ProspectorPete
Regular
Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]
- Status
Phil said 4.5 times here.
?http://www.gamespot.com/articles/project-scorpio-is-45x-as-powerful-as-xbox-one-phi/1100-6440965/?
56CU's with 54CU's active. I lke your 4X L2 too 14CU X 4 maybe?
Transponster
Newcomer
Scorpio is not going to have any more CUs than little Vega 11, Vega 10 is 12.5 TF, that's 64CUs with over 1.5GHz clock speed, Vega 11 is going to have 40 CU tops, even 36 CUs with higher clocks should be enough to achieve 6TF, but not much past that.
ProspectorPete
Regular
Phil Spencer said a lot of things. He also mentioned uncompressed pictures, unlimited cloud power, best lineup ever, and much more. Never trust a PR guy, especially not for technical information.
Rikimaru
Veteran
1.3TF * 4.5 == 5.9TF
Michellstar
Regular
How so?
They are custom socs, they don´t have to match any Vega, Polaris configuraton. not necessarily I mean.
Of the three things you list, Phil Spencer actually only said one of them ("best lineup ever") and that's subjective. "Uncompressed pixels" was a quote from one of the hardware engineers working on Scorpio and "unlimited cloud power" was never said by anyone but snarky people on the internet.
The 4.5x is an objective measure and specifically refers to the approximate TF difference between XBOne and Scorpio.
Transponster
Newcomer
And i'm not saying they necessarily have to, but it will. I think there's a reason they used either full or cut down existing gpus with all of the current gen consoles, going for more custom, exclusive solution obviously raises cost.
Not to mention that Vega 11 would suffice, as going for higher clocks, even if it's a little bit north of the efficiency sweet spot, is better(/cheaper) than going for more CUs.
Some people seem to think that a console would implode if its gpu reaches 1GHz, even if desktop version is fine at 1.6GHz(Notebook 1080 is achieves over 1.7GHz, btw).
What about 56CUs with 14 per block with 6 disabled it total for increasing yields?
Six disabled would mean six 'blocks' of shaders, current max for GCN is four. Would also increase die area allocated to redundancy. One CU per block for redundancy, max of four blocks. At least, going by history!
Scorpio could easily have more CUs than little Vega 11 if Vega 11 has only 40 or 36 CUs.
GPUs are very modular, custom number of CUs is well within the bounds of what would be sensible in order to manage power and die area considerations. Consoles have additional factors here as they are APUs and not GPUs. Layout will have to be heavily customised and potentially on a different process as it is. And X1 is 14 CUs, PS4 Pro is most likely 40: have these configurations been used elsewhere?
PC GPUs always go higher up the power curve than consoles, because they it makes economic sense. Crappy, noisy coolers that can't maintain turbo under load, expensive coolers for enthusiast models, high air flow, large cases and ATX for factors, practically unlimited power from massively overspecced power supplies and going out of PCI-E specifications ....
Scorpio isn't going to have anything like the clocks of Vega. Vega is 300W with HBM saving several tens of Watts at the very least over a GDDR5 equivalent arrangement. Knock 20% off those Vega clocks (more if the 12+ TF is based on "if lucky" boost clocks) and you might be in the Scorpio ballpark.
Transponster
Newcomer
What about Fiji Pro and XT?Current max for GCN is four. At least, going by history!
And i'm not saying it couldn't, maybe just not (as) "easily".
Which is what i was talking about.
X1 is cut-down Bonaire(260, 360 cards), Pro(36 CUs) is full Polaris 10(RX 480). Like i said, it's hardly a coincidence.
I already did, ~1.2GHz is hardly outlandish given arch gains and node improvements. Likely min. potential of PC Vega is ~1.8GHz, which is still well below that of Pascal.
P.S. If anything, <6TF is more likely than some "56" CUs, after all, 5.5+ rounds up to 6
4 Shader Engines
Fiji is 4 x 16 CUs, same as Vega 10 will be iirc. So far, it's an architectural limit of GCN.
Pro is 36 active CUs (4 x 9), and will be 40 including redundant CUs. Polaris 10 is 36 total - 480 uses all 36 while 470 uses 32.
Custom configurations aren't so uncommon for consoles - both the 360 and the PS3 (additional redundant CUs) had GPU arrangements unlike anything else. It shouldn't be difficult for MS or Sony to have custom CU / shader engine ratios if it suits their design goals.
I had forgotten that Bonaire XT was 14 active CUs and not 16 CUs though tbh.
To hit 12 TF Vega would need to be around 1.46 gHz, if I did my maths correctly. Assuming that's a turbo speed (like the 480's 5.8 TF) then it's a ~14% bump in clock, not bad considering it's on the same node. I wouldn't hold my breath for a 1.8 gHz Vega, short of exotic overclocking.
[Edit: And Vega has the luxury of 300W to the board, with HMB2, and no CPU to factor into things!]
P.S. If anything, <6TF is more likely than some "56" CUs, after all, 5.5+ rounds up to 6
I'm guessing something in the region of 40, 44 or 48 active CUs at 1.0 ~ 1.2 gHz. PS4Pro is less than 75% of the boost clocks of the 480, seems reasonable to expect something similar for Scoprio compared to Vega IMO.
Pro is 36 active CUs (4 x 9), and will be 40 including redundant CUs. Polaris 10 is 36 total - 480 uses all 36 while 470 uses 32.
Custom configurations aren't so uncommon for consoles - both the 360 and the PS3 (additional redundant CUs) had GPU arrangements unlike anything else. It shouldn't be difficult for MS or Sony to have custom CU / shader engine ratios if it suits their design goals.
I had forgotten that Bonaire XT was 14 active CUs and not 16 CUs though tbh.
http://www.eurogamer.net/articles/d...tation-4-pro-how-sony-made-a-4k-games-machine
Sounds counter-intuitive with respect to yields for the 4Pro, but the implication here seems clear enough.
One would almost wonder about taking the Hawaii/R9 290X configuration.
On the other hand, it may just be purely coincidental that FH3 clustered renderer tests were done @4K on such a GPU (closest existing GPU target).
¯\_(ツ)_/¯
Salt.
Interesting! Well, Polaris landed well before PS4 Pro, and he's talking about something coming along in the future ... so 4 shader engine, 40 CU Vega 11?
Maybe Bonaire was the result of years of work on the X1, and same for Pitcairn and PS4....
.... a coincidence, you say .... ⊙﹏⊙
R9 290X did also launch with...
...320GB/s.
Then there's DCC bandwidth compression for graphics & potentially the Vegan™ L2 cache consuming the RBE/ROP output alleviating the shared bandwidth inefficiencies of APU...
/tinfoil hat
edit:
44CU @ 1066MHz ~6TF.
Last edited:
I feel the article author's description of the situation furthers a mischaracterization of the architectures involved, and Cerny's statement is itself difficult to fully square with what we saw.
The PS4 and Xbox One were labelled by console architects as being generally of the CI generation, which Bonaire is the first example to come to market--before both consoles.
Calling the PS4 Pitcairn-based seems to be based on some of the more superficial elements of the GPU like the readily-changed CU count.
The reality seems to be more complicated, with IP elements of the platforms seemingly implemented and deployed on the individual schedules of AMD and the semi-custom clients. We can see things like the DSP block for Trueadio (or similarly semi-separate custom silicon blocks in the consoles), the broader ACE front end, the CI ISA, ESRAM, Sony's volatile flag, Polaris' discard accellerator, DCC, 2x FP16, and other items reaching the public in varying orders.
AMD could opt to roll out certain base items earlier, as evidenced by Bonaire and Polaris having elements that showed up in consoles, even as the consoles had elements that belonged to later GPUs.
seems like they pulled the article =( can't read it now
D
Deleted member 11852
Guest
ScorpioPete says don't trust ScorpioPhil.
Thanks. When you put it that way, then yeah the situation does appear rather more nuanced. I look forward to the inevitable Scoprio leaks to see where the various parts of the system sit in comparison to the PC timelines.
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?
It's still up for me. Perhaps a network filtering issue?
... not to mention the possibility of tiled rasterizing and esram on teh stacked dGPU.
My best guess too, as they'll be keen to stay under 200W. Saving 10 mm^2 is scant consolation if you need a $40 cooler and you sound like a launch PS4. Or worse ... like a launch 360.
- Status
