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

[Alexko]

The top to bottom phrasing is not particularly well-written, although I can see several ways that a new product range could have HBM throughout its covered range even if that range does not include every part of AMD's overall market.

Fiji has 4096 shaders and is about 600mm². One may (somewhat simplistically) suppose that on 14nm, it would be about 300mm², therefore that a 14nm semi-Fiji would be around 150mm², with 2048 shaders, just like Tahiti. With a single stack of HBM2, it would have about 256GB/s of bandwidth, which is a good bit more than Tonga but slightly less than Tahiti in its fastest variant.

So yes, I think it's reasonable to assume that AMD won't bother making a new GPU under 150mm², and probably just recycle Pitcairn—again!—and Bonaire.

 

[silent_guy]

what is dumb about it ?

Check my earlier posts...

It doesn't make sense to favor an expensive technology for size when the same performance can be reached much cheaper with GDDR5. It makes much more sense to gradually make HBM migrate down the stack with process improvements:
28nm: HBM doesn't make much sense, except as band-aid for an inefficient core architecture (mostly power, the perf benefit remains an open question.)
16/14nm: HBM2 makes sense for the highest-end silicon.
10nm: HBM2 makes sense for the highest end and the level below it.

The biggest counter-argument is indeed size, if that suddenly becomes a bigger deal than cost. IMO, the FuryX doesn't apply here because it still requires an external fan, and the Fury shows that this argument doesn't work for air cooling. Furthermore, the Fury Nano will likely have terrible perf/mm2. Let's also not forget that miniature size GTX 970 cards have existed as well that are similarly sized as a FuryX. I don't think physical size has been a major point of engineering attention up to now, and that significant improvement is possible with non-HBM solutions if they put their mind to it.

 

[Alexko]

Check my earlier posts...

It doesn't make sense to favor an expensive technology for size when the same performance can be reached much cheaper with GDDR5. It makes much more sense to gradually make HBM migrate down the stack with process improvements:
28nm: HBM doesn't make much sense, except as band-aid for an inefficient core architecture (mostly power, the perf benefit remains an open question.)
16/14nm: HBM2 makes sense for the highest-end silicon.
10nm: HBM2 makes sense for the highest end and the level below it.

The biggest counter-argument is indeed size, if that suddenly becomes a bigger deal than cost. IMO, the FuryX doesn't apply here because it still requires an external fan, and the Fury shows that this argument doesn't work for air cooling. Furthermore, the Fury Nano will likely have terrible perf/mm2. Let's also not forget that miniature size GTX 970 cards have existed as well that are similarly sized as a FuryX. I don't think physical size has been a major point of engineering attention up to now, and that significant improvement is possible with non-HBM solutions if they put their mind to it.

You're still neglecting laptops and thermal constraints.

 

[silent_guy]

You're still neglecting laptops and thermal constraints.

Yes, I do.

 

[tunafish]

Fiji has 4096 shaders and is about 600mm². One may (somewhat simplistically) suppose that on 14nm, it would be about 300mm², therefore that a 14nm semi-Fiji would be around 150mm², with 2048 shaders, just like Tahiti. With a single stack of HBM2, it would have about 256GB/s of bandwidth, which is a good bit more than Tonga but slightly less than Tahiti in its fastest variant.

So yes, I think it's reasonable to assume that AMD won't bother making a new GPU under 150mm², and probably just recycle Pitcairn—again!—and Bonaire.

This very much depends on the economics of 14nm vs 28nm, and the expected size of the laptop market that would like Pitcairn-level perf at much lower power use. If that market is small enough (just Apple?) then satisfying it with that 150mm² die severely underclocked like the Fury Nano might be a good idea. If it's bigger, reducing the manufacturing cost might be worth it.

In any case, the design + mask work per die type for 14nm will be much more than it was at 28nm, it's sane to expect fewer designs even if that means being slightly less efficient per mm². A full lineup made of only two distinct dies (with plenty of harvested variants) might not be that insane.

 

[3dilettante]

An unanswered question at this time would be volumes at the traditional low end, for which 16nm and HBM may have their individual uncertainties in scaling from the numbers of Fury chips to the volumes of AMD's typical cheap mobile products or the low-end discrete cards.

There does seem to be a number of players that have been waiting a long time for the next useful node, so getting the necessary allocation may be challenging.
Perhaps AMD's numbers are just that low in an absolute sense, or they could sell things through in brackets that wouldn't see much benefit, like dual-graphics pairings with their 28nm APUs.

 

[Deleted member 13524]

So yes, I think it's reasonable to assume that AMD won't bother making a new GPU under 150mm², and probably just recycle Pitcairn—again!—and Bonaire.

Why not? 150mm^2 in 28nm gets you a GM107, which is ~3x faster than the fastest that Intel has ever put out there as an iGPU (for ridiculous prices), and at 14nm you should be able to fit a much larger and better performing GPU (Tonga-class?). That's a level of performance that Intel will probably take many years to reach, given how conservative they are with the iGPU area proportions.

You're still neglecting laptops and thermal constraints.

And an eventual x86 APU that uses HBM exclusively, allowing for a much higher performance iGPU within a much smaller PCB.
And the fact that he's completely making up the supposedly enormous price difference between GDDR5 and HBM, that's supposedly so big that no one should be using HBM at all. But it's been going on and on like this for the entire thread, so mentioning the subject is like beating a dead horse.

 

[Alexko]

Why not? 150mm^2 in 28nm gets you a GM107, which is ~3x faster than the fastest that Intel has ever put out there as an iGPU (for ridiculous prices), and at 14nm you should be able to fit a much larger and better performing GPU (Tonga-class?). That's a level of performance that Intel will probably take many years to reach, given how conservative they are with the iGPU area proportions.

Yes, 150mm² is a good target for low-end GPUs. But I'm not sure going significantly lower than that would be worth it when today's (and especially tomorrow's) APUs tend to be larger than this and mostly made up of graphics silicon.

 

[eastmen]

Check my earlier posts...

It doesn't make sense to favor an expensive technology for size when the same performance can be reached much cheaper with GDDR5. It makes much more sense to gradually make HBM migrate down the stack with process improvements:
28nm: HBM doesn't make much sense, except as band-aid for an inefficient core architecture (mostly power, the perf benefit remains an open question.)
16/14nm: HBM2 makes sense for the highest-end silicon.
10nm: HBM2 makes sense for the highest end and the level below it.

The biggest counter-argument is indeed size, if that suddenly becomes a bigger deal than cost. IMO, the FuryX doesn't apply here because it still requires an external fan, and the Fury shows that this argument doesn't work for air cooling. Furthermore, the Fury Nano will likely have terrible perf/mm2. Let's also not forget that miniature size GTX 970 cards have existed as well that are similarly sized as a FuryX. I don't think physical size has been a major point of engineering attention up to now, and that significant improvement is possible with non-HBM solutions if they put their mind to it.

well the board itself would cost less and you wouldn't need all the traces to the ram if its contained in the chip. Your also going to reduce power consumption which can be a very big deal for smaller systems.

The real question is the cost of HBM vs the cost of GDR 5 / Larger board / traces

 

[silent_guy]

well the board itself would cost less and you wouldn't need all the traces to the ram if its contained in the chip.

There is no direct cost to the number of traces (you don't pay more or less for the amount of metal that gets stripped away), but it can determine the number of layers that are required. 8 layers is probably the lowest number you can reasonably expect on a small GPU, irrespective of the memory technology. I can't imagine they need more than 12. (Chip IOs are placed with PCB layout in mind.) Given the large size compared to, say, cell phone, there's not going to be a need for ultra small features or hidden vias etc.
In mass volume, the cost difference between 8 and 12 layers is not very high and the base cost is already on the order of only tens of cents per square inch. Asus has a GTX 970 mini that's about the same size as a FuryX.
So don't worry about the difference in PCB cost: it's going to be small for the large GPUs and even less for the small ones.

Your also going to reduce power consumption which can be a very big deal for smaller systems.

If you were asked to rank the reasons why people buy a small discrete GPU board, which one would come first: cost, performance, or power consumption? Power consumption is important, but it's not the most important driver.

The real question is the cost of HBM vs the cost of GDR 5 / Larger board / traces

Yes.

 

[silent_guy]

And an eventual x86 APU that uses HBM exclusively, allowing for a much higher performance iGPU within a much smaller PCB.

I remain a skeptic on that one. It's not as if Intel has issues with PCB size with their latest offerings.

And the fact that he's completely making up the supposedly enormous price difference between GDDR5 and HBM, that's supposedly so big that no one should be using HBM at all.

It makes sense to use it when it provides performance benefits or when it's the method of last resort to make a competitive product.
In the case of Fiji, AMD must have assumed that they would blow away Kepler but were blindsided by Maxwell. Which only left the last resort part. Still better than not having anything at all...

 

[fellix]

n the case of Fiji, AMD must have assumed that they would blow away Kepler but were blindsided by Maxwell. Which only left the last resort part. Still better than not having anything at all...

Kepler would have been too easy target. It's the pedestrian GM200 SKU is what took them off guard. Titan X was a halo product, more like Quadro for the masses. But the real issue with AMD is the architecture development stagnation. Relying only on raw power by scaling up isn't enough.

 

[UniversalTruth]

The real issue is your negativisim. It is very clearly stated in the good article that Arctic Islands will come with a new architecture.

Would you be please so kind to stop offtopic in this thread and turn to the proper one ?

 

[silent_guy]

The real issue is your negativisim.

What negativism? I'm only giving rational arguments why HBM isn't a slam dunk for upcoming smaller GPUs.

It is very clearly stated in the good article that Arctic Islands will come with a new architecture.

I've stated many times that I expect AMD to come up with architectural improvement for their next generation: they must have done *something* in the last 3 years. If all goes well, this new architecture will negate the requirement to use HBM, just the way it does for Nvidia.

 

[UniversalTruth]

What negativism? I'm only giving rational arguments why HBM isn't a slam dunk for upcoming smaller GPUs.

I was speaking to fellix about his expectation that AMD aren't busy with anything regarding new GCN architecture.

Yes, we understood what you are giving - the thing is that articles appear which makes me think that AMD will do otherwise, regardless of your arguments.

 

[Alexko]

It makes sense to use it when it provides performance benefits or when it's the method of last resort to make a competitive product.
In the case of Fiji, AMD must have assumed that they would blow away Kepler but were blindsided by Maxwell. Which only left the last resort part. Still better than not having anything at all...

In power-constrained environments (i.e., everywhere, now) everything that saves power increases performance.

 

[fellix]

In power-constrained environments (i.e., everywhere, now) everything that saves power increases performance.

Well, not in the case of Fury, but that's the price for the early adoption pain. Hopefully, HBM2 will provide much higher capacities to enable APUs with only two stacks for full system memory replacement in a compact form factor.

 

[Xmas]

If all goes well, this new architecture will negate the requirement to use HBM, just the way it does for Nvidia.

If all goes well, the new architecture will go beyond what is realistically achievable with GDDR5 at the high end.

 

[Alexko]

Well, not in the case of Fury, but that's the price for the early adoption pain. Hopefully, HBM2 will provide much higher capacities to enable APUs with only two stacks for full system memory replacement in a compact form factor.

It sure holds for Fury too. Without HBM, it would draw much more power, and AMD would have had to clock it lower. In fact, it probably wouldn't have developed the GPU at all, because at lower clocks it would barely surpass Hawaii—which is probably exactly what the Fury Nano will be, albeit with HBM of course.

 

[fellix]

The goal of HBM in Fury is definitely to allow the GPU die to "borrow" the TDP savings from the memory, with the net result of shifted balance and more power to the core to consume the extra "low-energy" bandwidth.