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

[Kaotik]

Don't be deceived: HBM helps decrease memory interface power, but that's a small fraction of overall GPU power. It will shave 30 W from Hawaii's 300 W budget, but 30 W of extra shaders is not sufficient.

To quote Tech Report, go few years back and the memories (including memory controller on GPU) could eat up to 70W, that's nearly 1/3 of total board power for most cards.
If/when HBM power savings translate to GPU's memory controller too, it can be a major advantage.

 

[iMacmatician]

While Tonga hasn't improved appreciably compared to Tahiti, I do hope better results for Grenada( if it's based on Tonga and beyond) vs. Hawaii.

I was under the impression that Grenada is essentially Hawaii, like Curacao is to Pitcairn.

Was GDDR4 worth having?

What were the advantages of GDDR4 over GDDR3? The GTX 285’s GDDR3 clock was higher than the GDDR4 clock of any part that AMD had, as far as I know, but I don’t know about power consumption or other factors.

 

[CarstenS]

To quote Tech Report, go few years back and the memories (including memory controller on GPU) could eat up to 70W, that's nearly 1/3 of total board power for most cards.
If/when HBM power savings translate to GPU's memory controller too, it can be a major advantage.

Your quote's gone, can't see it.

30 watts is the number AMD cites (10,66 watts/GB/s, 320 GB/s baseline) for a fully loaded memory subsystem on Hawaii for the memory chips alone. And only in regard to this were numbers of perf/watt improvements given. So, whether or not a 4.096 Bit memory interface at 500 MHz (or slightly more) is consuming more power than 512 Bits of 1.250 MHz is the question.

 

[Esrever]

GDDR4 wasn't worth the cost since its improvement over GDDR3 wasn't significant if I remember correctly.

 

[AlNom]

Your quote's gone, can't see it.

30 watts is the number AMD cites (10,66 watts/GB/s, 320 GB/s baseline) for a fully loaded memory subsystem on Hawaii for the memory chips alone. And only in regard to this were numbers of perf/watt improvements given. So, whether or not a 4.096 Bit memory interface at 500 MHz (or slightly more) is consuming more power than 512 Bits of 1.250 MHz is the question.

Shouldn't the comparison be made against similar bandwidth? How much power would they have to burn to achieve >500GB/s with GDDR5?

 

[CarstenS]

What AMD sorely needs is geometry performance, at least to counter the geometry viruses we've been finding in Gameworks titles.

AMD needs more graceful handling of fat control points which are suffocating the rest of their pipeline. Pure geometry performance is fine.

 

[silent_guy]

Shouldn't the comparison be made against similar bandwidth? How much power would they have to burn to achieve >500GB/s with GDDR5?

It depends what you want to compare. BW/W is important when performance is highly correlated with BW. Absolute power is more important when it isn't.

We'll have to see how Fiji performance turns out, but if it's similar to a Titan X which has less than 50% of the Fiji BW, then it's obvious that Fiji is ridiculously overbuilt in terms of BW with no way of making good use of it. In that case, it doesn't matter whether the BW is 500GBps or 1000GBps. What matters then is that it's useless from a performance point of view, but at least you get some absolute power reduction.

I'm incredibly excited that, just like Maxwell, Fiji is on 28nm as well, since it will allow us once again to make better apples for apples comparisons. I'll be even happier if there'll be a Fiji Light SKU with only 2 or 3 instead of 4 HBM stacks to see how BW influences performance. But if the memory size limitations turn out to be true, that's probably a stretch.

 

[silent_guy]

AMD needs more graceful handling of fat control points which are suffocating the rest of their pipeline. Pure geometry performance is fine.

What do you mean by fat control points?

 

[silent_guy]

Where did you get 30W?

From TechReport's article on HBM:

Macri did say that GDDR5 consumes roughly one watt per 10 GB/s of bandwidth. That would work out to about 32W on a Radeon R9 290X. If HBM delivers on AMD's claims of more than 35 GB/s per watt, then Fiji's 512 GB/s subsystem ought to consume under 15W at peak. A rough savings of 15-17W in memory power is a fine thing, I suppose, but it's still only about five percent of a high-end graphics cards's total power budget. Then again, the power-efficiency numbers Macri provided only include the power used by the DRAMs themselves. The power savings on the GPU from the simpler PHYs and such may be considerable.

If it's 17W on one side, then 30W on both sides is probably a pretty decent estimate.

 

[silent_guy]

GCN is behind maxwell, but not that far behind. Per shader efficiency of Maxwell should be about, figures takes from hardware.fr,

[(105/96)*(2816/2048)] / [(1240/1040)] = 1.26 times of GCN.

Not sure which hardware.fr article you're referencing, so could you explain the 105/96 factor? Is that a benchmark score?

You're adjusting for clocks. That's fine to get out IPC, but it filters out the fact that the Maxwell SM is able to run at much higher speeds yet at the same time consume much less power, which is a major factor when looking at efficiency.

 

[Alexko]

From TechReport's article on HBM:
If it's 17W on one side, then 30W on both sides is probably a pretty decent estimate.

OK. With 60% more bandwidth, that's pretty decent.

 

[Infinisearch]

With regards to HBM, does anyone know what the burst transfer length is? All I know is the width per stack is 1024bits.

 

[RecessionCone]

OK. With 60% more bandwidth, that's pretty decent.

Right. But not nearly sufficient to give Fiji the shader power it needs to compete against Titan X. To do that, AMD needs much more power efficient shaders, and/or a vastly increased thermal budget.

So the big story about Fiji will revolve around AMD's treatment of power issues, not HBM.

 

[silent_guy]

If the total power is less than 400W, a water cooler shouldn't have too much trouble with that.

At that point they'll be 2 stories: on one hand, whether or not it's a good GPU for consumers. That mostly depends on price for performance.
And the other one will be about technical aspects, whether or not it's an elegant technical solution.

If the current leaks are true, it will probably outperform a Titan X with a bit of margin for a lower price, which will make AMD fans and those who want the fastest GPU at all cost happy.

And it will probably do that , using tons of power and an insane amount of BW., a brute force solution lacking architectural finesse, much like GTX 480. (And if it doesn't outperform: R600?)

Either way, it will provide plenty of opportunity for discussions.

 

[gamervivek]

Not sure which hardware.fr article you're referencing, so could you explain the 105/96 factor? Is that a benchmark score?

You're adjusting for clocks. That's fine to get out IPC, but it filters out the fact that the Maxwell SM is able to run at much higher speeds yet at the same time consume much less power, which is a major factor when looking at efficiency.

Titan X review, they have clocks in question. And I agree the biggest problem AMD have right now is that they don't have the clockspeeds. GCN is fine atm.

 

[AlNom]

It depends what you want to compare. BW/W is important when performance is highly correlated with BW. Absolute power is more important when it isn't.

We'll have to see how Fiji performance turns out, but if it's similar to a Titan X which has less than 50% of the Fiji BW, then it's obvious that Fiji is ridiculously overbuilt in terms of BW with no way of making good use of it. In that case, it doesn't matter whether the BW is 500GBps or 1000GBps. What matters then is that it's useless from a performance point of view, but at least you get some absolute power reduction.

mm... fair enough. I suppose we're all waiting to see if this massive bandwidth is going to make a difference...

 

[gamervivek]

Samples are still going around for ~$1250

 

[RecessionCone]

If the total power is less than 400W, a water cooler shouldn't have too much trouble with that.

At that point they'll be 2 stories: on one hand, whether or not it's a good GPU for consumers. That mostly depends on price for performance.
And the other one will be about technical aspects, whether or not it's an elegant technical solution.

If the current leaks are true, it will probably outperform a Titan X with a bit of margin for a lower price, which will make AMD fans and those who want the fastest GPU at all cost happy.

And it will probably do that , using tons of power and an insane amount of BW., a brute force solution lacking architectural finesse, much like GTX 480. (And if it doesn't outperform: R600?)

Either way, it will provide plenty of opportunity for discussions.

TPU has Titan X performing at 43% faster than 290X. So if we scale up Hawaii's shader array to match Titan X performance, subtracting 30 W for memory interface savings, we get (1.43 * 300) - 30 = 400 W.

However, 290X is not the most power efficient Hawaii SKU - it gets 18.67 GFlops/W, while 295X2 gets 22.6 GFlops/W. So let's assume Fiji gets the same power efficiency as the best Hawaii GPUs. Then Fiji burns (1.43 * 300 * 18.67/22.6) - 30 = 325 W.

So, it seems that for Fiji to beat Titan X, it will need somewhere between 325-400 W (assuming AMD doesn't have other performance/watt improvements in their shader array).

I'm guessing that Fiji will be a 350 W part, and more or less match Titan X performance. I also think Nvidia will allow its partners to release various overclocked GM200 based parts that are somewhere between 15-40% faster than Titan X (at correspondingly increased power budgets).

 

[silent_guy]

Running a Titan X 40% faster seems to be a bit of a stretch.

 

[Grall]

We'll have to see how Fiji performance turns out, but if it's similar to a Titan X which has less than 50% of the Fiji BW, then it's obvious that Fiji is ridiculously overbuilt in terms of BW with no way of making good use of it.

Alpha blending/particles eats b/w like *analogy of your choice* at high resolutions; framerates on fat GPUs sag even in older games due to this. If fiji has a cure for it, then I'm all for it.