

 

[Tickle My Tendies]

I am just glad ethomaz was there to put that dork in his place. I mean, so what if he engineered the system. What the fuck does he know?

GAF Engineers build reality. All that Xbox loser did was build a console.

 

[yurinka]

Neither console aligns to 100% to RDNA1 or RDNA2 GPUs from AMD.
Does that mean it's not "full RDNA2"?
This depends what someone sees as "full RDNA2", it's at least not identical to the way AMD builds their hardware and there are some differences but ultimately does that matter?

AMD, Sony and MS said both consoles are custom RDNA2.

This means that starting from RDNA2, they add, remove or change stuff they consider for their insterests: like to enhace or to add something they consider important for them, to cut something they won't use to reduce costs, to replace something for their own implementation because they think it will benefit them, etc.

Real world results we saw in native next gen games show pretty similar results between both consoles, sometimes slightly better in one of them (more frequently on the PS5) and pretty similar to more or less equivalent RDNA2 stuff. Which leads to think they both are RDNA2 and if they removed something it's because they don't need it or because they replaced it for something better for them.

 

[longdi]

Do you think there is a possibility it could do both? That would be pretty neat. Not sure if even possible but i would imagine that would be a huge boon for MS

I don't see why not. Somewhere in this thread, it shows the difference between rDNa1 and rDNA2 rops, are the extra hardware to double up the speed of rops and extra vrs tier 2. Things were improved and not replaced

 

[Garani]

I am just glad ethomaz was there to put that dork in his place. I mean, so what if he engineered the system. What the fuck does he know?

GAF Engineers build reality. All that Xbox loser did was build a console.

Same GAF engineers that state that PS5 is based on RDNA1.

 

[SenjutsuSage]

Since your reference for XSX absolute power over PS5 is Hitman 3, so, i guess, you believe Digital Foundry, right?



at 12:50

Yea, but the basic arrangement doesn't tell you anything because there wasn't much of a visible major change in the basic compute unit design from RDNA 1 to RDNA 2. But what we do know is that there were a lot of other changes around that, which is what has enabled newer features and power optimizations. And as of HotChips in August RDNA 2 hadn't been fully detailed by AMD yet to begin with, so DF went with what they best understood then. At that time they didn't know whether to expect something that looked the same or totally different.

Since then, however, we've learned quite a bit more, and we now have a Microsoft engineer in James Stanard putting any beliefs of RDNA 1 portions in the GPU to rest with comments that are mere days old, and he was specifically challenged on the compute units.

And as to Hitman 3 on XSX vs PS5, those are just what the results told us. I don't exactly blindly follow DF, but I do trust DF's analysis and their sources, even if sometimes I take issue with how they do things.