Trinity vs Ivy Bridge

[Mintmaster]

Llanos iGPU allocation is much bigger (28-30% vs 40% of the overall chip). Even without the shrink the iGPU part would be far smaller than Llanos.

Did you see the link fellix posted? The SB die on the left is 216 mm². If you scale IVB up to match the CPU core size (i.e. if we had a 32nm IVB), it would look like the middle picture, and be substantially bigger than Llano.

On 32nm, IVB's iGPU would be roughly the same size as Llano's.

 

[nAo]

On 32nm, IVB's iGPU would be roughly the same size as Llano's.

What fraction of the total area of the die are you using for IVB and LLano GPUs? What area multiplier to compare "iso"process?

 

[Paran]

Did you see the link fellix posted? The SB die on the left is 216 mm². If you scale IVB up to match the CPU core size (i.e. if we had a 32nm IVB), it would look like the middle picture, and be substantially bigger than Llano.

On 32nm, IVB's iGPU would be roughly the same size as Llano's.

Intels CPU parts is much bigger, it means you cannot compare the overall size. You have to isolate the GPU part of the die. And based from the IVB wafer the die size is 160 mm² big in size and the GPU part 45-50 mm² (minus the free space 45 mm² more likely). I have no idea how chip-architect measured it.

 

[Mintmaster]

The retail 3770K die shown on this website shows 160mm2 though.

Retail dies don't have "INTEL CONFIDENTIAL" written on them...

What fraction of the total area of the die are you using for IVB and LLano GPUs? What area multiplier to compare "iso"process?

Intels CPU parts is much bigger, it means you cannot compare the overall size. You have to isolate the GPU part of the die.

I am. Using fellix's link, it seems that a scaled up IVB would be 280 mm2, and the iGPU is 31%.

We're looking at maybe 87mm2 on a theoretical 32nm IVB vs 91mm2 for Llano's IGP. I guess DX11 cost them quite a bit of space. It's possible that it's not the architecture that's at fault but rather poorer scaling of the GPU to 22nm, but I don't know if that makes sense.

Anyway, Trinity seems to have improved on Llano substantially, if we're to believe the released/leaked numbers.

 

[DavidC]

I am. Using fellix's link, it seems that a scaled up IVB would be 280 mm2, and the iGPU is 31%.

I am not sure if it works exactly like that. To compare at equal process to Llano we first need to see how the GPU in there compare with 40nm discrete GPU equivalents. But I can guess.

If this is the real die shot of Tegra 2, the 1MB L2 there would only take 3.8mm2: http://img121.imageshack.us/img121/559/tegra2.jpg

Bulldozer's L3 cache on 32nm takes 3.85mm2/MB, and Sandy Bridge takes 3.1mm2/MB.

Looking at pure density, the order from least dense to greatest dense becomes: AMD>Intel>TSMC. For Intel and TSMC the difference is probably due to optimization targets. Intel uses larger cells but much faster ones, TSMC uses smaller cells but much slower.

The reason I want to see direct comparisons between discrete 40nm and Llano's 32nm is to verify that there's no other factors influencing density. What if experience in designing to such processes affect the outcome? Bobcat's L2 is less dense than Tegra 2's despite having the same process technology at TSMC. Maybe because Nvidia is used to having TSMC made CPUs longer than AMD. What about AMD's experience with GPUs? Would that change anything?

 

[Mintmaster]

Not sure how those observations would put my claim into question. Is there any reason to think that a CPU would scale down substantially better than a GPU?

That's the only way the IVB IGP would be smaller at 32nm than is implied by that picture. The SB and IVB die shots show almost exactly the same features within each CPU core, so it looks very close to being a straight die shrink with some minor tweaks. I can't see any legit case against their use in determining scaling.

 

[Deleted member 13524]

I think these didn't show up in this thread yet:

APU with HD7660G + HD7670M (480sp VLIW5 Turks @ 600MHz) results in +75% performance over single HD7670M:

I don't know what that "GT550" is, that GTS450 seems to be the desktop version (GF106/GF116), which has about the same performance as the modern-day GTX560M for "large" notebooks.
I'd pay good money for that kind of performance on a slim 13.3" notebook (subsequently ditching the optical drive, I don't need that anymore).


And these slides:

 

[Deleted member 13524]

Follow-up: the rest of the slides. (why a 6 images-per-post limit?)

Nonetheless, it seems that Trinity+HD7600M will be able to provide performance/TDP ratios that Intel can't match, whatever IvyBridge+600M series they put up.
Let's hope AMD and their partners take this in mind and launch lots of truly portable gaming notebooks using Trinity, even though I'm definitely seeing a trend on the opposite way.

 

[fellix]

Wow! Piledriver touches almost every aspect of the CPU architecture. I hope it pays off.

 

[Mintmaster]

I assume MM07 is MobileMark '07, so isn't 2.1W really good? It seems Zacate uses 1.8W, so we should get fantastic battery life.

We damn well better see some ~3lb notebooks with Trinity. I'm crossing my fingers for seeing it in a 2lb W8 tablet (though I probably won't buy one).

 

[Blazkowicz]

Tottentranz, that's interesting. but I remember seeing benchmarks with crossfire between an AMD Llano and a discrete GPU, results were totally dysmal.

on a laptop we can have integrated GPU and discrete GPU better matched and drivers can be slightly reworked so that crossfire actually works, but I'm wary of it.
I guess I'd like a lone Trinity, just one big chip to cool off, paired either with a cheap SSD or a hard drive or both.

"IOMMU v2" is tasty though, less than 1% of the population will care about it but it might enable you to run a hypervisor with a linux guest using the integrated GPU, and a windows guest using the dedicated GPU, both at the same time.

 

[Alexko]

Tottentranz, that's interesting. but I remember seeing benchmarks with crossfire between an AMD Llano and a discrete GPU, results were totally dysmal.

I remember that too, but that was at launch time. As anyone reviewed "dual-graphics" since? Perhaps there have been improvements.

After all, AMD appears to be committed to this.

 

[Lightman]

I remember that too, but that was at launch time. As anyone reviewed "dual-graphics" since? Perhaps there have been improvements.

After all, AMD appears to be committed to this.

We had (and still have) only few games able to use hybrid CF because it was limited to DX10+.
Things should get better with time, but I'm more interested in APU alone or APU + GPU with switchable graphics.

 

[mczak]

I remember that too, but that was at launch time. As anyone reviewed "dual-graphics" since? Perhaps there have been improvements.

After all, AMD appears to be committed to this.

That might be true, but still if you need another discrete chip anyway you soon have the problem it might be more attractive to use a slightly more powerful gpu without CF and just forget about the apu. And at that point Trinity probably doesn't look that attractive anymore (though if it manages to be quite a bit faster than Llano at the cpu side for typical usage that should certainly help).
Even if Trinity paired with Turks gets you 75% more performance (probably often quite a bit less) baby Kepler (gk107, at least some versions of it) or mobile CV will still beat it easily (and using CV in a crossfire configuration won't help much again).
So I'm also thinking lone Trinity is probably quite nice, but dual-graphics might not be all that useful (of course for AMD it's probably pretty important as they could hit higher price points not to mention sell both a APU and a GPU). I'll reserve final judgement though until seeing more bechmarks.

 

[Alexko]

That might be true, but still if you need another discrete chip anyway you soon have the problem it might be more attractive to use a slightly more powerful gpu without CF and just forget about the apu. And at that point Trinity probably doesn't look that attractive anymore (though if it manages to be quite a bit faster than Llano at the cpu side for typical usage that should certainly help).
Even if Trinity paired with Turks gets you 75% more performance (probably often quite a bit less) baby Kepler (gk107, at least some versions of it) or mobile CV will still beat it easily (and using CV in a crossfire configuration won't help much again).
So I'm also thinking lone Trinity is probably quite nice, but dual-graphics might not be all that useful (of course for AMD it's probably pretty important as they could hit higher price points not to mention sell both a APU and a GPU). I'll reserve final judgement though until seeing more bechmarks.

GK107 might do better, but at a higher cost.

Plus, this has more to do with AMD not having a small GPU in this generation than anything else, although I suppose the 7750M could be a decent fit for dual graphics. But this does make dual-graphics slightly awkward for this generation.

From what we've seen, Kaveri will bring AMD's APUs from 384 VLIW4 SPs to 512 GCN ones, so that should make dual-graphics a lot more balanced with Kaveri + Cape Verde or its successor.

 

[mczak]

GK107 might do better, but at a higher cost.

Yes this is not quite sure, but if we take that 75% faster than just Trinity alone 3dmark11 score at face value it's just about the level of a GTS450 - top gk107 might not be all that much faster but this benchmark typically shows good SLI/CF scaling.
Cost might indeed be higher but I think there's not that much of an advantage. Granted gk107 might need gddr5 memory to beat that dual graphics solution. But if you compare just Trinity APU vs. say low-end gk107 (with low clocks, ddr3 etc.) the price advantage is likely going to be higher than if you compare APU+Turks vs. high-end gk107.

Plus, this has more to do with AMD not having a small GPU in this generation than anything else, although I suppose the 7750M could be a decent fit for dual graphics. But this does make dual-graphics slightly awkward for this generation.

It might make dual-graphics slightly awkward but if you can overcome the hurdles it's probably not much of a problem performance-wise (not much more than the usual sli/cf trouble). The biggest issue still is you really need a gpu which has about the same performance, so a ddr3-based Turks probably really makes the most sense, go over that and benefits start to diminish.

From what we've seen, Kaveri will bring AMD's APUs from 384 VLIW4 SPs to 512 GCN ones, so that should make dual-graphics a lot more balanced with Kaveri + Cape Verde or its successor.

Yes but even so for best scaling the slowest Cave Verde available today is already quite a bit too fast - well slap ddr3 on it and it might be a good match . Though maybe amd can compensate the bandwidth disadvantage of the Kaveri APU by using shared L3 cache.

 

[Alexko]

It might make dual-graphics slightly awkward but if you can overcome the hurdles it's probably not much of a problem performance-wise (not much more than the usual sli/cf trouble). The biggest issue still is you really need a gpu which has about the same performance, so a ddr3-based Turks probably really makes the most sense, go over that and benefits start to diminish.

What I meant is that Turks isn't ideal because it's still on 40nm, and VLIW5, so not the most power-efficient GPU in the world. If AMD had released a sort of half-Cape Verde, it would be an ideal partner for Trinity.

Yes but even so for best scaling the slowest Cave Verde available today is already quite a bit too fast - well slap ddr3 on it and it might be a good match . Though maybe amd can compensate the bandwidth disadvantage of the Kaveri APU by using shared L3 cache.

For some reason, all mobile Cape Verdes have the same bandwidth, in spite of large variation in arithmetic throughput and most other metrics. I'm really not sure why, maybe GCN GPUs no longer have DDR3 support.

 

[mboeller]

Plus, this has more to do with AMD not having a small GPU in this generation than anything else, although I suppose the 7750M could be a decent fit for dual graphics. But this does make dual-graphics slightly awkward for this generation.

From what we've seen, Kaveri will bring AMD's APUs from 384 VLIW4 SPs to 512 GCN ones, so that should make dual-graphics a lot more balanced with Kaveri + Cape Verde or its successor.

Why do you think that the combination of a 7670m and a Trinity APU is not balanced?

Trinity has 96 VLIW4 Shader + 24 TMUs @ (up to) 685MHz
a 7670m has 96 VLIW5 Shader + 24 TMUs @ 600MHz

so regarding shader count the system seems well balanced. ROPs (?) and especially bandwidth could still be a problem.

 

[Alexko]

Why do you think that the combination of a 7670m and a Trinity APU is not balanced?

Trinity has 96 VLIW4 Shader + 24 TMUs @ (up to) 685MHz
a 7670m has 96 VLIW5 Shader + 24 TMUs @ 600MHz

so regarding shader count the system seems well balanced. ROPs (?) and especially bandwidth could still be a problem.

Trinity + Cape Verde is unbalanced.

Trinity + Turks is very well balanced, but Turks is still on 40nm, so it's not very power efficient, probably much less than GK107, for instance.

 

[Paran]

Retail dies don't have "INTEL CONFIDENTIAL" written on them...
I am. Using fellix's link, it seems that a scaled up IVB would be 280 mm2, and the iGPU is 31%.

We're looking at maybe 87mm2 on a theoretical 32nm IVB vs 91mm2 for Llano's IGP. I guess DX11 cost them quite a bit of space. It's possible that it's not the architecture that's at fault but rather poorer scaling of the GPU to 22nm, but I don't know if that makes sense.

IVB GPU size is 45 mm² (~28% of the overall die) according to official wafer and die pictures. If you add a 28% big GPU to the Sandy Bridge non iGPU space you end up way smaller. Llanos iGPU is much bigger, it reaches around 40% of the overall die of 228 mm².