Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]

Status
Not open for further replies.
They at least shouldn't want to go below PS4 Pro clocks because front-end of the pipe.
 
Last edited:
They at least shouldn't want to go below PS4 Pro clocks because front-end of the pipe.

Indeed, I think they'd want it higher. That's one of the reasons I put a guestimate upper bound at 48, lower at 40 (because I don't have it in me to hope for more than that for frequency jumps from AMD).

Well, like I mentioned above, I don't know why Vega 10 is seemingly gigantic despite the node shift and with similar compute as Fiji. Something is bloating somewhere.

Yeah, it's frikken huge. Those figures of 530~560 mm^2 have to be off (much more than 2x Polaris 10, and HMB interface should be pretty small) ... or there's some kind of wizards playground on there.
 
While you're here .... could you share your thoughts on CPUs? In particular, regarding Jaguar's future?

Phil "Phil Spencer" Spencer said last year that CPU was one of the reason they were waiting for 2017 instead of trying to go in 2016. As this was already long after the Pro leaks, one might assume that this meant he was looking to something faster than a 2.1 gHz cat core. With Zen looking unlikely based on Lisa Su's comments, and Earth Mover cores being relatively big, low IPC relative to Zen and never being seen on anything below 28 nm, what do you think AMD/MS could do to improve Jaguar with least effort?

Looking at the earth mover cores, it seems that shrinking and compacting the cores, allowed them to increase L1 size and reduce L2 latency (even more after reducing L2 size). This lead to increased IPC. With Jaguar now on 16 nm, could AMD easily play around with L1 sizes, L2 latency and L2 throughput (no ULP models to be concerned with now)? Could this lead to a worthwhile increase in IPC, similar to how it did for the earth movers?

The least-effort improvement for the cat cores is the shrink to 14nm and upclock, and that'd probably eclipse most core tweaks. Jaguar's architectural balance is pretty happy where it is, with few notable deficiencies but also few avenues for significant improvement without a larger re-working of the design.
AMD has allegedly lost most of the engineering resources that built Jaguar, so whether there's much left to do that kind of work just for Scorpio is unclear.

How much more of a jump Scorpio is supposed to be is vague, but anything Jaguar-based and 8-core doesn't seem like it will present that big of a leap--not compared to the GPU. It's a relatively narrow core that is sized at the point where getting bigger starts to get expensive in terms of decode and execution, and its caches are good enough for the execution width it has and are at a design point below where increased capability becomes increasingly expensive.

On the topic of the construction cores, an Excavator module would actually be smaller than a Jaguar one at 28nm. Steamroller was the rumored competition to Jaguar for the PS4, and if Carrizo were available then I'd suspect it would have won. Even so, the actual IPC increase from Steamroller to Excavator was something like 10-15% despite the physical and core rework (bigger L1, modified L2, better L1 Icache, bug fixes).

I'm not sure at this point what Spencer's statements add up to. His early statements didn't really single out the CPU.
A later claim about "balance" if taken literally means the GPU being 4x stronger should be paired with 4x the CPU capability.
I suppose that could be an oblique reference to Zen's "balance", but the disclosed Zen might be 3-3.5x larger if Jaguar happens to shrink by 2x (not certain). I would wonder if a 14nm Excavator would be in a similar situation to Zen as Jaguar was to Steamroller, if such a thing were to happen.
 
@DSoup so boost mode is PS4 Pro, running PS4 setup, but at Pro speeds. So it locks it's 18/36 CU, but raises clocks to what PS4 Pro would run at.

I'll need to watch more, going through this video took quickly, but looks like architecture changes being withheld here.

Our initial reservations about boost mode turned out to be somewhat correct then, and more persuasion that GNM is very specific.
 
Our initial reservations about boost mode turned out to be somewhat correct then, and more persuasion that GNM is very specific.
The question raised is what is about about accessing PS4's hardware that is hardware locked such that GPU code doesn't just scale to a bigger GPU? Are there effectively two discrete GPUs in there? It was certainly described that way. Which quite frankly is a waste and a fairly poor design, at least from the outset.
 
Are there effectively two discrete GPUs in there? It was certainly described that way. Which quite frankly is a waste and a fairly poor design, at least from the outset.
They're obviously not two discrete GPUs in the sense of the CPU having to communicate with two different devices and all games having to use SFR or AFR while each one taking its share of memory from the unified pool. That would be a total clusterfuck because actual VRAM available would decrease and you would've heard developers complaining about it.

As for being a GPU that can have half of the rendering pipeline completely disabled while maintaining total functionality, I don't see why it couldn't happen. For example, the first generation of the Geforce GTX 965M was a GM204 with half the SMs, TMUs and ROPs disabled.



As for the Boost mode, it looks like it may be activating the "second half" after all. Some games like Project Cars are showing FPS increases so large (up to ~45%) that maybe the game is prepared to scale with better hardware even in the console version. Maybe because it's using a good chunk of code from the Windows version, for example.
 
As for the Boost mode, it looks like it may be activating the "second half" after all. Some games like Project Cars are showing FPS increases so large (up to ~45%)
In the video they say that it seems to be the gpu running at normal pro speed (14% faster than ps4 non boost mode), also the cpu is clocked back up to normal speed.
What your seeing in project cars is both gpu, cpu & possibly bandwidth limited situations, so you get the biggest overall gain.
So it seems that it's still using only half the gpu.
 
They're obviously not two discrete GPUs in the sense of the CPU having to communicate with two different devices and all games having to use SFR or AFR.
No, obviously not. But the hardware could be discrete. Cerny's comments were interpreted by DF as double CUs, but his actual remark was doubling the whole GPU.

"First, we doubled the GPU size by essentially placing it next to a mirrored version of itself, sort of like the wings of a butterfly. That gives us an extremely clean way to support the existing 700 titles," Cerny explains, detailing how the Pro switches into its 'base' compatibility mode. "We just turn off half the GPU and run it at something quite close to the original GPU."

So we don't know how much was doubled. The most obvious explanation is 36 CUs, half disabled, but it'd be nice to have some clarification on how and why so we know why more CUs aren't used for Boost mode.
 
No, obviously not. But the hardware could be discrete. Cerny's comments were interpreted by DF as double CUs, but his actual remark was doubling the whole GPU.

"First, we doubled the GPU size by essentially placing it next to a mirrored version of itself, sort of like the wings of a butterfly. That gives us an extremely clean way to support the existing 700 titles," Cerny explains, detailing how the Pro switches into its 'base' compatibility mode. "We just turn off half the GPU and run it at something quite close to the original GPU."

So we don't know how much was doubled. The most obvious explanation is 36 CUs, half disabled, but it'd be nice to have some clarification on how and why so we know why more CUs aren't used for Boost mode.

That's right. TMUs were probably doubled since they're attached to CUs in all GCN architectures so far. Geometry engines were probably doubled as well because they're attached to shader engines and Pitcairn had two while Polaris 10 has 4.
The biggest question is if they doubled the ROPs as well.
 
That's right. TMUs were probably doubled since they're attached to CUs in all GCN architectures so far. Geometry engines were probably doubled as well because they're attached to shader engines and Pitcairn had two while Polaris 10 has 4.
The biggest question is if they doubled the ROPs as well.

i do wish Cerny would have gone through another public technology briefing like he did with OG PS4 but alas I don't think it's happening.

It does raise some additional questions; with this new knowledge of boost mode, are we confident/not confident that Scorpio isn't of similar design (for non patched games)
 
Personally I'm confident Scorpio is designed around a different philosophy and is FC done right. The VM and close relationship with Windows suggests that's the smart direction of MS. Worst case, Scorpio won't transparently play improved XB1 games but will be a DX12 box playing Windows games and paving the way for the future.
 
So if PS4 Pro was a straight doubling of compute units and along with an increase in clocks to get 2.3X Flops then maybe AMD has similar strategy for Scorpio to get to aprrx 4.5X flops.

Xbox One had 12 active CUs correct?

So maybe it is as simple as 12X4 to give 48 CUs.

Xbox One gpu was 856MHz I believe?

So clock 48 CUs @980MHz would give about 6TFlops...

Seems like a pretty straightforward way to double or quadruple the power..especially if your primary goal resolution boost.
 
So if PS4 Pro was a straight doubling of compute units and along with an increase in clocks to get 2.3X Flops then maybe AMD has similar strategy for Scorpio to get to aprrx 4.5X flops.

Xbox One had 12 active CUs correct?

So maybe it is as simple as 12X4 to give 48 CUs.

Xbox One gpu was 856MHz I believe?

So clock 48 CUs @980MHz would give about 6TFlops...

Seems like a pretty straightforward way to double or quadruple the power..especially if your primary goal resolution boost.

Quadrupling the Xbone GPU while getting rid of the EDRAM seems logical to me.
Wherever the Xbone can do 1080p, Scorpio can do native 4K. Wherever the Xbone can do 900p, Scorpio can do 1800p + upscale.

It would be a straightforward 4K Xbone, at least for the beginning of Scorpio's lifetime.
 
This may be unrealistic or unfair.
But I expect the Scorpio to play unpatched x1 games at locked max res of the game, with no frame drops.
If there's frame drops then it's either weak cpu (to achieve this) or just a very very badly optimised game that was unplayable on the x1.
Nice benefit would also be if it was able to stream the higher res assets in quicker as can be seen in some 4pro boosted games.

I also don't expect that you would need to select boost mode either, unless it's to show just how much difference it makes. Bit like the Halo anniversary og mode switch.
 
This may be unrealistic or unfair.
But I expect the Scorpio to play unpatched x1 games at locked max res of the game, with no frame drops.
If there's frame drops then it's either weak cpu (to achieve this) or just a very very badly optimised game that was unplayable on the x1.
Nice benefit would also be if it was able to stream the higher res assets in quicker as can be seen in some 4pro boosted games.

I also don't expect that you would need to select boost mode either, unless it's to show just how much difference it makes. Bit like the Halo anniversary og mode switch.

Yeah...but I'm fully expecting a game like Halo 5 will have it's own Scorpio version though...probably at native 4K... that will be a "launch title" for Scorpio release.
 
This may be unrealistic or unfair.
But I expect the Scorpio to play unpatched x1 games at locked max res of the game, with no frame drops.
If there's frame drops then it's either weak cpu (to achieve this) or just a very very badly optimised game that was unplayable on the x1.
Nice benefit would also be if it was able to stream the higher res assets in quicker as can be seen in some 4pro boosted games.

I also don't expect that you would need to select boost mode either, unless it's to show just how much difference it makes. Bit like the Halo anniversary og mode switch.

I never said anything about unpatched games. I only suggested that Scorpio will be a much better "4K" upgrade to Xbone than the Pro is to the PS4.
First because it's significantly more powerful than the Pro, second because the original Xbone is significantly less powerful than the original PS4.
 
Just thinking that PS4 launched in November but Sony announced in February I believe....including specs. Sort of wish Microsoft would do something similar since this is so out in the open now anyway..
 
I never said anything about unpatched games. I only suggested that Scorpio will be a much better "4K" upgrade to Xbone than the Pro is to the PS4.
First because it's significantly more powerful than the Pro, second because the original Xbone is significantly less powerful than the original PS4.
sorry, my post had nothing to do with yours, it was just a follow on from the discussion about boost mode, and what Ms will do with Scorpio. I probably should've found someone to quote to make it clearer... :)
 
Yeah...but I'm fully expecting a game like Halo 5 will have it's own Scorpio version though...probably at native 4K... that will be a "launch title" for Scorpio release.
I would be surprised if most of the first party games don't get at least a simple patch, apart from old forza's and horizon 2, and couple others. If I was them I'd even try to get some of the third party exclusives patched up, ryse, qb.

lots of games have pc versions to get assets from, etc.
it's a simple way to somewhat show the difference and have Scorpio patched games at launch, without having to invest too much. May actually be offset a bit by sales, as many under performed, but I suspect some people would pick them up. Not saying will suddenly sell millions.
 
Status
Not open for further replies.
Back
Top