Xbox Series X [XBSX] [Release November 10 2020]

[BRiT]

Only if it is fast and does fit the resource footprints.

What they have done so far, is and does based on what I've experienced on my OneX with puny jaguar cores.

 

[Kugai Calo]

What they have done so far, is and does based on what I've experienced on my OneX with puny jaguar cores.

And that’s the store, which is an app only active on the foreground. And the animations are very limited, nowhere near as creative as what DirectComposition or Windows.UI.Composition can offer.

Oh yeah and see @zed ’s post.

 

[BRiT]

But that creativity was likely causing slowdown on Jaguar based systems, which is likely why they removed a lot of animation effects over the years with each new dashboard update for The Guide and Home screens.

 

[Kugai Calo]

But that creativity was likely causing slowdown on Jaguar based systems, which is likely why they removed a lot of animation effects over the years with each new dashboard update for The Guide and Home screens.

If Xbox 360 with its 3 Pentium 3s and 512MB of total RAM (and PS4 with the same/slightly worse hardware) can deliver smoother animation and more responsive experience, I really don’t know what good excuse there is to regress, other than they want to be lazy.

 

[Jay]

I don't know if anyone has seen this article about Ray Tracing on the XSX. It seems they don't expect a big uptake on developers using this feature so they didn't add much hardware to support it.

It sounds more like their interpretation of what is actually said

"We do support DirectX Raytracing acceleration, for the ultimate in realism™, but in this generation developers still want to use traditional rendering techniques, developed over decades, without a performance penalty," says Grossman sadly. "They can apply ray tracing selectively, where materials and environments demand, so we wanted a good balance of die resources dedicated to the two techniques."

It sounds like what is expected, don't expect full RT games, they will be used where appropriate, needed a ballance between standard and RT hardware.
Is there anybody that was expecting games to be fully RT only and not be some form of hybrid approach?

Their interpretation of what is said as him being sad, or that ms felt they skimped on RT, that the hardware is worse for it, doesn't match my interpretation of what is said.

 

[Ike Turner]

Is there anybody that was expecting games to be fully RT only and not be some form of hybrid approach?

Just take a look at half of the GPU/Graphics related posts here around the time Turing was announced/released..."RTX: It just works" had a lot of traction....
The same PR/Kool Aid cycle is about restart in a few days btw..

 

[Nisaaru]

MS s

If Xbox 360 with its 3 Pentium 3s and 512MB of total RAM (and PS4 with the same/slightly worse hardware) can deliver smoother animation and more responsive experience, I really don’t know what good excuse there is to regress, other than they want to be lazy.

Pentium3?

They had some in-order SMT 64Bit PowerPC cores.

 

[Kugai Calo]

MS s


Pentium3?

They had some in-order SMT 64Bit PowerPC cores.

Pentium 3 (P6) was out-of-order, superscalar, IIRC.

 

[Nisaaru]

Pentium 3 (P6) was out-of-order, superscalar, IIRC.

Maybe you should re-read my post until you get the meaning

A little hint. PowerPC is not a x86 CPU like a Pentium3.

 

[Deleted member 11852]

A little hint. PowerPC is not a x86 CPU like a Pentium3.

No, PowerPC was good.

 

[Allandor]

No, PowerPC was good.

Well... when the xbox 360 came out there was a comparison with an athlon64 at around 1.4 Ghz worth of processing power for the whole xbox 360 cpu. Well we will never know if that is correct, but we know that the CPU was not that much fun to work with.

btw, while we are at "power". Do we know how much power the xbox series x needs?
Because of the 1x hint of the hot-chips presentation some seem to think that it needs about 140W. I really can't believe that, because of the whole tower-cooling design thing (and the gigantic ps4). Just 140W for the APU should be much easier to cool and you don't need such a big tower.

 

[anexanhume]

Digital Foundry video on the Hot Chips talk

 

[fehu]

I remember Spencer saying that the soc has over 70MB of cache, so 8MB for the cpu, 5MB for the gpu, and the rest?

 

[AlNom]

I remember Spencer saying that the soc has over 70MB of cache, so 8MB for the cpu, 5MB for the gpu, and the rest?

I don’t have my post handy, but there are also the L0 to L2 caches across the cores and CUs. I think it was up to 40MB or around there that were “obvious”, but I couldn’t figure out the rest, even with registers and the sort.

edit: nvm

Here it was:

Zen 2
-------
L1 = (32kB I$+ 32kB D$) * 8 = 512kB
L2 = 512kB * 8 = 4MB
L3 = 2*4MB = 8MB
Total CPU = 12.5MB
Registers/AVX etc. ???

Anaconda GPU
4 SIMD per WGP
-----------------------
I$ = 32kB per WGP
K$ = 16kB per WGP
Vector register file = 128kB*4 SIMD
L0 Vector cache = 2 * 16kB per WGP
Scalar write-back cache = 16kB*4 SIMD
Scalar RF = 10kB *4 SIMD
LDS = 2*64kB per WGP

L1 = 128kB per shader array * 4 arrays
^very confused at trying to read the whitepaper

28 WGP = ~23MB ???
GPU L2 = 5MB

Render Back End caches = 128kB per RBE ??? (guesswork - I think sebbbi did some tests back with GCN, but hard to say if there are any changes while I can't remember the amount he deduced)


Total: ~41MB accounted for

 

[scently]

I don’t have my post handy, but there are also the L0 to L2 caches across the cores and CUs. I think it was up to 40MB or around there that were “obvious”, but I couldn’t figure out the rest, even with registers and the sort.

edit: nvm

Here it was:

Zen 2
-------
L1 = (32kB I$+ 32kB D$) * 8 = 512kB
L2 = 512kB * 8 = 4MB
L3 = 2*4MB = 8MB
Total CPU = 12.5MB
Registers/AVX etc. ???

Anaconda GPU
4 SIMD per WGP
-----------------------
I$ = 32kB per WGP
K$ = 16kB per WGP
Vector register file = 128kB*4 SIMD
L0 Vector cache = 2 * 16kB per WGP
Scalar write-back cache = 16kB*4 SIMD
Scalar RF = 10kB *4 SIMD
LDS = 2*64kB per WGP

L1 = 128kB per shader array * 4 arrays
^very confused at trying to read the whitepaper

28 WGP = ~23MB ???
GPU L2 = 5MB

Render Back End caches = 128kB per RBE ??? (guesswork - I think sebbbi did some tests back with GCN, but hard to say if there are any changes while I can't remember the amount he deduced)


Total: ~41MB accounted for

I'd imagine the audio engines, the decompression blocks, and such all have caches of varying sizes. Those will take up some more SRAM.

One piece of information that I find interesting is that the devkits have 40gb of RAM. I suspect that the initial plan was to have 20gb of RAM in the retail hardware but cost is an issue in order to hit a certain price point. Certainly, the system as it is is most likely has a higher BOM than the X1, X1X, or X360 relative to their BOM at launch. Plus I'd also think they want to have a certain amount flexibility wrt the retail price given the "cat and mouse/ who blinks first" game they are playing with Sony. 20gb would have reduced that amount of flexibility even further.

All in all, it seems a very well thought out system.

 

[Kugai Calo]

Maybe you should re-read my post until you get the meaning

A little hint. PowerPC is not a x86 CPU like a Pentium3.

A little hint, read @Allandor 's post.

Well... when the xbox 360 came out there was a comparison with an athlon64 at around 1.4 Ghz worth of processing power for the whole xbox 360 cpu. Well we will never now if that is correct, but we know that the CPU was not that much fun to work with.

If still not obvious, I was referring to it as "Pentium 3s" because each core in the Xenon probably has the equivalent performance level as a Pentium III.

 

[Kugai Calo]

I don’t have my post handy, but there are also the L0 to L2 caches across the cores and CUs. I think it was up to 40MB or around there that were “obvious”, but I couldn’t figure out the rest, even with registers and the sort.

edit: nvm

Here it was:

Zen 2
-------
L1 = (32kB I$+ 32kB D$) * 8 = 512kB
L2 = 512kB * 8 = 4MB
L3 = 2*4MB = 8MB
Total CPU = 12.5MB
Registers/AVX etc. ???

Anaconda GPU
4 SIMD per WGP
-----------------------
I$ = 32kB per WGP
K$ = 16kB per WGP
Vector register file = 128kB*4 SIMD
L0 Vector cache = 2 * 16kB per WGP
Scalar write-back cache = 16kB*4 SIMD
Scalar RF = 10kB *4 SIMD
LDS = 2*64kB per WGP

L1 = 128kB per shader array * 4 arrays
^very confused at trying to read the whitepaper

28 WGP = ~23MB ???
GPU L2 = 5MB

Render Back End caches = 128kB per RBE ??? (guesswork - I think sebbbi did some tests back with GCN, but hard to say if there are any changes while I can't remember the amount he deduced)


Total: ~41MB accounted for

There's probably some TLBs littered here and there, and caches that aren't normally mentioned like constant cache, texture cache... So I wouldn't doubt the over 70MB of SRAM number either~!

Also render backends have cache? I thought they'd go straight to the VRAM.

edit:
I see you included the constant caches.

 

[Allandor]

I don’t have my post handy, but there are also the L0 to L2 caches across the cores and CUs. I think it was up to 40MB or around there that were “obvious”, but I couldn’t figure out the rest, even with registers and the sort.

edit: nvm

Here it was:

Zen 2
-------
L1 = (32kB I$+ 32kB D$) * 8 = 512kB
L2 = 512kB * 8 = 4MB
L3 = 2*4MB = 8MB
Total CPU = 12.5MB
Registers/AVX etc. ???

Anaconda GPU
4 SIMD per WGP
-----------------------
I$ = 32kB per WGP
K$ = 16kB per WGP
Vector register file = 128kB*4 SIMD
L0 Vector cache = 2 * 16kB per WGP
Scalar write-back cache = 16kB*4 SIMD
Scalar RF = 10kB *4 SIMD
LDS = 2*64kB per WGP

L1 = 128kB per shader array * 4 arrays
^very confused at trying to read the whitepaper

28 WGP = ~23MB ???
GPU L2 = 5MB

Render Back End caches = 128kB per RBE ??? (guesswork - I think sebbbi did some tests back with GCN, but hard to say if there are any changes while I can't remember the amount he deduced)


Total: ~41MB accounted for

It was 76MB SRAM on the soc. ~24MB are now "missing" because of the smaller L2 caches. A part of this will sure be in the Audio and compression blocks. But 24MB are a lot. But maybe RDNA2 Caches (with all the registers and the enhanced functionality because of RT) is bigger than we thought.

 

[Jay]

One piece of information that I find interesting is that the devkits have 40gb of RAM. I suspect that the initial plan was to have 20gb of RAM in the retail hardware but cost is an issue in order to hit a certain price point. Certainly, the system as it is is most likely has a higher BOM than the X1, X1X, or X360 relative to their BOM at launch. Plus I'd also think they want to have a certain amount flexibility wrt the retail price given the "cat and mouse/ who blinks first" game they are playing with Sony. 20gb would have reduced that amount of flexibility even further.

I very much doubt 20GB was ever on the cards, they've shown that they knew from the start that ram pricing was going to be an issue, hence all the investment in ssd and VA.
Dev kits always have more memory and is more powerful.
I'm surprised by it having 40GB also though.

 

[scently]

I very much doubt 20GB was ever on the cards, they've shown that they knew from the start that ram pricing was going to be an issue, hence all the investment in ssd and VA.
Dev kits always have more memory and is more powerful.
I'm surprised by it having 40GB also though.

It is also possible that the devkit is based on the Azure models. It will certainly need a lot of RAM to be able to run 4 X1 instances at once. I don't know, just spitballing.