Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]

[BoardBonobo]

Thinking about it I'd be surprised if they ever actually launch the scorpio in the form they're talking about. I mean at the moment it's just vaporware and a very useful tool in attempting to derail any uptake of the PS4Pro (which seems to have done very well in pre-orders) as once that hits the market MS haven't got anything current in the console space apart from XB1 and that is already lagging far behind. There's a good chance their market share will go the way of their mobile business.

 

[Jubei]

Vega 10 = 64 CU, 12 TF, 16 GB HBM2 and 225W TDP

http://videocardz.com/63700/exclusive-first-details-about-amd-vega10-and-vega20

Looks like it will be a massive chip.

 

[Shifty Geezer]

According to the article, it's from the server roadmap, so a server processor. Way too big and expensive for a console, suggesting Scorpio is going to be ~same architecture as XB1/PS4 Pro.

 

[function]

They'll be compared against Sony 1st party offerings, tragically, but also 3rd party games on Scorpio. And if these look better than MS's first party because they eschew native rendering for reconstruction, MS will just look a bit stupid - either by first party not being as pretty as third party, or back-tracking on this statement.

I think in the initial stages, to early adopters, just knowing that their console is doing native 4K will have some value. Particularly as people looking for differences will be sticking their noses right up against 55" screens. It also means that MS can't be called out for abandoning X1 gamers - these will be X1 games going from 900/1080p to 4K.

And while it may be a waste of power in the general case, and it's not what I want, it still won't be that half of the machines GPU power is wasted. It'll still produce getter results than a half res upscale, with a sharper image for both edges and textures (going by PC comparisons so far at least), it's just that diminishing returns mean that you could better have spent the power elsewhere ....

 

[function]

According to the article, it's from the server roadmap, so a server processor. Way too big and expensive for a console, suggesting Scorpio is going to be ~same architecture as XB1/PS4 Pro.

New architectures normally hit the consumer space well before they hit the server/workstation market. I think all AMD's GPUs due next year are graphics generation 9, compares to Polaris' gen 8.1.

It's possible Scorpio will be based on the older Polaris, but given how close AMD's consoles have usually been to the latest architectures and that Scorpio is coming out months after Vega I think Vega is much more likely.

No idea what advances it might bring though!

 

[Jubei]

I interpreted it as the roadmap came from AMDs internal server. But nevertheless i agree with your point

 

[Silent_Buddha]

Vega 10 = 64 CU, 12 TF, 16 GB HBM2 and 225W TDP

http://videocardz.com/63700/exclusive-first-details-about-amd-vega10-and-vega20

Looks like it will be a massive chip.

That's not the interesting thing from that article.

The piece that's perhaps most relevant to Project Scorpio.

We have also learned that AMD is planning to replace Polaris 10 with Vega 11 next year.

That should indicate at least some significant change between Vega and Polaris. Otherwise AMD would just continue to use Polaris 10 in that market segment as they have done for the past few generations. For example, R9 290 and R9 390 use the same GPU.

As well considering the price bracket that Polaris targets, it's highly unlikely that HBM 2 will be used for Vega 11 (replacing Polaris 10) compared to Vega 10. Hence indicating that Vega would have no problems using some form of GDDR which will be important for Project Scorpio.

Wild speculation here. But it's possible that Vega 11 will have a significant perf/watt advantage over Polaris 10 making Polaris 10 redundant. That would also make it far more feasible for Microsoft to have a 6 TFLOP GPU in Project Scorpio without going over 200 watts. Depending on how much of an improvement there is, it may be possible to get near the power envelope of the PS4-P.

Regards,
SB

 

[Shifty Geezer]

New architectures normally hit the consumer space well before they hit the server/workstation market. I think all AMD's GPUs due next year are graphics generation 9, compares to Polaris' gen 8.1

Yeah, looking around it seems this spec is on the consumer roadmap, although it'll be a high end card. I don't know what scaling options there'll be though. AFAIK there's no word on a 'Vega' APU (Zen + Vega?).

 

[Silent_Buddha]

Yeah, looking around it seems this spec is on the consumer roadmap, although it'll be a high end card. I don't know what scaling options there'll be though. AFAIK there's no word on a 'Vega' APU (Zen + Vega?).

The article also includes information that Vega 11 will be replacing Polaris 10 in the product stack. My post probably came in at the same time you posted.

Regards,
SB

 

[function]

Yeah, looking around it seems this spec is on the consumer roadmap, although it'll be a high end card. I don't know what scaling options there'll be though. AFAIK there's no word on a 'Vega' APU (Zen + Vega?).

There's a Gen 9 (Vega generation) "Greenland" HPC APU, but that's a beast with a big GPU with interposer and two stacks of HBM2 on the same chip package as a 16 core Zen (with it's own 256-bit memory interface). It would make an incredible console, at incredibly unaffordable prices!

There is a Zen SoC APU for somewhere around the middle of next year, which is a single quad core (8 thread) Zen module on the same chip as 12 CUs (same number as active in X1), presumably with only a 128-bit DDR4 bus to feed it. With DDR4 3200 and colour compression, this should be a big step up from current APUs, and in many ways could exceed the X1 ... but in practice it's likely to be hampered enormously by very limited bandwidth (less than the X1 even if you exclude the esram and then factor in ideal use of the colour compression). Don't know it his is gen 9 graphics (Vega) or 8.1 (Polaris) though, as APUs can trail a little on implementation of new graphics IP.

Edit: low cost HBM that simplifies the stack and removes the need for an interposer can't come soon enough for AMD. Their next gen APUs would blitz X1S and be on the tail of the 460 if they simply had the BW.

 

[function]

I dont see a significant perf/watt improvement for Vega, RX480 is 5,5 TF, 150 watts and 32 CUs, Vega 10 is 64 CU, 12 TF and 225 watts

Double the CUs and you have 11 TF, increase clockspeed and you have 12 TF, add in HBM2 and you get less power draw and much more bandwith.

Well to be fair, the 480 was actually hitting 170+ W and going beyond spec for both the PCI-E slot and the PCI-E 6 pin power connector .... at the same time.

If Vega 10 can stay under 225 W (8 pin + slot?) in normal use then I'd say that's a pretty big jump in per/watt. Greater than 100% more processing power, for less than 50 % more power consumption. Given that there's no node jump here, I'd say that would be very good.

 

[Jubei]

Well to be fair, the 480 was actually hitting 170+ W and going beyond spec for both the PCI-E slot and the PCI-E 6 pin power connector .... at the same time.

If Vega 10 can stay under 225 W (8 pin + slot?) in normal use then I'd say that's a pretty big jump in per/watt. Greater than 100% more processing power, for less than 50 % more power consumption. Given that there's no node jump here, I'd say that would be very good.

But Vega 10 is not using GDDR5, its using HBM2 wich draws far less power. So how can you conclude its a pretty big jump in perf/watt without excluding that variable? It would be a fantastic feat by AMD to release a new architecture only a year after their last and get a significant improvement on perf/watt on the same node

 

[Shifty Geezer]

There is a Zen SoC APU for somewhere around the middle of next year, which is a single quad core (8 thread) Zen module on the same chip as 12 CUs (same number as active in X1), presumably with only a 128-bit DDR4 bus to feed it

So nothing Scorpio level? Will Scorpio have to be discrete CPU and GPU?

 

[function]

But Vega 10 is not using GDDR5, its using HBM2 wich draws far less power. So how can you conclude its a pretty big jump in perf/watt without excluding that variable? It would be a fantastic feat by AMD to release a new architecture only a year after their last and get a significant improvement on perf/watt on the same node

It's the total board power that I think is most interesting - and probably most interesting to customers - as the chip is useless without the memory it's architected for. By putting the work into enabling HBM2 (and indeed pioneering HBM use) they're still responsible for the power drop from their products.

I don't see any point in excluding HBM2 from any attempt to look at the chip as a whole (looking a individual subsystems is a different matter).

 

[function]

So nothing Scorpio level? Will Scorpio have to be discrete CPU and GPU?

Doesn't appear to be anything, Scorpio will have to be a semi-custom solution and nothing off-the-shelf.

Scorpio could be a separate GPU and CPU, but the render from the reveal and the comments of "6 TF SoC" in the video do seem to indicate it's a single chip solution, and probably quite a large one.

 

[Jubei]

It's the total board power that I think is most interesting - and probably most interesting to customers - as the chip is useless without the memory it's architected for. By putting the work into enabling HBM2 (and indeed pioneering HBM use) they're still responsible for the power drop from their products.

I don't see any point in excluding HBM2 from any attempt to look at the chip as a whole (looking a individual subsystems is a different matter).

Thats fair but the original argument was in regards to a console sized box that wont be using HBM2, i think you would then agree that Vega wont have a signficant improvement in perf/watt over Polaris without HBM2

Assuming the videocardz source is accurate, you can see that the HBM2 is actually only 512 GB bandwith instead of 1 TB wich suggests that AMD is clocking the HBM2 memory very low to reach that 225 watt TDP.

 

[mpg1]

Most mind blowing is Vega 20...7nm..24Tflops..32GB HBM2...TDP of 150??!

What year is that coming out? Has to before Navi right?

That's your next gen console right there.

Edit: wait got mixed up. 24 Tflops is for Vega 10.

 

[function]

Thats fair but the original argument was in regards to a console sized box that wont be using HBM2, i think you would then agree that Vega wont have a signficant improvement in perf/watt over Polaris without HBM2

Won't know until we see a GDDR5 variant of the architecture. My expectation would be that it's more power efficient that Polaris, but that without HBM2 perf/watt will definitely drop.

Assuming the videocardz source is accurate, you can see that the HBM2 is actually only 512 GB bandwith instead of 1 TB wich suggests that AMD is clocking the HBM2 memory very low to reach that 225 watt TDP.

Could also be two HBM2 stacks at full speed, which would give you up to 16 GB of memory. This is actually what Greenland (the HPC APU) appears to use as its memory arrangement, going by leaked slides.

 

[function]

Most mind blowing is Vega 20...7nm..24Tflops..32GB HBM2...TDP of 150??!

What year is that coming out? Has to before Navi right?

That's your next gen console right there.

That's about as likely as Vega 10 being in Scorpio next year.

 

[mpg1]

How the hell do you get 24 Tflops? Must be dual GPU card.

Edit: nevermind article says dual Vega 10 is coming in fall 2017.

Now more confused.