Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]

[GrimThorne]

It's a given, 64 for cut(56-64 CU), 96 for full(88-96 CU) Vega.

How, exactly? That wouldn't even be 6TF gpu anymore
MS is hardly going to pay for custom(not even semi-custom like X1's) chip.

This discussion reminds me a little of how the 360's architecture eventually worked out. Most who follow this sort of news knew it was going to be in the 1000 family but were pleasantly surprised when it was revealed to be based on an X1800. It also surpassed the expectations of the Xenos leaks. And right after the 360 was launched, ATi released the X1900s.

So it's going to be Vega10 in Q1 next year, followed by Vega 11(confirmed in the Open CL driver in the latest Crimson edition) which will probably be the 88-96 CU card you're talking about? Now I'n not saying it's going to be as exotic as Xenos, but it's probably no accident that Scorpio and Vega 10 are going to be revealed around the same time next year. 6TFs may be the goal but if Vega is a significant mover in performance, no one should be surprised if Scorpio surpasses the proposed hardware specs.

 

[upnorthsox]

And what is Vega? How is it different than Polaris?

 

[Kaotik]

More CUs, texture units and ROPS, sure, clocks? Probably not. Safest bet, I think, is 48 CU / 1GHz

 

[Kaotik]

And what is Vega? How is it different than Polaris?

Vega is AMD's next "aimed at high end architecture". All we know for sure atm is HBM2 support, and that it's gfx ip 9.x, while Polaris was 8.x (just like Tonga and Fiji are)

 

[upnorthsox]

So we know nothing about it and have no idea why MS would pay more for it.

For all we know, Vega = Big and hot, which do not necessarily matter much on a high end architecture but are not desirable qualities in a console gpu.

Now it may well be a Vega gpu but we have nothing but wishful thinking on what that means

 

[Transponster]

You're joking, right? You expect MS to use a standard AMD APU available in 2017?

I meant that they're not gonna pay more for it, than they did for X1's APU, custom design costs more than semi-custom, doesn't it?
And something like that:

More CUs, texture units and ROPS, sure, clocks? Probably not. Safest bet, I think, is 48 CU / 1GHz

screams: "Custom!".


I think the safest bet for Scorpio is an improved Polaris easily capable of achieving 6TF in a console, maybe on a different(not 10nm(inb4 misunderstanding)) node? Still 32 ROPs.
Much like Pitcairn, Polaris is probably here to stay with us until the shrink to 7nm, both Vega and Navi are shooting for higher end, so we should at least expect some refresh of it either way.

 

[Jubei]

So we know nothing about it and have no idea why MS would pay more for it.

For all we know, Vega = Big and hot, which do not necessarily matter much on a high end architecture but are not desirable qualities in a console gpu.

Now it may well be a Vega gpu but we have nothing but wishful thinking on what that means

People are speculating its Vega because it would be wiser to go for a bigger chip and more CUs with a low clock than trying to overclock RX480 in a console sized box. The only drawback would be costs since a bigger die would cost more

Im also not completely sold on Vega being in Scorpio, regardless if they underclock it or not, its going to be sold for 400+ dollars a few months before Scorpio goes into production, granted a big part of the cost is HBM2 but even without that its a big expensive card

 

[dobwal]

I meant that they're not gonna pay more for it, than they did for X1's APU, custom design costs more than semi-custom, doesn't it?
And something like that:

screams: "Custom!".


I think the safest bet for Scorpio is an improved Polaris easily capable of achieving 6TF in a console, maybe on a different(not 10nm(inb4 misunderstanding)) node? Still 32 ROPs.
Much like Pitcairn, Polaris is probably here to stay with us until the shrink to 7nm, both Vega and Navi are shooting for higher end, so we should at least expect some refresh of it either way.

How many amd APUs support 32 mb of esram? The apu in the Xbox one is one of the most customized apu produced by AMD. What makes u think something has changed?

As long as the part is a derivative of AMD PC APUs and mostly use AMD IP it's considered semi-custom. Full custom would probably be a new novel exclusive arch (like the 360 gpu) not just a change in the number of ROPs or CUs.

 

[Kaotik]

People are speculating its Vega because it would be wiser to go for a bigger chip and more CUs with a low clock than trying to overclock RX480 in a console sized box. The only drawback would be costs since a bigger die would cost more

Im also not completely sold on Vega being in Scorpio, regardless if they underclock it or not, its going to be sold for 400+ dollars a few months before Scorpio goes into production, granted a big part of the cost is HBM2 but even without that its a big expensive card

People really need to realize that architectures and chips made on said architecture are not the one and the same.
Polaris isn't RX 480, even Polaris 10 isn't RX 480, RX 480 is just one implementation of Polaris 10 which is one implementation of Polaris architecture. And no console will use RX 480 in any form, ever, anywhere.
Same goes for Vega, obviously. No console will use RX something, they may or may not use Vega architecture, it may or may not have as many CUs as some specific dGPU part, but it will not be that dGPU part regardless.
You could punch in 100 Polaris CUs, or you could throw in just 10 Vega CUs. Architecture does not dictate how big or small implementation you can do with it.

 

[3dilettante]

Architecture will set some parameters for scaling. As an example, Fiji's unit counts happen to be bound by die size and the architectural limits of GCN. 16 ROPs max per shader engine, 16 CUs max per shader engine, 4 shader engines max per GPU.

http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/4

At the time of the Hawaii launch AMD told us that the GCN 1.1 architecture had a maximum scalability of 4 shader engines, and Fiji’s implementation is consistent with that. While I don’t expect AMD will never go beyond 4 shader engines – there are always changes that can be made to increase scalability – given what we know of GCN 1.1’s limitations, it looks like AMD has not attempted to increase their limits with GCN 1.2. What this means is that Fiji is likely the largest possible implementation of GCN 1.2, with as many resources as the architecture can scale out to without more radical changes under the hood to support more scalability.

Polaris has not shown a change for that design limit with the implementations provided so far. Whether Vega will is uncertain.

 

[Kaotik]

Architecture will set some parameters for scaling. As an example, Fiji's unit counts happen to be bound by die size and the architectural limits of GCN. 16 ROPs max per shader engine, 16 CUs max per shader engine, 4 shader engines max per GPU.

http://www.anandtech.com/show/9390/the-amd-radeon-r9-fury-x-review/4

Polaris has not shown a change for that design limit with the implementations provided so far. Whether Vega will is uncertain.

Yeah, sorry for exaggerating there, point being that whether it's Vega or Polaris doesn't mean anything in CU counts etc, and RX 480 is just one iteration of Polaris 10, which is one iteration of Polaris

 

[GrimThorne]

Im also not completely sold on Vega being in Scorpio, regardless if they underclock it or not, its going to be sold for 400+ dollars a few months before Scorpio goes into production, granted a big part of the cost is HBM2 but even without that its a big expensive card

I wasn't immediately sold on Vega either, but there are both statements and upcoming schedules that reinforce the very real possibility that MS is going with Vega.

Phil Spencer @ Eurogamer -

"We could have done a new update this year. We actually looked at it. We went all the way to, we had the spec in front of us, should we ship something that’s less than Scorpio this year, but in truth you can’t do a true 4K console this year. And I just didn’t think anything between what is effectively a 1080p console and the 4K console, like, from a consumer television standpoint there’s nothing in the middle. So let’s go focus on 4K and next year was the right year to do that."

He then talks a little about some of the engineering choices they're making for Scorpio

"Balance in the design of the system is really important. So when you think about the CPU and the GPU that are on the SOC, the optical disc drive and the speed of that, which we haven’t talked about too much, but I’ll just say that, the speed of RAM, all of these things when you’re designing the system, making sure they’re all in balance, so one piece doesn’t get overpowered relative to the other, or you’ve got a lot of GPU but you’re not able to feed all of the assets to it out of the RAM fast enough, all of those things were critical for us when thinking about the design.

All of this is why we picked the design point we did of 4K and the date we did. Like I said, we looked at other designs that might be coming earlier, and we don’t think they really delivered on a full point."

Are these definitive statements regarding Vega? Hardly. But he's making his point very clear, there's better tech coming down the road in 2017 that would fit more into their design for what they have planned with Scorpio.

As for size and heat.............well right now the R9 Nano at 28nm is a 6inch card with a 175wt TDP. If it were to receive the node shrink benefits of 14/FF as a Vega 10 processor (which for Scorpio is likely to be a cut down version) , then it's very easy to comprehend why it might be so appealing to Spencer.

And while we're at it, such a card would also be a game changer for PC gaming as well

 

[Deleted member 13524]

Most who follow this sort of news knew it was going to be in the 1000 family but were pleasantly surprised when it was revealed to be based on an X1800.

Xenos isn't based on the R520/X1800 (which isn't even an unified shader architecture) and it's quite far from 2007's Terascale architecture.
The closest non-console architecture there is to Xenos are Qualcomm's Adreno GPUs, at least the 200 series.

And what is Vega? How is it different than Polaris?

From what we know so far, Vega is simply Polaris with the GDDR5 controller replaced with a HBM2 one.
That would make Scorpio (which seems to be using GDDR5) essentially a Polaris architecture + whatever (semi-)custom features they decide to add.

 

[AlNom]

One would hope they'd be pushing updated FeatureLevel support.

 

[Kaotik]

From what we know so far, Vega is simply Polaris with the GDDR5 controller replaced with a HBM2 one.
That would make Scorpio (which seems to be using GDDR5) essentially a Polaris architecture + whatever (semi-)custom features they decide to add.

No, it isn't. AMD's OpenCL drivers already confirmed that Vega is gfx ip 9.x (which was claimed ages ago at linkedin profile of ex(?)-AMD employee), while Polaris is gfx ip 8.x just like tonga and fiji

 

[Deleted member 13524]

No, it isn't. AMD's OpenCL drivers already confirmed that Vega is gfx ip 9.x (which was claimed ages ago at linkedin profile of ex(?)-AMD employee), while Polaris is gfx ip 8.x just like tonga and fiji

Okay, I wasn't aware of that.
But does that mean anything? Did it mean anything in the past?

 

[Jubei]

People really need to realize that architectures and chips made on said architecture are not the one and the same.
Polaris isn't RX 480, even Polaris 10 isn't RX 480, RX 480 is just one implementation of Polaris 10 which is one implementation of Polaris architecture. And no console will use RX 480 in any form, ever, anywhere.
Same goes for Vega, obviously. No console will use RX something, they may or may not use Vega architecture, it may or may not have as many CUs as some specific dGPU part, but it will not be that dGPU part regardless.
You could punch in 100 Polaris CUs, or you could throw in just 10 Vega CUs. Architecture does not dictate how big or small implementation you can do with it.

You are correct but imo we should be able to use layman terms in this subforum even if they are technically incorrect. Xbox One having a 7790 with 2 CUs deactivated is easier to understand and compare to than saying its using a custom Cape Verde

So while Neo obviously wont be using the exact RX480 GPU being sold right now, its easier to understand by comparing to it rather than Polaris uarch with x amount of CUs clocked at x speed. Its just simpler to say its a downlocked custom 480

 

[GrimThorne]

Xenos isn't based on the R520/X1800 (which isn't even an unified shader architecture) and it's quite far from 2007's Terascale architecture.
The closest non-console architecture there is to Xenos are Qualcomm's Adreno GPUs, at least the 200 series.

Really?! I was always under the assumption it was a X1800 XL. What was it based upon if you don't mind sharing? I read through numerous articles back then, T-Report, Ars, Anand, Semi, [H]ardocp etc. Xenos was always described as unique but I always thought it was based on something in the ATi 1000 series.

 

[Jubei]

I wasn't immediately sold on Vega either, but there are both statements and upcoming schedules that reinforce the very real possibility that MS is going with Vega.



Are these definitive statements regarding Vega? Hardly. But he's making his point very clear, there's better tech coming down the road in 2017 that would fit more into their design for what they have planned with Scorpio.

As for size and heat.............well right now the R9 Nano at 28nm is a 6inch card with a 175wt TDP. If it were to receive the node shrink benefits of 14/FF as a Vega 10 processor (which for Scorpio is likely to be a cut down version) , then it's very easy to comprehend why it might be so appealing to Spencer.

And while we're at it, such a card would also be a game changer for PC gaming as well

Right but the Nano was still a 650 dollar card. The Vega successor to R9 Nano is probably not going to be 199 dollars. And while it will improve perf/watt drastically compared to R9 Nano, i think total wattage is still going to be atleast 150 watts with HBM2.

To reach 6 TF in 2017, it would require a big die (in comparison to Neo or past consoles) with a lot of CUs and its still going to draw more power than the 36 CUs in Neo, thats just logical if the uarch is similar

And thats without Zen cores, i have no idea how much they add in power, the 8 core desktop Zen is supposedly 95 watts, the only leaked APU so far was 45 watts with 4 cores and 11 CU Polaris GPU

 

[Deleted member 13524]

What was it based upon if you don't mind sharing?

Well some of this is my own speculation on the matter:
The R500 IIRC was a project that was initially meant for replacing the R400 in the PC space (which by itself was a rather small evolution over R300, meaning ATi's resources had been centered elsewhere), but was put aside circa 2004. Supposedly, back then the transition to unified shaders and an effective use of eDRAM in Windows could be a huge hassle driver-wise, so instead they sold the VEC4+Scalar architecture to Microsoft, who was looking for something with a lot of potential inside a closed environment. Then R520 and R580 came with non-unified shaders and by late 2006 they figured Terascale (VLIW5) would be more appropriate for the desktop.
Xenos' VEC4+Scalar was still used for ATi's Imageon line of GPUs for handhelds (back then almost exclusively used by Intel for their ARM-based SoCs called XScale), specifically for their next-generation OpenGL ES 2.0 Imageon called Z430, or Mini-Xenos. Qualcomm used ATi's Imageon for their MSM7xxx SoCs and the Z430 went into MSM7227 and the first-ever mobile SoC with a 1GHz CPU called Snapdragon.
When Qualcomm bought the handheld GPU division from ATi (by then already a part of AMD IIRC), they changed the Z430 name to "Adreno", which is actually an anagram of "Radeon".