Why can't Intel compete with AMD in GPU Sector?

[witeken]

We are talking about the design of the iGPU not the density of the process.

Performance per mm² is directly related to process used.

 

[witeken]

Iris Pro(HD5200) iGPU is close to 100-110mm2, add 85mm2 for the eDRAM and we are close to 200mm2 vs 110-120mm2 for Kaveri.

Intel Iris (HD5000) without eDRAM is considerable slower than Kaveri at the same die size with a full node advantage. Kaveris iGPU at 22nm would be smaller than Intels HD5000 and perform faster.

You could as well have added the other caches to your die size calculation...

 

[SAAA]

You could as well have added the other caches to your die size calculation...

Well he got the point: without eDRAM GT3 GPU isn't all that good so it's essential to include that area in the computation.

Oh I forgot before but not only the problem seems to be in perf./mm^2 but particularly in transitors/mm^2 of these IGP parts: just compare Haswell dual core+GT3 vs quad core+GT2 dies, they are about the same density, and while it also contains strong computation/logic parts overall perf/mm^2 is really low compared to Nvidia or AMD GPUs.

At any point in time and at any node GPUs have been much denser than CPUs by a large factor, close to a full node shrink, but that isn't happening for Intel iGPUs for some reason.

 

[AtenRa]

Performance per mm² is directly related to process used.

Again, we are not talking about process density.

You could as well have added the other caches to your die size calculation...

What other caches ??

HD5000 is the same iGPU as Iris Pro HD5200 but without the eDRAM. HD5000 is slower than both Iris Pro and Kaveri.

 

[witeken]

Again, we are not talking about process density.

You don't decide what's being talked about. Someone made a comment about Intel's IGP's performance per mm². I replied.

What other caches ??

HD5000 is the same iGPU as Iris Pro HD5200 but without the eDRAM. HD5000 is slower than both Iris Pro and Kaveri.

You cherry picked Intel's L4 cache, which has only indirectly to do with the performance. You could as well have added L1, L2 and L3, or while you're at it, the whole die.

 

[AtenRa]

You don't decide what's being talked about. Someone made a comment about Intel's IGP's performance per mm². I replied.

Take it easy and pay attention, Im not the one making the argument,

The most dreadful thing about Intel IGP is performance/mm^2 right now. Looking at the competitors it seems they are 5-10 years behind in this and they even have a decent process to hide the problem!
Imagine the same architecture on 28 or 32nm process...

Performance per mm² isn't stellar because Intel didn't focus on density until 14nm.

He was talking about the iGPU density(architecture, design of the IC etc) in relationship to other iGPUs, you replied about node process density. Not the same.

You cherry picked Intel's L4 cache, which has only indirectly to do with the performance. You could as well have added L1, L2 and L3, or while you're at it, the whole die.

This is not cherry picking, L4 cache gives the performance Iris Pro has. Without it GT3 iGPU is slower, simple as that. So when people are talking about Iris Pro performance they mean GT3e dies, the one with ~85mm2 L4 eDRAM.

 

[Erenhardt]

You could as well have added the other caches to your die size calculation...

If you are willing to skip memory enhancements that boost the igp performance, why not apply the same standards for both?

If you insist on keeping eDRAM iris pro results, compare it against kaveri without memory bottleneck, which is the same as HD7750. Here is HD6750 and it crushes everything by a huge margin:

I won't even metion HD7750 is about 50% faster than HD6750 ...
That is what kaveri is capable of with bandwidth bottleneck gone - HBM can't come soon enough.

It shows how far behind intel is. It also shows how much bandwidth starved kaveri is.

 

[witeken]

It shows how far behind intel is.

This is how it always goes. It's like the comparison GCN vs Maxwell; Nvidia's GPU architecture is 1 iteration ahead. There's always some bias: when it launched, Iris Pro was pretty decent, and just after Broadwell-K's launch, it will also look good... until AMD comes with a new generation to leapfrog Intel etc.

 

[Enigmoid]

People are comparing HD 5000 in 28W parts against desktop parts? Really?

 

[Erenhardt]

This is how it always goes. It's like the comparison GCN vs Maxwell; Nvidia's GPU architecture is 1 iteration ahead. There's always some bias: when it launched, Iris Pro was pretty decent, and just after Broadwell-K's launch, it will also look good... until AMD comes with a new generation to leapfrog Intel etc.

GCN was released in 2012. Intel cant match GCN with their latest GPU arch outside of bottlenecked scenarions.

The only reason why kaveri and iris pro have similar performance is because one has a lot less memory bandwidth.

 

[Exophase]

Iris Pro is a good contender among IGPs but that doesn't necessarily mean it's trivial to scale it up to high-end dGPU class and still be as competitive. Crystalwell, for example, gives them an advantage in the IGP space that doesn't translate to a similar advantage for high-end dGPUs. At least it wouldn't as it exists today.

Intel has the manufacturing advantage and a lot of resources, but that's not everything. AMD and nVidia have a lot more experience with high end designs and more experience with GPU design in general, which counts for a lot. For such an entrenched and shrinking market space to really be attractive to Intel they'd have to have a solution that is far better than the competition, not merely competitive. That's a pretty tall order.

 

[Phynaz]

If you insist on keeping eDRAM iris pro results, compare it against kaveri without memory bottleneck, which is the same as HD7750.

Compare it to a product that doesn't exist except in your mind...

 

[Erenhardt]

Compare it to a product that doesn't exist except in your mind...

Last time I checked HD7750 was a real product. (It has 20% shaders disabled)

 

[Phynaz]

Last time I checked HD7750 was a real product. (It has 20% shaders disabled)

Kaveri without memory bandwidth limits doesn't exist except in your mind.

It a sign of desperation when you have to make up hypothetical products to make your case against a real shipping product.

 

[Erenhardt]

Kaveri without memory bandwidth limits doesn't exist except in your mind.

It a sign of desperation when you have to make up hypothetical products to make your case against a real shipping product.

I advise you to get back to topic: "Why can't Intel compete with AMD in GPU Sector?"

We are not discussing SoC/APU performance here.
We discuss graphics architectures here.

Comparing intel's igp perfromance/die space to amd GCN performance/die space shouldn't be hard to grasp concept.

 

[Phynaz]

You're the one who brought up Kaveri, not me.

 

[Erenhardt]

You're the one who brought up Kaveri, not me.

No. Some here were comparing performance of intel igp to kaveri with conclusion that intel graphics architecture is as good (or even better) than gcn since kaveri have more space dedicated to gpu than iris pro.

Its not unusual for people here to show amd in worst possible scenario. Here they put gcn in the most bottlenecked flavour available and dance their victory dance.

If they want to compare intel graphics to amd graphics, at least be square.

Comparing ~85mm2 iris pro to 90mm2 gcn part to draw perf/mm2 seems reasonable.

 

[mrmt]

No. Some here were comparing performance of intel igp to kaveri with conclusion that intel graphics architecture is as good (or even better) than gcn since kaveri have more space dedicated to gpu than iris pro.

Still a real product against an unreal product.

 

[Enigmoid]

No. Some here were comparing performance of intel igp to kaveri with conclusion that intel graphics architecture is as good (or even better) than gcn since kaveri have more space dedicated to gpu than iris pro.

Its not unusual for people here to show amd in worst possible scenario. Here they put gcn in the most bottlenecked flavour available and dance their victory dance.

If they want to compare intel graphics to amd graphics, at least be square.

Comparing ~85mm2 iris pro to 90mm2 gcn part to draw perf/mm2 seems reasonable.

To be fair, I thought an on die/APU comparison would be best. That is what is being sold after all and implementation frequently is more important than technical.

To really make a comparison the performance area of the gcn must be compared to the performance area of Gen 7.5 graphics. With process scaling and driver handicapps taken into account.


In an implementation form AMD or Intel can do stupid things. Like putting 512 shaders when you only have the bandwidth to feed 384.

Ie) http://www.anandtech.com/show/8067/amd-am1-kabini-part-2-athlon-53505150-and-sempron-38502650-tested

Given that one core is 3.1 mm2, extrapolating out gives the size of the die at 31.4x the size of a single core, or 97.3 mm2. The GPU area is approximately 5.2x the size of a core, giving ~16.1 mm2 for 128 GCN cores, compared to 12.4 mm2 for CPU cores. The Video Codec Engine and Unified Video Decoder are not part of these totals, located on other parts of the APU. The memory controller clocks in at ~9.4 mm2 and the display/IO portion runs at ~7.3 mm2.

16.1 mm^2 for 128 sp -> 64.4 mm^2 for 512 sp.

Something similar must be done for intel.

 

[Erenhardt]

To be fair, I thought an on die/APU comparison would be best. That is what is being sold after all and implementation frequently is more important than technical.

To really make a comparison the performance area of the gcn must be compared to the performance area of Gen 7.5 graphics. With process scaling and driver handicapps taken into account.


In an implementation form AMD or Intel can do stupid things. Like putting 512 shaders when you only have the bandwidth to feed 384.

Ie) http://www.anandtech.com/show/8067/amd-am1-kabini-part-2-athlon-53505150-and-sempron-38502650-tested


16.1 mm^2 for 128 sp -> 64.4 mm^2 for 512 sp.

Something similar must be done for intel.

Yes. Because of that a8-7600 have much better performance/mm2 (I know its the same die - count working area) than 7850k despite both sporting the same GCN cores.
Comparing intel performance boosted by eDRAM to kaveri bottlenecked performance is quite misleading. If someone would really try hard to make gcn look terrible - single channel benchmark would do the trick.

iris pro wouldn't loose much performance thanks to additional bandwidth from eDRAM. Kaveri would loose 50% of its performance because of single channel memory.

This leads to wrong conclusions about the architecture's competitiveness.

 

[CoPhotoGuy]

Intel could...they would just have to do what AMD did and do something like buy Nvidia. But they don't care to do that.