AMD - Piledriver chitchat

[CAT-THE-FIFTH]

They're less about Ryzen and more about infinity fabric, which looks very interesting indeed. A coherent data and control fabric that sits in their entire product range and handles all the connectivity between segments of chips looks great, and the data fabric is clearly what Vega is building on with its HBM2 "high bandwidth cache" with coherent access to system RAM and non-volatile storage too. And I note that raven ridge is getting the fabric ("high performance scalable bus") - wonder what that'll mean for APUs? Possible on-package HBM cache might not be a pipe dream after all Even a single stack of HBM1 would be an amazing last-level cache...

It does make me wonder whether AMD was trolling everybody by saying they would have a 40% IPC increase with Ryzen and now it apparently has 55%,or a 40% further improvement over their initial target.

[watercooled]

AMD sampling 4C/4T Ryzen:

Ughhhh why bother? That's Intel-style, artificial product segmentation nonsense. With 4, 6 and 8 cores they already have enough real variants to flesh out their product line, surely?

I guess it could allow them to address more of the market before the Zen APUs arrive, but I'd imagine there's a limit to how low they can go (profit-wise) with a native 8C die unless they're actually planning to release a 4C die too?

Edit: Just spotted this: http://www.pcgamesn.com/amd/amd-ryzen-overclocking

Seems Ryzen F4 stepping is at 3.6/4GHz - higher base and identical boost to the 6900k now! Assuming it's for the 8C variant, that's promising given the seemingly similar IPC.

[kalniel]

Ughhhh why bother? That's Intel-style, artificial product segmentation nonsense. With 4, 6 and 8 cores they already have enough real variants to flesh out their product line, surely?

I guess it could allow them to address more of the market before the Zen APUs arrive, but I'd imagine there's a limit to how low they can go (profit-wise) with a native 8C die unless they're actually planning to release a 4C die too?

I think it'll be to hit a particular TDP target if so - they might have just been close and need the extra headroom disabling SMT gives.

[watercooled]

Ah so maybe it's more of a power-budgeted part rather than just a restricted-performance part? I hadn't looked at it that way.

[watercooled]

Seems we'll be able to buy a fully-enabled Polaris 11 soon: http://videocardz.com/65594/sapphire...eam-processors

[Xlucine]

Doesn't bode well for the hopes of a 460X or 465 card, since it'd be odd for sapphire to beat AMD to the punch with a card with the wrong name

[scaryjim]

... it'd be odd for sapphire to beat AMD to the punch with a card with the wrong name

GPU manufacturers have been releasing all sorts of hybrid cards for a few generations now. The hexus reporting of that story speculates that the card will only be released in China.

It does seem like the chinese market gets a much wider degree of product differentiation than others - it already has the 1792sp RX 470D, after all. Possibly it's seen as a testing ground for the potential popularity of a product/market segment.

Given the RX 465 should have to be a genuine revision, it could easily be that any 1024sp RX 460s will have to keep the original naming, and it won't be until we get later/improved steppings that we'll get the naming updated. They may also have to be partner-specific cards, as I doubt the fully enabled die can run at AMD's specified clockspeeds within a 75W TDP. That's the main reason I suspect we'll eventually see a fully-enabled official part with the RX 465 designation - AMD will want it to remain a nominally bus-powered card so they'll refresh the naming when they hit a stepping that'll give them a fully enabled, bus-powered card running at 1200MHz+...

[CAT-THE-FIFTH]

Actually we have already seen an "RX570" appear in laptops which is basically an RX470,so it is quite possible when Vega launches,the 1024 shader RX460 might be called an RX560 or RX550.

[Xlucine]

It'll be the 3XX series all over again!

[watercooled]

It's not unusual to see 'refreshes' like that on the mobile side - as I understand it, the manufacturers all but demand it to match their release cadence so it looks like they've changed something since last year's version. So I wouldn't assume seeing it in a laptop implies a similar thing on the desktop. Not unless they plan to change *something*, and it would risk taking the wind out of an actual GPU release unless they don't have one planned for a while.

Then again, maybe they want a refresh to go along with the AM4 platform release?

[watercooled]

Has anyone been more in-the-loop WRT Vega? Of course, we haven't seen much yet but some of the stuff we have seen doesn't seem to add up, and these videos raise a few interesting points:
https://www.youtube.com/watch?v=UuchUscHWSw
https://www.youtube.com/watch?v=C4BUb6wSSXk

I can't really disagree with what he's saying about the demos and AMD's typical marketing stance (i.e. not sandbagging), but the performance of that DOOM demo doesn't add up, or at least his interpretation that it's close to final performance.

First of all, it's a really large die and it's on 14nm - it really doesn't make sense that its performance should be where it is; less space is taken up for HBM2 memory controllers vs GDDR5. Nvidia does seem a bit more space efficient than Polaris given Pascal's considerably higher achievable clock speeds, but still in the right ballpark. E.g. 480 vs 1060 - similar die size, similar performance. Here, you're talking about something well over twice the size of the 480, with ~50% higher clock speed, HBM2 memory, vast architectural changes... having less than twice the performance??? Nah.

Even if they'd literally just shrunk Fiji and clocked it higher, we should be getting better performance than we're seeing, and the die size is FAR too big for that to be the case. I'd estimate a 4096 shader version of Polaris to be something like 400mm2, and even smaller with HBM2. Methinks non-final clocks and/or very, *very* early drivers. Meh, we're missing something either way.

However, what is concerning is the 8GB HBM2 we keep seeing. I hope either that's wrong, or their caching system works well, or they're just leaving themselves open to attack like happened with Fiji.

[Xlucine]

8GB isn't too bad - people are still selling 3GB cards for more than £200, after all. 1080 has 8GB too, and seems to be the prime competitor for vega

[DanceswithUnix]

I think these are more of a compute part than a gaming part, in which case you can't compare gaming performance per unit area.

It also sounds like improvements have been made on memory usage, so the 8GB will be more effective even though it is enough for the foreseeable future anyway.

[watercooled]

8GB isn't too bad - people are still selling 3GB cards for more than £200, after all. 1080 has 8GB too, and seems to be the prime competitor for vega

I'm thinking purely in marketing terms. IIRC the Fury showed no signs of struggling with 4GB but it was still given the thumbs down in reviews with no real justification.

I think these are more of a compute part than a gaming part, in which case you can't compare gaming performance per unit area.

Perhaps. Even with that in mind the die size doesn't seem to make sense though, and nor does the performance - even with a clock speed bump and identical performance per clock it should be doing far better than it is.

It also sounds like improvements have been made on memory usage, so the 8GB will be more effective even though it is enough for the foreseeable future anyway.

I agree. People would still complain about it though.

[DanceswithUnix]

S/A has some technical stuff about Vega: http://semiaccurate.com/2017/01/17/a...ga-high-level/

[CAT-THE-FIFTH]

I just saw this review of the RX470 tested with the G4560 and a few Core i5 and Core i7 CPUs:

https://www.youtube.com/watch?v=dichjs9HXTg




Apart from one game,there seems to be virtually no difference in the percentage drop between the G4560 and a Core i7 7700K using the RX470 or a GTX1060,and in some cases the GTX1060 has a worse drop in framerates(!!).

I remember reading somewhere,the ReLive drivers improved GPU overhead,so it does seem this might be the case.