Digital Foundry Article Technical Discussion [2022]

Status
Not open for further replies.
So I've been trying to get some kind of rough idea of how the SM 6.4 version of XeSS might run on Series S. And yes, my entire hypothesis is based off of this one part of the DF video (y) :

XeSS.png

The RTX 3070 is shown upscaling using XeSS from 1080p to 4K in 3.8 ms. With default clock it has specs of, I think:
- 20.3 TF fp32
- 10.2 TOPs int32
- (afaik) 40.6 non tensor TOPs int8 (I'm multiplying int32 x 4)

Xbox Series S has specs of:
- 4 TF fp32
- 4 TOPs int32
- 16 TOPs int8

So just going by the numbers (which is always risky!) you're looking at XSS being ~ 1/5th speed at fp32 and int32, and ~ 2/5th speed at int8. So depending on the balance of instructions used, and assuming the 3.8 ms represents all the cost of XeSS, you might expect the XSS to be 20% to 40% as fast in this scenario, upscaling from 1080p to 4K. So taking 9.5 to 19ms where the 3070 takes 3.8 ms.

If DP4a is leant on heavily in XeSS, then that might be expected to skew towards the better end of the range for Series S, as it is in less of a deficit for int ops than flops.

However, you probably wouldn't be trying to scale from 1080p to 4K on the Series S. Assuming that XeSS workload scales broadly linearly with resolution, upscales from 540p to 1080p (2.4 ~ 4.8 ms?) or even 720p to 1440p might be within reach.

All very highly speculative of course, so take with a big 'ol pinch of salt!
 
function said:
So I've been trying to get some kind of rough idea of how the SM 6.4 version of XeSS might run on Series S. And yes, my entire hypothesis is based off of this one part of the DF video (y) :

View attachment 6939

The RTX 3070 is shown upscaling using XeSS from 1080p to 4K in 3.8 ms. With default clock it has specs of, I think:
- 20.3 TF fp32
- 10.2 TOPs int32
- (afaik) 40.6 non tensor TOPs int8 (I'm multiplying int32 x 4)

Xbox Series S has specs of:
- 4 TF fp32
- 4 TOPs int32
- 16 TOPs int8

So just going by the numbers (which is always risky!) you're looking at XSS being ~ 1/5th speed at fp32 and int32, and ~ 2/5th speed at int8. So depending on the balance of instructions used, and assuming the 3.8 ms represents all the cost of XeSS, you might expect the XSS to be 20% to 40% as fast in this scenario, upscaling from 1080p to 4K. So taking 9.5 to 19ms where the 3070 takes 3.8 ms.

If DP4a is leant on heavily in XeSS, then that might be expected to skew towards the better end of the range for Series S, as it is in less of a deficit for int ops than flops.

However, you probably wouldn't be trying to scale from 1080p to 4K on the Series S. Assuming that XeSS workload scales broadly linearly with resolution, upscales from 540p to 1080p (2.4 ~ 4.8 ms?) or even 720p to 1440p might be within reach.

All very highly speculative of course, so take with a big 'ol pinch of salt!
Click to expand...

On consoles its probably more worthwhile to go with existing upscaling methods.
 
function said:
However, you probably wouldn't be trying to scale from 1080p to 4K on the Series S. Assuming that XeSS workload scales broadly linearly with resolution, upscales from 540p to 1080p (2.4 ~ 4.8 ms?) or even 720p to 1440p might be within reach.
Click to expand...
It would be very tight for 30fps, but possible. The lower resolution scaling seems fairly doable, if they can get it below 10ms.

Quality would be painful though. Upscaling from so few pixels becomes a very large problem in terms of inference. But if the quality is incredible, that would be amazing.
I suspect something has to be in the works for consoles, but likely not ready to be prime time until next generation, and in an ideal world they may release it earlier in the later parts of this generation.
Though who would they partner with, I suspect only AMD is willing to help push this for them given the nature of the hardware and the likelihood that they are the supplier for next gen, so I think I'm indirectly ruling out XeSS.

With AMD being the partner both Sony and MS and other will benefit if they pool resources. I don't see MS or Sony taking on ML upscaling on their own.
 
iroboto said:
It would be very tight for 30fps, but possible. The lower resolution scaling seems fairly doable, if they can get it below 10ms.

Quality would be painful though. Upscaling from so few pixels becomes a very large problem in terms of inference. But if the quality is incredible, that would be amazing.
I suspect something has to be in the works for consoles, but likely not ready to be prime time until next generation, and in an ideal world they may release it earlier in the later parts of this generation.
Though who would they partner with, I suspect only AMD is willing to help push this for them given the nature of the hardware and the likelihood that they are the supplier for next gen, so I think I'm indirectly ruling out XeSS.

With AMD being the partner both Sony and MS and other will benefit if they pool resources. I don't see MS or Sony taking on ML upscaling on their own.
Click to expand...

RDNA3+ is (rumored?) to be including hardware AI acceleration just like Intel and NVs solutions.
 
davis.anthony said:
The issue with XSS is its performance.

I genuinely don't think it has the performance in its tiny GPU to proves and AI upscale in the small frame times that are needed.


I'm sure it has like 1/8 of the INT performance of an RTX2060 OS something like that.
Click to expand...
The question is how powerful the Xe iGpu performance is compared to the XSS. I'm talking performance in general not DP4a peformance. The very name gives it away, XeSS. Intel likely always had iGPu in mind when they created it. In its current form, XeSS on the XSS would use the SM6.4 path, but if they wanted to, either Microsoft or Intel could make an optimized version for the GDK. With that said, I wondered why Alex didn't run test on a GTX 1060 to see how well the SM6.4 path worked on a low end gpu. The S was never meant to be a 4k console so 1440 would be the target.
 
Last edited:
  • Like
Reactions: Jay
Dr. Nick said:
The question is how powerful the Xe iGpu performance is compared to the XSS. I'm talking performance in general not DP4a peformance. The very name gives it away, XeSS. Intel likely always had iGPu in mind when they created it. In its current form, XeSS on the XSS would use the SM6.4 path, but if they wanted to, either Microsoft or Intel could make an optimized version for the GDK. With that said, I wondered why Alex didn't run test on a GTX 1060 to see how well the SM6.4 path worked on a low end gpu.
Click to expand...

Intel have said it's compatible with their iGPU's, they've haven't said how performant or useful it is in reality.
 
Dr. Nick said:
With that said, I wondered why Alex didn't run test on a GTX 1060 to see how well the SM6.4 path worked on a low end gpu.
Click to expand...
I am sure he will cover that when it releases.
I would be interested in the following comparisons:
GTX 1060 vs RX 580
2060 Super vs RX 5700

Quality and Performance.

Low end vs midrange, DP4a vs FP16 fallback across different architectures.

And of course a comparison against FSR 2 on the same GPUs, if a title releases soon that supports both.

The most interesting stuff about it is still left to be uncovered!
 
Dampf said:
I am sure he will cover that when it releases.
I would be interested in the following comparisons:
GTX 1060 vs RX 580
2060 Super vs RX 5700

Quality and Performance.

Low end vs midrange, DP4a vs FP16 fallback across different architectures.

And of course a comparison against FSR 2 on the same GPUs, if a title releases soon that supports both.

The most interesting stuff about it is still left to be uncovered!
Click to expand...

Alex has something to do going forward ;)
 
iroboto said:
Quality would be painful though. Upscaling from so few pixels becomes a very large problem in terms of inference. But if the quality is incredible, that would be amazing.
Click to expand...
We've seen that 720p -> 1080p & 860p -> 1440p gives decent results sometimes lot better than native.
davis.anthony said:
Intel have said it's compatible with their iGPU's, they've haven't said how performant or useful it is in reality.
Click to expand...
It's obviously not much to go on but Tom (is that the Intel guys name) was pretty upbeat about the DP4a implementation on their chips.

XSS (DP4a)
Cons:
Cost of the upscale @60fps could be very tight.

Pros:
Potentially IQ
Releaving pressure on bandwidth etc (which could be a big win?) by being able to render at a lower resolution than would otherwise be desirable.

I'll not bother explain my views here but there's good but different reasons for both MS & Intel to want XeSS on XS. So I would expect it to be the DP4a version.
 
BRiT said:
I couldn't disagree more with their statement "the feature works - and works well." Sure, if you ignore all of these game breaking and horrible potential user issues it works fine... Come on, no company in their right mind would ever ship anything so unstable in a flagship title. That's the honest reality.
Click to expand...
John's conclusion of 'it's actually great' and 'very playable' and 'surprisingly stable' on XB1 is particularly baffling to me. He would absolutely tear apart any other game running that badly, and would undoubtedly be calling it 'unplayable' and 'a mess'.

I mean, either such framerates are playable or they aren't when we're talking about a given type of game. I understand giving allowances for hardware in the case of something like Switch, but for a TV-based big box console, I dont think this would be considered acceptable by many at all. I can totally believe this is a(or the) big reason they've ditched it.
 
Seanspeed said:
John's conclusion of 'it's actually great' and 'very playable' and 'surprisingly stable' on XB1 is particularly baffling to me. He would absolutely tear apart any other game running that badly, and would undoubtedly be calling it 'unplayable' and 'a mess'.

I mean, either such framerates are playable or they aren't when we're talking about a given type of game. I understand giving allowances for hardware in the case of something like Switch, but for a TV-based big box console, I dont think this would be considered acceptable by many at all. I can totally believe this is a(or the) big reason they've ditched it.
Click to expand...
I would have to rewatch it, but I'm sure he said as a beta/early access feature.
Was it much worse than the current campaign etc apart from the bits he said would need to be fixed?
 
BRiT said:
I couldn't disagree more with their statement "the feature works - and works well." Sure, if you ignore all of these game breaking and horrible potential user issues it works fine... Come on, no company in their right mind would ever ship anything so unstable in a flagship title. That's the honest reality.
Click to expand...
You mean platformer owner? Many games release in completely broken states these days. It'll be curious what MS do wth Bethesda who are pretty notorious for their buggy titles.
 

AyaNeo AirAyaNeo Air ProSteam Deck
ProcessorRyzen 5 5560URyzen 5 5560U/Ryzen 7 5825UCustom AMD 'Van Gogh'
CPUZen3+: Six Cores/12 ThreadsZen3+: Six Cores/12 Threads or Eight Cores/16 ThreadsZen 2: Four/Eight Threads
GPUVega 6: Six Compute Units at up to 1600MHzVega 6: Six Compute Units at up to 1600MHz/ Vega 8: Eight CUs at up to 1600MHzRDNA 2: Eight Compute Units at up to 1600MHz
Memory8GB LPDDR4X-3200 or 16GB LPDDR4X-42008GB or 16GB LPDDR4X-420016GB LPDDR5-5500
Screen5.5-inch 1080p OLED5.5-inch 1080p OLED7-Inch IPS LCD
Storage128GB/512GB SSD512GB/1TB/2TB SSD64GB Flash/256GB/512GB SSD
 
Last edited by a moderator:
My one question is does Richard Leadbetter own more than one t-shirt (DF branded) and hoodie (grey) ??? I am vaguely sure I saw him wearing an actual shirt once, but in retrospect, that may have been CGI.
 
Status
Not open for further replies.

Similar threads

Shifty Geezer
Digital Foundry Article Technical Discussion [2024]
53 54 55
Replies
1K
Views
85K
Flappy Pannus
F
BRiT
  • Locked
Digital Foundry Article Technical Discussion [2023]
252 253 254
Replies
5K
Views
467K
chris1515
C
T
2023 IQ is unacceptable [spawn]
2 3 4
Replies
65
Views
7K
Seanspeed
S
BRiT
Digital Foundry Retro Discussion [2022]
4 5 6
Replies
117
Views
21K
BRiT
BRiT
BRiT
  • Locked
Digital Foundry Article Technical Discussion [2021]
207 208 209
Replies
4K
Views
512K
BRiT
BRiT
Back
Top