Digital Foundry Microsoft Xbox Scorpio Reveal [2017: 04-06, 04-11, 04-15, 04-16]

[mpg1]

Aren't they effectively brute forcing esram emulation. Would 218GB/s be enough?

Also pretty sure in the Digital Foundry article one of the Microsoft engineers said after profiling for 4K an 4K assests they came to the conclusion they needed 300GB/s. But who knows maybe that's theoretical number?

 

[Ike Turner]

I think the article has a number of factual errors and shaky suppositions.

Indeed. The first 4 drafts of the article (it was update 4 times..) where unlike anything I've seen on AnandTech ever. It started by saying that the CPU was Ryzen based and the GPU was Vega based even though all info was based on DF piece that everybody was reading & which was clearly stating the opposite....so god knows how Ian came to these conclusions?? I guess they where in a hurry for the clicks given that it was updated 3 times in the first 2 hours of it being online..
From the comments

Sorry for the extra additionals making the flow seem a bit odd. As more and more of this story was passing my eyes, it was changing some of my assumptions as I went along and required new analysis. We'll reach out to MS and perhaps get some clarification on some of the details and publish a new piece in due course. Ideally I would love to got the same access DF had. I bet MS would love my questions...

And this etc..

 

[Cyan]

There's no real max. The max comes from the slowest timing path in the design and the process node. With process scaling the same design should be able to hit higher clocks, with re-balancing the timing paths you can further increase clock. Sometimes that's not practical if it were to be a significant change, or the schedule doesn't allow it even if it's easily doable. This is probably one of the things that AMD improved in the year between the Pro and Scorpio.

very interesting, thanks. PS4 and Xbox One CPUs seemed to be the same, and I remember that when the console was called Durango, engineers said that for them going past 1,75GHz was counterproductive because of the diminishing returns and increased latency. So going past that speed was not beneficial.

So I wonder which is the exact model of Jaguar CPU for the console, because they say it has minimum latency and it is close to the Jaguar line max speed, which was like 2.4GHz, officially. Maybe the CPU is based on the A6-6310 processor, with up to 2,4GHz max speed.

Tablets line -maybe the A10 Micro 6700T is the CPU the PS4 Pro is based on-;

Laptops line:

 

[iroboto]

Aren't they effectively brute forcing esram emulation. Would 218GB/s be enough?

Also pretty sure in the Digital Foundry article one of the Microsoft engineers said after profiling for 4K an 4K assests they came to the conclusion they needed 300GB/s. But who knows maybe that's theoretical number?

That 300 GB/s is probably near the point in which ROPS and bandwidth intersect. 326GB/s seems to just clear that. It's definitely on purpose, everything has a purpose

As long as there is enough bandwidth to do complete 4K@60 with 16xAF. I assume this is proven with Forza 6 @ ultra. I think to do it you require about 300 GB/s.

 

[AlNom]

The ROPs will need ~280GB/s for just RGBA16F writes. Given bandwidth utilization efficiency, it's probably close enough.

 

[Scott_Arm]

Xbox One had one thing going for it, which was dedicated GPU memory in ESRAM. I imagine Scorpio will see the same drop off in memory bandwidth as PS4/Pro when the CPU is accessing memory. Maybe they slightly over-spec to deal with cases of contention?

 

[sebbbi]

Shaky suppositions: suggesting crossbar may bring no extra BW; saying 16 ROPs were a "big part of the reason" that X1 commonly uses 900p ("it has been a pretty universal suspicion").

I am always smiling inside when someone implies that ROPs alone are enough to scale to higher resolution (or the limit resolution scaling). ROPs are pretty dumb devices that write data to memory. Somebody must provide this data. More pixels = more data. Also more data to write = more bandwidth required. Doubling the ROPs only brings linear gains if you also double the FLOPs, double the TMUs and double the memory bandwidth.

Most of the modern AAA game shaders are so complex that writing the result to memory (ROP) isn't the bottleneck. ROPs are the bottleneck only in passes that don't use pixel shaders (shadow maps, z-prepass) or in passes that use very simple pixel shaders. Modern AAA games can also avoid ROP bottlenecks completely in lighting and post processing, etc steps by using compute shaders. Compute shaders don't use ROPs at all.

 

[turkey]

For the bandwidth they have what options are there? I don't understand memory hardware at all so apologies for the probably faceplam question.

I assume 12 GB chips offers certain set bandwidth offerings, how much more than a 250 GBS solution does this 326 one cost?

 

[Rangers]

Well, it's about two factors...RAM speed and the memory bus. RAM speed will generally be in the sameish ballpark for a console we assume. So it's mainly down to using a 384 bit bus vs the 256 one in PS4 pro. That wider bus gives you a significant bump in BW.

It does require a certain minimum die size, and I'm sure eats up some amount of area on it's own. However the Scorpio GPU is surely big enough on it's own where choice of 384 bus is not forcing it to be bigger than it otherwise would be. At least much. However if they wish to die shrink it in the future it could cause problems. For a ostensibly "high end" iterative "thing" like Scorpio, they probably aren't too worried about bottom line pricing in future.

 

[Rangers]

Indeed. The first 4 drafts of the article (it was update 4 times..) where unlike anything I've seen on AnandTech ever. It started by saying that the CPU was Ryzen based and the GPU was Vega based even though all info was based on DF piece that everybody was reading & which was clearly stating the opposite....so god knows how Ian came to these conclusions?? I guess they where in a hurry for the clicks given that it was updated 3 times in the first 2 hours of it being online..
From the comments


And this etc..

Anand La Shimpi left Anandtech some time ago (to work at Apple apparently). Site has not been the same quality, as you might expect when the founder leaves I suppose. In particular the CPU side (Ian Cutress who wrote that Scorpio thing) is questionable IMO. GPU side under Ryan Smith, better.

 

[jiaolu]

the forza 4k on scorpio looks really great. the lighting seem greatly enhanced too?
http://www.eurogamer.net/articles/d...-motorsport-on-project-scorpio-the-full-story

 

[iroboto]

View attachment 1950

the forza 4k on scorpio looks really great. the lighting seem greatly enhanced too?
http://www.eurogamer.net/articles/d...-motorsport-on-project-scorpio-the-full-story

Yup. I think you're looking at tech from Forza horizon 3 which has some of those lighting elements in them.

 

[Silent_Buddha]

View attachment 1950

the forza 4k on scorpio looks really great. the lighting seem greatly enhanced too?
http://www.eurogamer.net/articles/d...-motorsport-on-project-scorpio-the-full-story

That's a pretty good read. Forza Motorsport is one of the games that isn't limited due to the CPU (the most CPU limiting functions already are run at 60 FPS or more, even on Forza Horizon 3).

"This is us. This is ForzaTech running 60 frames a second, 4K," Tector says proudly. "We're still running with settings that we would have used in Forza 6. Since it's Xbox, we're using EQAA, so it's like a 4:2 EQAA. That's the actual GP utilisation so we're only using 60 per cent of the compute to get to this. Importantly, I know I've just said it's like a Forza 6 set-up, but this is also including 4K content, so all of our build system - we've got authored assets for this set of the models, cars, tracks everything. We pushed it through and made sure the 4K textures were flowing through."

That's with XBO quality settings but using increased texture quality (larger size, higher fidelity) and models (?) appropriate for 4k. That's a lot of GPU utilization that can still be used to improve the visual quality of the game on Project Scorpio if they want. As an example,

"The crazy story here is that we've gone over our PC ultra settings and for everything that's GPU-related, we've been able to max it - and that's what we're running at, 88 per cent," says Tector, pointing to the utilisation data at the top of the screen. Right beneath it is the anti-aliasing setting - 4x, or rather 8:4x using the Radeon EQAA hardware AA.

"This is rendering the player LOD for every car, so you won't see a single LOD pop. [It's the] top-level model you'd see in race, one below what you see in Autovista, the model you usually only see for the player. And then we balance out LODs across the scene. It's a disgusting abuse of GPU power is what it is, right?"

And it's still not maxing out the GPU. Hence rendering time never pops above 16.6 ms so it maintains a rock steady 60 FPS.

The other really impressive thing?

The process of getting ForzaTech up and running on the makeshift Project Scorpio hardware took just two days, and according to Chris Tector, the majority of that time was adapting the codebase from an earlier XDK [the PC-based development environment] that dated to just after Forza Motorsport 6's ship-date.

Just 2 days work using the codebase from 2015. It'll be interesting to see what can be done for 1st party exclusives. But apparently 3rd party developers have been in to mess around with the hardware and have seen similar results with their engines on Project Scorpio. Implication being that just naïve scaling to full 4k (non-checkerboard) still leaves a lot of GPU headroom for them to do whatever they want with it depending on the games original resolution on XBO. IE - a 1080p game will have lots of headroom for additional graphics IQ while a 900p game will have less. And below that there probably isn't much headroom if any, but should still be able to do 4k just by scaling up the resolution.

Anyway, people should really read that article. Some good info on the use of PIX and a funny anecdote about the development of Project Scorpio.

Regards,
SB

 

[Prophecy2k]

This is all still pretty pointless with no price - we're still left with so much uncertainty...sure it's a 'no brainer' for XBO owners to upgrade but will be price dependent. Also, for the best version of cross plats games, if you really must have those best versions - again, what price for what upgrade over Pro?

It's a bit like going to a car show room and the salesman showing you a really nice car but not asking what your budget is or what you will use the car for!

I'm pretty sure Leadbetter was right when he said $499.

It's niche hw targeting the hardcore Xbox owning market. It's a premium 4k xbox for gamers would would be willing to shell out $150 for an Xbox Elite controller.

When has MS ever released a premium targeted product at an non-premium price? It'll be $449 minimum, imho.

I am curious too, how much the "power-profiled motherboard components" and vapour chamber cooling solution (a first for a console - and likely for a good reason) will add to the price?

 

[Scott_Arm]

$500 if it included an elite controller might be an interesting package. I imagine the margin on the controller is massive, so that way they can keep the price high enough they don't lose money on the hardware, but also make it seem like the bundle is a deal.

 

[Jay]

Just 2 days work using the codebase from 2015. It'll be interesting to see what can be done for 1st party exclusives. But apparently 3rd party developers have been in to mess around with the hardware and have seen similar results with their engines on Project Scorpio. Implication being that just naïve scaling to full 4k (non-checkerboard) still leaves a lot of GPU headroom for them to do whatever they want with it depending on the games original resolution on XBO. IE - a 1080p game will have lots of headroom for additional graphics IQ while a 900p game will have less. And below that there probably isn't much headroom if any, but should still be able to do 4k just by scaling up the resolution.

I think that importing 4k assets is a highlight, but if a game doesn't already have any available, then a quick boost to res, maybe turn up some filtering would be a reasonable patch for many games. If the cost of doing it is a couple days of work. Good for Scorpio, and not a bad bit of pr for older titles.

Even if you don't hit native 4k, 1800p or 1400p (at the worst) would give good image quality down sampled to 1080p, and a lot better image at 4k upscale if needed.

Be interesting to know what it takes to increase the res, if ms has added something in the xdk to simplify the process, e.g. developer selectable esram emulation size. Maybe with something like that, they could set it to say 256mb, leave most of the code the same and just increase frame buffer size etc. Otherwise you would probably need to remove esram handling and logic which could be a larger bit of work?

 

[Scott_Arm]

I think 3rd parties will probably do whatever they're doing on PS4 Pro. So if they're doing checkberboard on PS4 Pro, they're going to do checkerboard on Scorpio. I could be wrong. Maybe it'll be easier to just do native 4k if it's performant. I'm expecting some games won't get there though.

I'd love to see an update to Quantum Break. Hope they show something at E3.

 

[iroboto]

I think that importing 4k assets is a highlight, but if a game doesn't already have any available, then a quick boost to res, maybe turn up some filtering would be a reasonable patch for many games. If the cost of doing it is a couple days of work. Good for Scorpio, and not a bad bit of pr for older titles.

Even if you don't hit native 4k, 1800p or 1400p (at the worst) would give good image quality down sampled to 1080p, and a lot better image at 4k upscale if needed.

Be interesting to know what it takes to increase the res, if ms has added something in the xdk to simplify the process, e.g. developer selectable esram emulation size. Maybe with something like that, they could set it to say 256mb, leave most of the code the same and just increase frame buffer size etc. Otherwise you would probably need to remove esram handling and logic which could be a larger bit of work?

unlikely. Code was designed with 32MB in mind. The esram remapping is only for non-patched games. Once you patch, you're going to toss esram out entirely which isn't that bad, (2 days of work)

 

[RancidLunchmeat]

if nobody but MS first party titles are using dx12 features on essentially a dx12 only console that would be an issue. That would really suggest that they didn't design dx12 very well. Lol. There are feature flags and other things. Considering both PS4 and XBO are effectively the same GPU this seems unlikely. The XPA initiative helps to combat the issue of dx12 on PC; MS tied two dx12 ecosystems together.

So is Sebbbi wrong here, or did I misunderstand his point? If PC developers aren't utilizing DX12 specific functions because their base is split between XP, 7 and 10, I can't see how console manufacturers are going to spend time doing it for Scorpio games. And if the GPU in the PS4/PRO also have the extra DX12 command units to save CPU overhead by 30%, and they are in the XB1 currently, why aren't we already seeing these gains?

https://forum.beyond3d.com/posts/1976117/

 

[Jay]

unlikely. Code was designed with 32MB in mind. The esram remapping is only for non-patched games. Once you patch, you're going to toss esram out entirely which isn't that bad, (2 days of work)

where did you get 2days of work from?
if it's the forza example, we don't actually know how that was achieved.
unless your saying it is that easy to remove esram logic for Scorpio path, which if that's the case then even better.
whatever is fastest and easiest

@Scott_Arm
it's a lot easier to do native than checker boarding.
in new games both may use checker boarding though, doubt for patches for older games in general though