Baseless Next Generation Rumors with no Technical Merits [post E3 2019, pre GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
11.6 TF can be achieved using 44CUs@2.06GHz as I calculated in the previous posts.
Yes, that can work. I can see Sony going with 44CU chip (48CUs on die), but I doubt it would be clocked at 2.0GHz then. In any case, if they can turn of CUs without turning SE, then 48CU design would work very well actually (you could have full 3x4WGPs per SE).

Would probably put SOC around ~330-340mm2.

I still think chances are they will go with 36CUs, but this one would be my second guess.

You can also get 11.6TF with 52CU@1.75Ghz which is a much more reasonable clock.
Would be a bit of a weird fit for 256bit bus and 56CUs.

My expectations for consoles are : 56CU (60 full die, 3x5WGP SE), 44CU (48CU full die, 3x4WGP SE) and 36CU (40CU full die).
 
Last edited:
Controlling die cost may make more sense than controlling cooler cost.

Boosting frequency also affects the chip's yield.

Its one thing for AMD to bin their chips based on multiple performance profiles and then sell to AIB who further bin those chips based on OC performance. You can get Navi 10s that clock to 2.0 Ghz.

Its another to produce 2.0 Ghz viable chips across the whole wafer (excluding defects).
 
Last edited:
Boosting frequency also affects the chip's yield.

Its one thing for AMD to bin their chips based on multiple performance profiles and then sell to AIB who further bin those chips based on OC performance. You can get Navi 10s that clock to 2.0 Ghz.

Its another to produce 2.0 Ghz viable chips across the whole wafer (excluding defects).

You’re forgetting that Sony helped them design Navi so expect another 2tf just from that. Thus 9tf becomes 11tf. Easy math.
 
Boosting frequency also affects the chip's yield.

Its one thing for AMD to bin their chips based on multiple performance profiles and then sell to AIB who further bin those chips based on OC performance. You can get Navi 10s that clock to 2.0 Ghz.

Its another to produce 2.0 Ghz viable chips across the whole wafer (excluding defects).
This is why MS came up with the ‘Hovis Method’ to account for die to die variability.
 
Yes, that can work. I can see Sony going with 44CU chip (48CUs on die), but I doubt it would be clocked at 2.0GHz then. In any case, if they can turn of CUs without turning SE, then 48CU design would work very well actually (you could have full 3x4WGPs per SE).
The story seems to be that PS5 GPU stuck at high 10TFs (44CUs at 1.9 GHz?). Recently Sony tunes the cooling and pushes PS5 to 11.6TF but louder.
11.5TF seems like their limitation.

You can also get 11.6TF with 52CU@1.75Ghz which is a much more reasonable clock.
PS5 is rumored to be $450 BOM which indicates a smaller SOC. 52CUs won’t be very different. Besides 1.75GHz doesn’t seem to cause some serious heat problem.


Boosting frequency also affects the chip's yield.

Its one thing for AMD to bin their chips based on multiple performance profiles and then sell to AIB who further bin those chips based.
RDNA2 has better power efficiency. This may improve yield compared to RDNA 1 NAVI 10.
 
I would point out we don’t know the conditions under which the 50% is true. They could have tested RDNA 1.0 at much higher clocks to match 2.0 where it is wildly inefficient and cause the gap to swell.
Which I assume they did, given that they promised "multi GHz clocks". I guess there will be big Navi with ~2-2.1GHz, and where difference in perf per watt will be biggest. But as you said, at those clocks, RDNA1 is already relatively inefficient.
 
Well, 8-core (+SMT) Zen 2 can do 1.8 GHz base 4.2 GHz boost when paired with 8 CU Vega @ 1750 MHz at 15W TDP (configurable TDP 10-25W)
I would suspect that it cannot offer all those at the same time. At least historically, APUs have allowed the individual elements to max out TDP on their own, and that doesn't leave room for them both in a sustained load case.
A single Zen 2 core at 4.0 GHz+ could potentially do that, and I would expect the GPU or part of it to be similarly capable. One side of the chip is usually significantly underutilized if there's thermal room for the other to run at turbo--which is itself not promised for a long duration of time.
One of Zen 2's improvements AMD presented was the ability to operate at voltage further below where Zen 1 had a voltage floor. Perhaps the 4000 series have done something similar for the Vega CUs, although in the case of Zen2 the improvements at the upper range were more modest. AMD's presentation effectively stated they were proud they didn't regress in terms of top-end performance, and it took quite some effort to work around some of the ways 7nm makes it harder to operate at the same frequencies.

Consoles have usually offered a reliable and consistent performance target, which is where I think AMD's products might not give applicable data points.
 
I would point out we don’t know the conditions under which the 50% is true. They could have tested RDNA 1.0 at much higher clocks to match 2.0 where it is wildly inefficient and cause the gap to swell.
I know facts have no room in this thread, but the claim for 50% is against 5700 XT (and RDNA's +50% is against Vega 64)
 
I don't really see the relevance to Brit's point. He wasn't suggesting even 20 MB/s is fairly good for streaming but that even though completely anaemic, clever design could make 20 MB/s fairly effective in some cases. Therefore, 100x that next gen means storage speed is clearly going be liberated for the first time for consoles.

I'm sure BRiT can speak for himself but I started my respond with "I don't disagree", I wasn't disagreeing with him but pointing out that it is easy to get lulled by well designed worlds and streaming solutions (Spider-Man) and assume that something 100x faster will be fine for the next 6-7 years. Most currentgen open world games lack variety of objects, variety of models for those objects and textures for those objects, not to mention density of objects in any individual place unless you collect all the identical cheesewheels.

I would be hugely disappointed if all of these aspects did not hugely improve nextgen. It's not just about streaming once the game is running and keeping up with a character moving about, but about loading the game (or loading from a different save) much quicker than we do today. That was the focus of Mark Cerny's April 2019 Wired demo.
 
I said before that I don't think the xsx BC upgrades would be doing things like shader replacement, out of the box for all games.
But what about taking x enhanced games, since they've already been signed off, should be able to have a simpler time getting signed off again for something like this.
And how much performance does it take to add RT in this fashion?

 
Status
Not open for further replies.
Back
Top