• 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.

Xbox Series X’s Advantage Could Lie in Its Machine Learning-Powered Shader Cores, Says Quantic Dream

longdi

Banned
Never losing a chance to take a conversation and making it an astroturfing piece (see bolded for a little gem there) ;). Enjoy your psychology 101 experiments in attempting to piss PS fans offs :p.

With that said, I do find your attempt to suggest PS5 will not improve over time as developers get more and more familiar with the HW cute though.

Having good first day DevTools does not mean there is no room for improvement or that devs experiments with HW do not extract more performance over time (especially if MS does plan to release more HW iterations more frequently and developers are disincentivized from optimising for any particular one as a result). That is the point of fixed specs console generations: single spec available to ALL the users and developers being able/having good RoI incentives to spend time to find novel and crazier ways to make it sing.

Not sure why you think i cared about ps fans in my post. 🤷‍♀️

All i was saying is that XSXS have a lot of potential to be tapped. The launch games and over the next 1 year, are just a small preview of what can be done on the system. The modern efficiency features are going to show its power over time.

That's what i meant earlier about the fine wine engineering of the systems
 
Last edited:

Thirty7ven

Banned
I will believe this when another site corroborates it. It’s looking likely that it might be true, but using this trash ass site as a source? Nah
 

turtlepowa

Banned
Overall, I think that the pure analysis of the hardware shows an advantage for Microsoft, but experience tells us that hardware is only part of the equation: Sony showed in the past that their consoles could deliver the best-looking games because their architecture and software were usually very consistent and efficient.

And this will repeat in the new gen.
We are talking about let's say a handful of high budget eyecandy games like TLOU2 while One x had the lead in most crossplatform games. Next Gen MS will have a shitload of first party games and i wouldn't be too sure that they are all ugly, i bet Flight Simulator will be sexy as hell on Series X.
 
Last edited:
The Nvidia silicon has far greater die area available for such tasks though. So while MS might be able to do something in this area compared to Sony, I wouldn’t expect any magic.
 

Panajev2001a

GAF's Pleasant Genius
Not sure why you think i cared about ps fans in my post. 🤷‍♀️

All i was saying is that XSXS have a lot of potential to be tapped. The launch games and over the next 1 year, are just a small preview of what can be done on the system. The modern efficiency features are going to show its power over time.

That's what i meant earlier about the fine wine engineering of the systems

There are efficiency features on both systems and sure we can expect GDK improvements will result in XSX games improvements, but it still remains a more generic dev approach meant for multiple platform reach over single platform efficiency (you are winking and nudging at the PS5 SDK and efficiency being set in stone and tapped out at launch... PSVita, PS4, and PS4 Pro have a history of balancing evolution and easy to code for at launch with revolution and elements to give longevity to the platform which is the fine balancing job where proper architecture does shine... I suspect PS5 will too).

All the other PR buzzwords (sustained SoC :LOL:) are there for console war stirring and getting the other lot mad ;).
 
Last edited:

Thirty7ven

Banned
The Nvidia silicon has far greater die area available for such tasks though. So while MS might be able to do something in this area compared to Sony, I wouldn’t expect any magic.

It has less INT cap than a 2060 but watch some of these peeps act like this is 3090 territory.
 

Njocky

Banned
Its likely true, question is what ML tasts is the correct question?

Procedural generation is now in Gerometry engine, and Digital foundry could not even tell if Spiderman was upscaled, and the we will all get the new AMD supersampling if Sony dont have their own of course.

It also confirms the shader cores and architecture are NOT the same, and Ps efficiencies are not the same.

Da da da

I was sure there was a way to spin this into a win for Sony.
 

geordiemp

Member
I was sure there was a way to spin this into a win for Sony.

Whats the win exactly ? ML is not used to give performance benefits on console games currently.

If dedicated cores are necessary for future techniques, then XSX will have a benefit here.

If Sony BELEIVES they not need dedicated ML cores for future, then they wont waste space using them.

MS likley needs more ML cores for server application anyway. So dedicatd ML is not wasted on XSX server requirements, its needed.

The result is unknown = future. So there is no win.
 
Last edited:

rnlval

Member
David Cage, CEO and founder of Quantic Dream, highlighted the Xbox Series X's shader cores as more suitable for machine learning tasks, which could allow the console to perform a DLSS-like performance-enhancing image reconstruction technique.

XSX GPU supports INT4 and INT8 tensor style workloads on the stream processors.

Yes, some aspects of the PS5 GPU are 20% faster. Pixel fillrate, rasterization rate, caches.
XSX GPU has 25% more L2 cache, 44% more LDS(local data storage), 44% more L0 cache, 44% more L1 cache, 44% more wavefront queue. More on-chip very fast SRAM storage leads to lower latency and fewer pipeline stalls.

Pixel fillrate is mostly bound by memory bandwidth and L2 cache. ROPS discussion is pointless without memory bandwidth discussion.

RX 6800 has a higher CU count(60 CU)/higher CU SRAM storage, high clock speed, 96 ROPS, and 128 MB very fast L3 cache.

RX 6800 can crush both XSX (52 CU at 1825 Mhz base clock) and PS5 (36 CU up to 2230Mhz).
 

Panajev2001a

GAF's Pleasant Genius
Whats the win exactly ? ML is not used to give performance benefits on console games currently.

If dedicated cores are necessary for future techniques, then XSX will have a benefit here.

If Sony BELEIVES they not need dedicated ML cores for future, then they wont waste space using them.

MS likley needs more ML cores for server application anyway. So dedicatd ML is not wasted on XSX server requirements, its needed.

The result is unknown = future. So there is no win.

The thing is that MS does not have dedicated ML cores in XSX like the tensor cores. They are taking their current vector ALU’s and splitting the 32 bit lanes into four 8 bits ones or eight 4 bits ones for parallel INT4/INT8... extending rapid packed math from FP16 which XOX fans derided (but now enjoy as XSX has this feature too) to INT4 and INT8. It does not have specialised HW like you can find on Google’s TPU, the Tensor cores on nVIDIA’s GPU’s, or Apple’s Neural Engine.

You want to run faster speed INT4/8 code you can and that portion of code will run at an increased throughput (4x as many INT4 operations as FP32 ones per CU), but you are using the CU’s you also need for shading and async compute.

Beyond FP16 RPM support we do not know if Sony has any similar enhancements as of yet.
 
Last edited:

Panajev2001a

GAF's Pleasant Genius
XSX GPU supports INT4 and INT8 tensor style workloads on the stream processors.


XSX GPU has 25% more L2 cache, 44% more LDS(local data storage), 44% more L0 cache, 44% more L1 cache, 44% more wavefront queue. More on-chip very fast SRAM storage leads to lower latency and fewer pipeline stalls.

Pixel fillrate is mostly bound by memory bandwidth and L2 cache. ROPS discussion is pointless without memory bandwidth discussion.

RX 6800 has a higher CU count(60 CU)/higher CU SRAM storage, high clock speed, 96 ROPS, and 128 MB very fast L3 cache.

RX 6800 can crush both XSX (52 CU at 1825 Mhz base clock) and PS5 (36 CU up to 2230Mhz).

Counting L2 and LDS is highly misleading as you are double counting the CU’s this way. Those resources are there because you have more CU’s and are needed to sustain the CU itself.

I cannot say I have two cores compared to your single core setup and also float about having more L1 cache and registers... c’mon...
 

rnlval

Member
Counting L2 and LDS is highly misleading as you are double counting the CU’s this way. Those resources are there because you have more CU’s and are needed to sustain the CU itself.

I cannot say I have two cores compared to your single core setup and also float about having more L1 cache and registers... c’mon...

LDS and L2 cache is not the same memory pool.

910fcae9-f40b-4ba1-9684-475d9dd4e46a.PNG
 

Panajev2001a

GAF's Pleasant Genius
so he basically confirms no 4bit and 8bit support on the PS5 GPU?

Not sure, but currently it would indicate that PS5 cannot run INT8/INT4 code faster than it does FP16 code while that portion of code can run faster on XSX/XSS in terms of pure throughput (same as FP16 on PS4 Pro).
 

rnlval

Member
The thing is that MS does not have dedicated ML cores in XSX like the tensor cores. They are taking their current vector ALU’s and splitting the 32 bit lanes into four 8 bits ones or eight 4 bits ones for parallel INT4/INT8... extending rapid packed math from FP16 which XOX fans derided (but now enjoy as XSX has this feature too) to INT4 and INT8. It does not have specialised HW like you can find on Google’s TPU, the Tensor cores on nVIDIA’s GPU’s, or Apple’s Neural Engine.

You want to run faster speed INT4/8 code you can and that portion of code will run at an increased throughput (4x as many INT4 operations as FP32 ones per CU), but you are using the CU’s you also need for shading and async compute.

Beyond FP16 RPM support we do not know if Sony has any similar enhancements as of yet.
There are pros and cons with tensors being integrated with stream processors vs another functional block away from the stream processors e.g. latency.
 

Chumpion

Member
If Sony really doesn't have INT8 then there's clearly potential for differentiation there if MS seizes the moment.
 
Yes its this generationals/GPU/BC approach that's why im bullish on XGS. You can say MS is forced to change how the console model work as fighting Sony on old grounds will be tough.

MS being MS, a hearty warchest helps but its good men like Phil, Andrew and Ronald being the final push.

Imo first gen XSXS games will be rough, late XDK, pushed by the raw ST cpu speeds and 12tf.

It's the 2nd gen XSXS games that will be beautiful. Making use of efficiency features of VRS, mesh shading, SFS, 16T, ML, sustained SoC performance. 1440p upscale to 4k plZ . :messenger_savoring:

A case study of using a hodgepodge of technical words without understanding any of it... :messenger_sunglasses: Sustained SOC performance...LOL
 

geordiemp

Member
Not sure, but currently it would indicate that PS5 cannot run INT8/INT4 code faster than it does FP16 code while that portion of code can run faster on XSX/XSS in terms of pure throughput (same as FP16 on PS4 Pro).

Correct, and the question is does ps5 need that functionality, its a dedicated console, does not need server functions.

Also in one of the ps5 Cerny patents, they start joining up pixel vertices before pixel shaders,, so it gets more confusing on ps5 requirements and what it needs.
 
That's complete bullshit, they were not "let go" but you know that already. QD becoming multiplat is actually quite a loss for Sony as Detroit sold 5M copies (twice as much as, say, Bloodborne).

The games were quite good tech wise, but vacuous and horribly, horribly written. It is not a loss when you decide to cut ties, especially after the Studio got blasted for its workplace toxic environment...
 
Last edited:

Lysandros

Member
Not sure, but currently it would indicate that PS5 cannot run INT8/INT4 code faster than it does FP16 code while that portion of code can run faster on XSX/XSS in terms of pure throughput (same as FP16 on PS4 Pro).
If true it seems like a big omission from Sony. Do we have confirmation for this?
 
Last edited:
"The CPU of the two consoles uses the same processor (slightly faster on Xbox Series X), the GPU of the Xbox also seems more powerful, as it is 16% faster than the PS5 GPU, with a bandwidth that is 25% faster. The transfer speed from the SSD is twice as fast on PS5."

All the exact same specs that was obvious to everyone, yet people try their hardest to ignore or downplay what is fundamentally true. Here is a Dev telling it like it is.

He is reading from the specs sheet, just like you would do. No developer under NDA would attempt a comparison, my friend. Two weeks to see the first results...
 

01011001

Banned
Correct, and the question is does ps5 need that functionality, its a dedicated console, does not need server functions.

if AMDs "Super Resolution" feature uses it it will come in handy. or Microsoft use it for their upscaling algorithm (not sure if they have a name for it...)
 

Panajev2001a

GAF's Pleasant Genius
LDS and L2 cache is not the same memory pool.

910fcae9-f40b-4ba1-9684-475d9dd4e46a.PNG

I was wrong on the L2, the L2 size is directly proportional to the external memory bus. The wider the bus the more Memory Controllers, the more MC’s you have the more total L2. Still feels a bit like double counting as you have the additional L2 to sustain the extra bandwidth.

You are incorrect about the LDS which is shared by CU’s inside a DCU (XSX has more DCU’s and inside each DCU we have essentially two CU’s combined and sharing some logic):

NsyFX5d.jpg


J3uypkp.jpg
 

Panajev2001a

GAF's Pleasant Genius
If true it seems like big omission from Sony. Do we have confirmation for this?

No, we do not. The only thing we know is that this is actually an optional configuration option for RDNA1 (so nothing new added just for RDNA2) and that MS chose to use it for XSX/XSS. Not sure how big of a differentiator it will be and how many developers will push it (like they pushed or did not push FP16 before). We do not know the features and/or customisations used by Sony yet.
 

Lysandros

Member
No, we do not. The only thing we know is that this is actually an optional configuration option for RDNA1 (so nothing new added just for RDNA2) and that MS chose to use it for XSX/XSS. Not sure how big of a differentiator it will be and how many developers will push it (like they pushed or did not push FP16 before). We do not know the features and/or customisations used by Sony yet.
Okay, thanks for the explanation. 👍
 

Panajev2001a

GAF's Pleasant Genius
There are pros and cons with tensors being integrated with stream processors vs another functional block away from the stream processors e.g. latency.

Indeed, but it could also be an additional dedicated special function unit inside the SP, but then again unless you get a bigger win than increased throughout you may just add more general SIMD units.

I was not just saying that there was only a tradeoff of ML compute vs shading/async compute When I talked about having dedicated cores for ML, but that these cores provide a much less versatile but also much more powerful setup which does bring a very Big Bang for your buck more than enough to offset the added latency (which you can design around/minimise).
 
Last edited:

Riky

$MSFT
"XSX GPU has 25% more L2 cache, 44% more LDS(local data storage), 44% more L0 cache, 44% more L1 cache, 44% more wavefront queue. More on-chip very fast SRAM storage leads to lower latency and fewer pipeline stalls."

Is this confirmed specs?
 

rnlval

Member
I was wrong on the L2, the L2 size is directly proportional to the external memory bus. The wider the bus the more Memory Controllers, the more MC’s you have the more total L2. Still feels a bit like double counting as you have the additional L2 to sustain the extra bandwidth.

You are incorrect about the LDS which is shared by CU’s inside a DCU (XSX has more DCU’s and inside each DCU we have essentially two CU’s combined and sharing some logic):
Each DCU has a Local Data Share which scales with DCU count. XSX's 26 DCU LDS / PS5's 18 DCU LDS = ~44% advantage for XSX. LOL

03737c08-7540-4a78-940e-a660ca7fdebf.PNG


More DCU has the following
1. more wave32 processing on the chip instead of outside the chip.
2. more Texture Filter Units
3. more texture load/store units
4. more L1 cache
5. more branch & message unit
6. more scalar units
7. more RT function blocks (RDNA 2)

Trips to the external memory bus have a higher cost.
 
Last edited:

01011001

Banned
"Wow a playstation centric developer being honest. It's actually quite refreshing to see"

Saying that sony dev houses are lying is not trolling?

so you already forgot the news that Sony/Bluepoint lied about raytracing in Demon's Souls? (hard to say where the lie came from, if Sony or Bluepoint were responsible for advertising it)

yes, developers closely tied to a first party should never be trusted without a big grain of salt.
 
Last edited:

longdi

Banned
There are efficiency features on both systems and sure we can expect GDK improvements will result in XSX games improvements, but it still remains a more generic dev approach meant for multiple platform reach over single platform efficiency (you are winking and nudging at the PS5 SDK and efficiency being set in stone and tapped out at launch... PSVita, PS4, and PS4 Pro have a history of balancing evolution and easy to code for at launch with revolution and elements to give longevity to the platform which is the fine balancing job where proper architecture does shine... I suspect PS5 will too).

All the other PR buzzwords (sustained SoC :LOL:) are there for console war stirring and getting the other lot mad ;).

🤷‍♀️

pr buzzwords? Those are efficiency features thats exclusive to XSXS. Can we at least acknowledge their designers?

It is also fair to say XSXS games will show a bigger uplift over time. The situation with the late XDK is what it is.
Now that we got another reminder/affirmation of hardware ML, it can only be better in the long run. Amd is working on super resolution shortly.

Just wait for the 2nd gen of XSXS games, im sure you will agree with my assessement.
 

rnlval

Member
"XSX GPU has 25% more L2 cache, 44% more LDS(local data storage), 44% more L0 cache, 44% more L1 cache, 44% more wavefront queue. More on-chip very fast SRAM storage leads to lower latency and fewer pipeline stalls."

Is this confirmed specs?
My statement assumes baseline RDNA v1 DCU and scaled to DCU count. I haven't factored in XSX APU's 76 MB SRAM storage.
 

rnlval

Member
Sony developer who was just bitching about the S says Xbox has an advantage. Interesting.
It's a shopping list for the future PlayStation. :messenger_grinning_squinting:

It wouldn't matter much since a good game story has its own advantages.
 
Last edited:

Panajev2001a

GAF's Pleasant Genius
Each DCU has a Local Data Share which scales with DCU count. XSX's 26 DCU LDS / PS5's 18 DCU LDS = ~44% advantage for XSX. LOL

03737c08-7540-4a78-940e-a660ca7fdebf.PNG


More DCU has the following
1. more wave32 processing on the chip instead of outside the chip.
2. more Texture Filter Units
3. more texture load/store units
4. more L1 cache
5. more branch & message unit
6. more scalar units
7. more RT function blocks (RDNA 2)

Trips to the external memory bus have a higher cost.

You are back to double counting things and making it now in a laundry list format (yet more DCU’s feeding off the same L1 does not seem to be a minus for some reasons ;)). You cannot quote more CPU cores and more CPU registers as if they were two separate things :LOL:.

... well unless you are just trying to have a e-peen numbers war and cannot accept that having “just” an 18% advantage is not bad even though it is not the monster advantage you may have wanted.
 
Last edited:

Panajev2001a

GAF's Pleasant Genius
🤷‍♀️

pr buzzwords? Those are efficiency features thats exclusive to XSXS. Can we at least acknowledge their designers?

It is also fair to say XSXS games will show a bigger uplift over time. The situation with the late XDK is what it is.
Now that we got another reminder/affirmation of hardware ML, it can only be better in the long run. Amd is working on super resolution shortly.

Just wait for the 2nd gen of XSXS games, im sure you will agree with my assessement.

I already agree 2nd generation games will look better than first generation ones and yes, you were throwing in buzzwords to make it all sound more impressive (sustained SoC :LOL:).
 

ThisIsMyDog

Member
To be honest, I hope the faster SSD of the PlayStation 5 will make a bigger difference than the stronger GPU of the Xbox Series X.

I care more about shorter loadings than native 4K VS upscaled 4K.

If you only care about the loading times, the SSD won't make much difference, a game that loads in 4s on PS5, will load in 8s on XSX, which is still very fast, you wouldn't really pay much attention to it.
 
Top Bottom