AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

[3dilettante]

4GB is going to hurt.

http://techreport.com/review/28294/amd-high-bandwidth-memory-explained/2


So, erm, AMD's promising it won't hurt. Driver technology for the win

I give kudos for his candor. I do not recall an AMD statement that included its drivers and the phrase "exceedingly poor" before. That was usually relegated to its customers and reviewers. It also admits assigning a couple of engineers, so I'm willing to bet there was a significant if temporary uptick/doubling/trebling in developer hours put into its drivers that share a neighborhood with the phrase "exceedingly poor", so *golfclap*.

I'd imagine there is some leeway at 4GB capacity for a while for much of the market, but the longevity is less assured with the threat of reverting to bus transfers hanging over every driver hiccup.
I have serious doubts as to what customers of the pricier Firestream cards would find tolerable.

 

[Jawed]

I give kudos for his candor. I do not recall an AMD statement that included its drivers and the phrase "exceedingly poor" before. That was usually relegated to its customers and reviewers. It also admits assigning a couple of engineers, so I'm willing to bet there was a significant if temporary uptick/doubling/trebling in developer hours put into its drivers that share a neighborhood with the phrase "exceedingly poor", so *golfclap*.

I'd imagine there is some leeway at 4GB capacity for a while for much of the market, but the longevity is less assured with the threat of reverting to bus transfers hanging over every driver hiccup.

He gives the impression that the driver is overtly wasteful. Which could imply that there's a lot of low-hanging fruit.

But everyone's going to be taking this GPU to the max in games, trying to make it slower than Titan X. So we'll see.

I'm reminded of how awful, for how long, GCN drivers (compiler particularly) were.

Well, perhaps AMD's spent the last 9 months working this out. That's how late I reckon it is...

I have serious doubts as to what customers of the pricier Firestream cards would find tolerable.

They'll prolly never see an HBM1 chip (Firestream doesn't exist any more, by the way - I'm assuming you're referring to compute cards like FirePro W9100, which has 16GB). Also, in compute, the driver doesn't get in the way to quite the extent it does in games, when it comes to memory (in theory).

Finally, the drivers for D3D12 need a huge overhaul, because game developers are likely to go "give me 75% of the memory" and then do their own allocation management within that blob.

Though one can argue that it's never possible to do this stuff without the driver having the final say.

 

[Jawed]

While I'm quoting TechReport, I suppose it's worth adding this bit:

One problem with HBM is especially an issue for large GPUs. High-end graphics chips have, in the past, pushed the boundaries of possible chip sizes right up to the edges of the reticle used in photolithography. Since HBM requires an interposer chip that's larger than the GPU alone, it could impose a size limitation on graphics processors. When asked about this issue, Macri noted that the fabrication of larger-than-reticle interposers might be possible using multiple exposures, but he acknowledged that doing so could become cost-prohibitive.

So it would seem very likely that Fiji is smaller than GM200.

 

[3dilettante]

He gives the impression that the driver is overtly wasteful. Which could imply that there's a lot of low-hanging fruit.

But everyone's going to be taking this GPU to the max in games, trying to make it slower than Titan X. So we'll see.

I'm reminded of how awful, for how long, GCN drivers (compiler particularly) were.

It may be wasteful enough for now, until titles or higher resolution modes make the genuine capacity demand exceed the limits of their optimizations.
I think things are generally not constrained enough outside of the most demanding cases--which may need a next-next gen GPU or Crossfire to make playable--but it's not an enviable position to be in.
Perhaps AMD is counting on DX12 to bail them out, although Mantle under the current WDDM showed increase memory consumption.

Nvidia could "pull an AMD" and jack up the recommended resolution until its competition if memory-constrained, although if I recall correctly AMD's cards were not as sensitive to this as Nvidia's were.

They'll prolly never see an HBM1 chip (Firestream doesn't exist any more, by the way - I'm assuming you're referring to compute cards like FirePro W9100, which has 16GB).

That was a mistake on my part. I was thinking any FirePro in the compute or professional graphics market that have expanded memory capacity.

 

[Kaotik]

While I'm quoting TechReport, I suppose it's worth adding this bit:

So it would seem very likely that Fiji is smaller than GM200.

We don't know which process they'll use to make the interposer, so how could we know the max reticle size for said process either? At least earlier, if my memory serves me, the interposer has been mentioned to use older process than the chips on top of it?

 

[AlNom]

We don't know which process they'll use to make the interposer, so how could we know the max reticle size for said process either? At least earlier, if my memory serves me, the interposer has been mentioned to use older process than the chips on top of it?

65nm seems to be thrown around.

 

[silent_guy]

Do the memory mismanaged comments from AMD sound like a case of under promising to prepare us for disappointment?

I'm still hoping on that special 8GB hack.

 

[3dilettante]

While I'm quoting TechReport, I suppose it's worth adding this bit:
So it would seem very likely that Fiji is smaller than GM200.

This seems to go back to the interposer dimensions being <<70mm in either dimension, the old assumptions of a roughly 860mm2 maximum are not being disproven.
AMD provides a 7x5mm HBM module in its slides, but this may be neglecting to mention the slightly larger base die's footprint.

We don't know which process they'll use to make the interposer, so how could we know the max reticle size for said process either? At least earlier, if my memory serves me, the interposer has been mentioned to use older process than the chips on top of it?

The optics of the equipment are constant. There are expensive wider-field methods, and there are products that connect wider than reticle wiring, but 65nm is several nodes too small compared to those.

 

[Ryan Smith]

That's not the message I get from Anandtech's article at all:



Ryan doesn't even hide the fact that AMD does have something up their sleeves. Otherwise, he'd write "we'll see if they have something".
Maybe I'm reading too much into it, but he's around to put me in my place if that's the case

I have no clue what AMD is going to do. All I know is that AMD hasn't announced final SKUs yet, so it's a bit early to make assumptions.

 

[3dilettante]

Do the memory mismanaged comments from AMD sound like a case of under promising to prepare us for disappointment?

I'm still hoping on that special 8GB hack.

Not knowing Macri's mannerisms, it might be a blunt characterization of the situation.
The most charitable interpretation is that it is his opinion that the vast majority of grapcics workloads have a lot of slack that they can work with for a while.
The enthusiast space that AMD is courting so heavily with everything else it tries to use to justify its high end may reach the end of that slack more quickly, shrinking that time window.
With Nvidia's desire to dump cold water on AMD's product launch, there may be further PR shrinkage.

If there is manipulation, it might be a bluff based on current truth.
State that a snapshot of how things are now is decently served by 4GB by painting a poor picture of the current driver management to create an impression that there's enough that can be done. Then don't talk about the impact to product longevity before either a new SKU or 2016 comes along.

It could be sandbagging, if there is any truth to the 8GB hack.

 

[LordEC911]

From raghu78:
www.amkor.com/go/Kalahari-Brochure

Test vehicle of a ~832mm2 interposer for a < 350mm2 main die.

 

[Jawed]

Yield figures from that PDF are interesting:

■ Interposer die yield (TSV + BEOL): > 99%
■ Interposer backside (TSV daisy chain) continuity and leakage test yield: > 99%
■ Interposer BSI yield: ~ 98%
■ Assembly yield: ~ 95%

Along with a date of December 2013 on the assembly, I dare say yield is not looking like a disaster.

 

[silent_guy]

It's going to be a fun month of June. Yum.

 

[eastmen]

I was actually getting interested in the r390x ... but 4GBs will kill it for me. Oh man this week is taking the wind out of my sales

 

[gamervivek]

Seems be doing quite the same for AMD.

 

[Razor1]

4gb isn't enough for next gen games at highest texture settings let alone any anti aliasing........

 

[silent_guy]

Along with a date of December 2013 on the assembly, I dare say yield is not looking like a disaster.

If those yields are cumulative then that results in a 93% or lower yield. That's not a disaster, but it's not great either, and quite a bit lower than typical packaging yields which are in the high nineties. In addition, a packaging failure may not only destroys the main die but all 4 HBM stacks as well. (I wonder if they package everything in one go, or if it you have separate phases, with continuity tests in between.)

 

[liquidboy]

4gb isn't enough for next gen games at highest texture settings let alone any anti aliasing........

??? what makes you think that ... a four stack HBM GPU should have more than enough BW and capacity to satisfy any modern game at 4K resolutions

 

[eastmen]

??? what makes you think that ... a four stack HBM GPU should have more than enough BW and capacity to satisfy any modern game at 4K resolutions

you think your going to be able to fit a 4K framebuffer and textures into 4gigs ? What about FSAA ?

 

[Alexko]

I was actually getting interested in the r390x ... but 4GBs will kill it for me. Oh man this week is taking the wind out of my sales

Are you some sort of wind-propelled retailer?