Next Generation Hardware Speculation with a Technical Spin [post E3 2019, pre GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
This is supposed to be an intelligent thread with sound, intelligent predictions. Keep random guesswork and silly ideas out of it. A pattern counted from a series of two precursors isn't any pattern at all. Given a complete change in GPU architecture, there are pretty much zero parallels to be drawn from the GCN PS4 series.
 
Why only 20.72 for IO and not 37.72 ?
Based on Proelite comment
You can remove ~17mm2 from the IO as the console one won't have some stuff.
And I found 1.25 for density gains of EUV on your 66CUs APU. 435.85 / 348.68 = 1.25
A 20% area reduction is X*0.8
435*0.8 = 348
Why 3 Shader Engines?
Because i think maintaining efficiency is important, so that way it will punch above its weight like the 5700
If you drastically increase the number of DCU per SE it will be less efficient than the 5700.
I think 3 SE won't be possible with RDNA because of the butterfly structure. It's either 2 or 4.
Why not? can you elaborate
 
Based on Proelite comment


A 20% area reduction is X*0.8
435*0.8 = 348
It's 20% higher density, not 20% area reduction. It's not the same. It's actually about 17% area reduction.

Why not? can you elaborate
The butterfly structure of RDNA needs (I think) a symmetrical layout (one SE for each sides). How do you do that with 3SE ?
 
It's 20% higher density, not 20% area reduction. It's not the same. It's actually about 17% area reduction.
Thanks for pointing it out, that went right over my head
Screenshot_2019-07-04 TSMC Reveals 6 nm Process Technology 7 nm with Higher Transistor Density.png
But still i think they could bring it closer to 350mm2, there seems to be wasted space on 5700
The butterfly structure of RDNA needs (I think) a symmetrical layout (one SE for each sides). How do you do that with 3SE ?
Where did you read RDNA needs a symmetrical layout? what do you assume is the benefit of such layout?
Active CU per SE must be odd, right?
5700XT has 20CU per SE (or 10DCU)
What matters is each SE has an equal number of DCU (2xCUs)
 
Just watched vid and he says 4x processing power than 1X.
So I think it is as I said before, overall processing power. Last time I'm sure they said gpu power.

This time they are purposely, and for good reason being vague about it.

4x more powerful than XBX is a combination of SSD + CPU + GPU + ram. Immediacy of gaming (shorter loadings) is now their top priority.
https://www.eurogamer.net/articles/...cross-gen-halo-infinite-and-the-lack-of-fable

If I can't convince you maybe Phil can :p

He has clarified that the 4x power statement that its purely CPU related:
PC Games Hardware: Does the already mentioned statement 4x relate to the performance of the complete console?

Phil Spencer: No, that's a pure CPU statement. It would also be a little too simplistic to refer to the whole system, as much as I would like to, because so many components flow into it. Take the Xbox One X: In its development, the memory bandwidth was the bottleneck. It had to be big enough to provide content to the GPU without idle time. We could have brought the console to market a year earlier, but we waited another year to get all 6 of the GPU's TFLOPS up and running.

https://www.pcgameshardware.de/Xbox.../Phil-Spencer-Das-Mega-Interview-1293543/amp/

Nah overall system performance can't be 4x, since the GPU represents the overwhelming majority of computational performance any gaming system nowdays. Ryzen2 will probably do somewhere around hundred of billion OPS in peak theoretical situations, GPU in Anaconda will likely be in the range of ten of trillion operations per second. You are talking about an order of magnitude difference of 1/100.

Of course it could be conflated oh this could mean CPU limited scenarios on XOX where the XOX cpu cores are at 100% utilization and the XOX gpu is only at ~50% utilization, then you'd get 4x the performance on Anaconda (assuming the double GPU performance ballpark is correct(I think it is) since they are moving onto Ryzen 2 which ceratainly could fall under the 4x the cpu performance of XOX's 2.1ghz mildly customized Jaguar cores.
 
@Pixel Yes, I know. I was just repeating what Matt Booty (not really a nobody at MS) had said in the Eurogamer interview. Phil clarified it well after that.

Not my fault if they don't got their stories straight. :rolleyes:
 
PS5 APU could look like this ?

kfsSxjM.jpg
 
You'd think it'd be easy to communicate that in the first place. :rolleyes:

:rolleyes::rolleyes::rolleyes::rolleyes:
It is easy :cry: unfortunately we both know this is the marketing game.
You play loose and fast with numbers, and it makes your product sound better. Sure techies like us see through it, but 99% of the viewers aren't going to know this and it makes the product sound better.

How is 4x overall performance even quantified? Given (obviously since we follow the semiconductor industry we know they won't be able to put in a 4x more powerful GPU, thus) most XO games won't be running 4x faster or 4x higher resolution or a mixture of the two on Scarlett than XOX unless a game is cpu bottlenecked on XOX and the XOX GPU is at less than 50-60% utilization and you don't get memory bandwidth/ROP bound.
Since companies are getting loose and fast with numbers, and PR spin is so common, especially in something like an E3 trailer, its never good to rely on vague percentages or vague multiplier claims from execs. Only when pressed do they give up the goods.

"The new F150 offers 15% more torque than the leading competitor"
"Freesync doesn't work" - Jensen Huang
"Fallout 3 will have over 200 endings" - Todd Howard (well yeah if you reeeaaally stretch the definition of what entails an ending really these are just tiny variations on 3 endings)
 
Last edited:
How is 4x even quantified?

Sebbbi said:
8x Zen2 cores is roughly 4x faster than 8x Jaguar. Roughly 2x IPC and roughly 2x clocks (conservative 3.2 GHz estimate). Also 8x faster for AVX workloads (Jag was 0.5 rate AVX, Zen2 is 2.0 rate) such as ISPC and Unity Burst

Sebbbi said:
That might be true. But then again Zen 1 was already over 2x IPC of Jaguar, and Zen 2 got noticeable IPC uplift. So should be 4x+ in total with any reasonable clock rate in general purpose code. And of course higher in AVX/AVX2 code. So get that ISPC or Burst compiler ready.


 
Yes exactly. But 4x cpu perf isn't 4x overall. I referenced those Sebbbi's tweets earlier. I'm saying 4x represents cpu perf not overall perf. Others were saying 4x was "overall" system performance, but how is 4x "overall" even quantified? Particularly when Scarlett won't be capable of running most XO games 4x the resolution/4x the fps*/or a mixture of the two (Since we know the Scarlett GPU won't be 4x (obviously) and will bottleneck games from running at 4x in whatever metric (res/fps), then how can a 4x overall perf be claimed.
Yes 4x cpu perf, but not 4x 'overall'.

*if the game/ game engine permit higher fps
 
It's never been possible to quantify how much faster a system is. Do you accumulate the multiplier improvements of each part, or take the highest single improvement, or multiply them all together, or take the lowest single improvement? If you have 2x the RAM speed and 3x the GPU Flops and 1.5x speed the CPU, is the system 1.5x faster or 3x faster or 9x faster or 6.5x faster? As there's no such thing as 'console power' to quantify, we can't thus quantify the 'total improvement'. You'd need a qualified metric to say which aspect is faster by how much.
 
Zen cores would be on the left side, not the display engine side. Also needs to have some paths from the GDDR6 controllers to the Zen cores.
What do you make of a 3 SE config? RDNA2 possibly could be rearranged to accommodate it or just beef up the current 2SEs config with extra cache, rops etc. needed to add 33DCUs (30 enabled)
Thoughts?

Also any idea how HBM memory controllers compare to GDDR6 (smaller or bigger?)
 
Status
Not open for further replies.
Back
Top