Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]

[iroboto]

https://mobile.twitter.com/XboxP3/status/823770575265038336

Posting this for timeline reasons. Jan 23rd early kit.
Somewhat correlated to the possibility that official SDKs probably still not ready for devs (excluding exceptions).

But from the sounds of the tweet it should be soon. He makes mentions of the looks, I assume this means 'tentative complete on box and fitting', I assume chips are ready for mass production; so the hardware can be officially declared locked from the information provided to us.

 

[mrcorbo]

Digital Foundry got a hold of a whitepaper sent to developers about future Xbox One/Scorpio/PC development.

Some tidbits:

No ESRAM in Scorpio.
Reiterates that developers will *not* be able to develop Scorpio exclusives. Any title developed for Scorpio must also run on OG Xbox One and they go on to suggest how that can be achieved while also maximizing Scorpio.
Suggestions that developers find ways to run their CPU stuff at half-rate (30hz) if they want to target 60hz which *may* indicate the lack of a significant CPU upgrade in Scorpio.

 

[Karamazov]

could they ship it sooner than expected ? Like before summer ? Like annoucing it at E3 and then say "available tomorrow !"

 

[Shifty Geezer]

Does this point to the low latency of the ESRAM not really being A Thing? Back in the day where this was tooted as an advantage of ESRAM, I argued it can't be that meaningful as no-one at MS etc. was talking about it.

 

[turkey]

Digital Foundry got a hold of a whitepaper sent to developers about future Xbox One/Scorpio/PC development.

Some tidbits:

No ESRAM in Scorpio.
Reiterates that developers will *not* be able to develop Scorpio exclusives. Any title developed for Scorpio must also run on OG Xbox One and they go on to suggest how that can be achieved while also maximizing Scorpio.
Suggestions that developers find ways to run their CPU stuff at half-rate (30hz) if they want to target 60hz which *may* indicate the lack of a significant CPU upgrade in Scorpio.

Or offer ways for the Xbox one to keep up, these techniques will help the lower power system more I assume. The paper seems as much about multi platform target and the original Xbox one as it is Scorpio.

 

[mrcorbo]

Does this point to the low latency of the ESRAM not really being A Thing? Back in the day where this was tooted as an advantage of ESRAM, I argued it can't be that meaningful as no-one at MS etc. was talking about it.

It might be a thing, but other things might be more important things that would be lesser things if ESRAM remained a thing in Scorpio.

 

[mrcorbo]

Or offer ways for the Xbox one to keep up, these techniques will help the lower power system more I assume. The paper seems as much about multi platform target and the original Xbox one as it is Scorpio.

Quoting DF, who are quoting from the Whitepaper:

"We acknowledge that developers may not wish to spend all of the additional GPU resource of Project Scorpio on resolution, and this is not mandated," the paper says. "To make the best games possible, developers will inevitably spend GPU resource on other quality improvements such as higher fidelity shadows, reflections, texture filtering and lower draw distances. Another option developers might consider is frame-rate upscaling - running graphics at 60Hz but the CPU at 30Hz and interpolating animation."

 

[Cyan]

Does this point to the low latency of the ESRAM not really being A Thing? Back in the day where this was tooted as an advantage of ESRAM, I argued it can't be that meaningful as no-one at MS etc. was talking about it.

You might recall that people argued in the forums the eSRAM's low latency could be interesting compared to GDDR5, as an advantage for XB1. In the end, nobody gave a damn it seems, nor developers ever mentioned that possible advantage. Maybe it was true, nobody seemed to notice it

 

[3dilettante]

Cross-referencing the leaked SDK with other presentations seemed to put the ESRAM at about half the latency of main memory. I thought that the general consensus was that it could help in specific scenarios, but that the GPU was generally tolerant of worse latency.
The interview with the Xbox One designers had the point made that the GPU was latency-tolerant when that point was brought up.
If there's enough hardware in the new GPU, the portions that might be somewhat affected could be brute-forced.

Less clear would be the corner cases where the access patterns are less than ideal for DRAM, however there's more memory channels and likely items like compression that could help reduce the amount of non-ideal access patterns.

 

[iroboto]

Does this point to the low latency of the ESRAM not really being A Thing? Back in the day where this was tooted as an advantage of ESRAM, I argued it can't be that meaningful as no-one at MS etc. was talking about it.

It may be more precise to say that latency was probably always a factor for XBO performance, but no one coded to take advantage of it (as latency tricks would not work on any other configuration).

I'm sure the XBO would perform worse with higher latency, with such a small space to work with you're going to have to access it more and more, higher latency would really just kill the performance here.

edit: I'll amend. Low Latency probably made esram more tolerant to less optimized coding. Latency hiding techniques should apply here as they would in GDDR5

 

[iroboto]

One thing to keep note of the release is that the information provided to DF was released to developers shortly after their E3 announcement. With the early SDK kits being completed now, we have no real knowledge if anything has changed since then, we only know that between E3 to Jan (approximately 6 months) changes could have been introduced. Unlikely, but possible, if possible likely minor.

 

[Cyan]

Cross-referencing the leaked SDK with other presentations seemed to put the ESRAM at about half the latency of main memory. I thought that the general consensus was that it could help in specific scenarios, but that the GPU was generally tolerant of worse latency.
The interview with the Xbox One designers had the point made that the GPU was latency-tolerant when that point was brought up.
If there's enough hardware in the new GPU, the portions that might be somewhat affected could be brute-forced.

Less clear would be the corner cases where the access patterns are less than ideal for DRAM, however there's more memory channels and likely items like compression that could help reduce the amount of non-ideal access patterns.

Now that you mention other types of RAM like DRAM, I also remember how DDR3 memory is known for having less latency than GDDR5, another comparison brought as an advantage for the X1 when people talked about the (then) new consoles in the very early days.

In the end, I think the most tangible difference I've seen favouring the X1 was CPU's frequency and having a dedicated sound chip thus freeing up CPU work.

 

[3dilettante]

Now that you mention other types of RAM like DRAM, I also remember how DDR3 memory is known for having less latency than GDDR5, another comparison brought as an advantage for the X1 when people talked about the (then) new consoles in the very early days.

DDR3 is roughly the same latency as GDDR5, however. When I cross-referenced the SDK for the ESRAM latency, the rough latencies for memory came out close. More came out of the delay getting out of the cache and through the uncore.

What usually happens is that GDDR5 is tied to a GPU memory subsystem that heavily emphasizes bandwidth at the expense of taking more time to coalesce and re-order accesses. DDR3 is usually tied to a device like a CPU that cannot tolerate that kind of latency, and so the memory subsystem fires off accesses sooner.
The internal arrays of GDDR5 and DDR3 are physically similar, if not the same. The difference is more in the choice of how many arrays the interface reads at once, and signalling. The interface and distance traveled doesn't add that much, so the dominant factors are the DRAM arrays (not different) or the memory controller and hierarchy (independent of the memory device).

edit:
For reference, here's where I tried to derive some of the values:
https://forum.beyond3d.com/threads/...are-investigation.53537/page-407#post-1816112

 

[turkey]

Quoting DF, who are quoting from the Whitepaper:

Yes undoubtedly this will happen but equally.

A 4.5x boost to compute power suggests that 1080p engines will scale nicely to 4K on Scorpio, but the reality is that many Xbox One titles render at a 900p base resolution. The leap to 4K therefore becomes a 5.76x increase in pixel-count and at the same time, developers may not wish to spend GPU power on pixels alone.

So the issue in some manner seems less Scorpio hitting 4k but the base game hitting 1080 at which point later 4k on scorpio becomes far easier.

The paper is about Scorpio but I suspect hopes are more advanced techniques can keep Xbox one closer to 1080 than 720 going forward to keep it viable for as long as possible.

 

[Jay]

lower latency was probably a nice positive, but I think memory contention was a bigger gain

 

[mrcorbo]

Yes undoubtedly this will happen but equally.



So the issue in some manner seems less Scorpio hitting 4k but the base game hitting 1080 at which point later 4k on scorpio becomes far easier.

The paper is about Scorpio but I suspect hopes are more advanced techniques can keep Xbox one closer to 1080 than 720 going forward to keep it viable for as long as possible.

I'm reading it as more, "You need to continue to consider how things will run on OG XBox One while developing on Scorpio and PC, but some of the techniques you need to use to get acceptable performance on Xbox One can also be leveraged to get better performance on Scorpio and PC.". It's a little like how the idea was put forward that as much as developing for PS3's architecture sucked (especially early on) since it took so much more effort to get acceptable performance it also led to programming in a way that ended up benefitting your performance on 360 and PC as well.

 

[iroboto]

Rumor mill:

I also saw Scorpio documents dated around E3 2016, things have changed.

https://twitter.com/JezCorden/status/823954818905604097

as always, take with caution until official channels. Considering the recent media swirl, and engagement from MS on Neogaf forums again, I can't imagine waiting too long.
edit: save you guys some time and reduce the noise, he is a senior editor at WindowsCentral

 

[turkey]

Clarification issued

https://twitter.com/JezCorden/status/823976558469771264

I didn't see the same docs as DF for clarity. But I saw *some* docs around E3 about Scorpio stuff that later changed. Plans change, etc. https://t.co/4sono7QTQO

 

[Rangers]

Does this point to the low latency of the ESRAM not really being A Thing? Back in the day where this was tooted as an advantage of ESRAM, I argued it can't be that meaningful as no-one at MS etc. was talking about it.

I believe it is an objective thing, pointed out by some specs somewhere.

But considering the only people who would probably really take advantage of it would be MS first party devs, the real world performance impact probably was slim. That just gets more so as more and more platforms get ported to (now PS4 pro, PC etc).

Platform specific optimizations likely just continue to decrease. Now you'll do PC, Scorpio, PS4, PS4 PRo, Xbox, Xbox S versions, to name a few.

Anyways, I never doubted ESRAM was gone but good to know. MS can finally get over their ill fated obsession with on die RAM.

Also I figured Zen wasn't likely, but it's still disappointing.

 

[Rangers]

Clarification issued

https://twitter.com/JezCorden/status/823976558469771264

There's no way they're suddenly going to be like "oh, now we're going to put Zen in there" with 6 months to go. Maybe if it was discrete CPU and GPU, but it's an APU. These things take years to plan out.

Unless you're speculating about things other than zen. I think possible GPU up-flops are a distinct possibility, while hardly to be counted on.