What if AMD partnered with ARM to cross-license AMD's GPU tech?

[Idontcare]

(I thought I'd put this thread here in CPU instead of in VC&G considering we talk more about these hypothetical AMD/ARM possibilities here in CPU)

Inspired by this ARM announcement, what if AMD licensed their GPU tech in an alliance with ARM?

ARM Holdings plc has announced the next core in its Mali line of graphics processing units, which is intended to start appearing in smartphone system-chips in 2013.

The Mali-T658 design supports up to 8 shader cores, compared with the Mali-T604's four shader cores, and ARM has also doubled the number of arithmetic pipelines per shader core from two to four.

Wouldn't that work to AMD's favor in terms of leveraging a key asset they have and others do not? Near immediate cashflow with negligible additional investment on AMD's behalf.

 

[cbn]

I don't pay attention to GPUs very much these days, but I did notice the Mali GPUs are getting pretty strong:

http://www.anandtech.com/show/5077/arms-malit658-gpu-in-2013-up-to-10x-faster-than-mali400

The first T604 based SoCs will be available in the second half of 2012 on 28/32nm silicon. ARM is promising up to 68 GFLOPS of compute from T604 (presumably that's for a 4-core configuration at high clocks).

68 GFLOPs?

Wow, that is right up there in Brazos territory!

The following Mali called the "T658" will be even stronger!

I just wonder where this puts AMD? Are we actually at the point where an AMD Bobcat GPU can be reduced in size with die shrinks and low power process tech to actually make it onto a smartphone SOC? Or are the smallest AMD GPUs still far too large for this kind of service?

 

[Olikan]

^ Gflops don't count in performance alone, yet a bobcat gpu for smartfones would be overkill


you are a bit late IDC...

http://semiaccurate.com/2011/06/22/amd-and-arm-join-forces-at-last/

 

[nenforcer]

ARM Norway is a dedicated GPU production team devoted to the MALI architecture.

They actually seem to be doing more mobile GPU development than Qualcomm is with the Adreno GPU's in San Diego. (which is the former ATI Imageon technology they sold) It looks like Adreno is going to be stuck with DirectX 9 technology for a while.

There is definitely room for a new mobile GPU architecture from AMD with support for DirectX 10 and 11 / OpenGL 4 features.

I doubt ARM is going to want to abandon their investment in Mali though, especially since AMD has shown nothing competitive in these ultra low (milli) wattage spaces.

That article even says they consider Imagination and Vivante (MIPS GPU) the competition, not even nVidia Tegra.

 

[cbn]

a bobcat gpu for smartfones would be overkill

The Bobcat GPU is 80 stream processors.

Can AMD break that down any smaller? I'm asking even though I am pretty sure I've been told 40 stream processors is not possible for the 5000 series?

How about the upcoming 28nm AMD's 7000 series GPU? Can that be broken down into a smaller GPU? Or does that one also require the smallest grouping (for lack of a better word) to be fairly large?

 

[Schmide]

They could get the next xbox contract?

 

[quest55720]

They could get the next xbox contract?

I would bet the mortgage on ADM getting the next xbox GPU part. After the orginal xbox I don't think MS will ever use NVIDIA again. I think IBM is almost a lock to be the CPU of choice.

 

[wlee15]

I don't pay attention to GPUs very much these days, but I did notice the Mali GPUs are getting pretty strong:

http://www.anandtech.com/show/5077/arms-malit658-gpu-in-2013-up-to-10x-faster-than-mali400



68 GFLOPs?

Wow, that is right up there in Brazos territory!

The following Mali called the "T658" will be even stronger!

I just wonder where this puts AMD? Are we actually at the point where an AMD Bobcat GPU can be reduced in size with die shrinks and low power process tech to actually make it onto a smartphone SOC? Or are the smallest AMD GPUs still far too large for this kind of service?

Ah but while were on the topic of basically meaningless specs I will note that Bobcat is capable of 280 to 500 million triangles, while the T604 is only be capable of 133 million.

Getting back on topic I think that AMD should develop (on their own) a low-power optimized GPU and license it to third parties. This has the benefit of allowing AMD to profit the growth of ARM while producing a GPU that allows AMD to scale down Bobcat derived SOC.

 

[Gikaseixas]

I think that AMD should develop (on their own) a low-power optimized GPU and license it to third parties.

This

The mobile segment is very lucrative and AMD needs to get into it asap, they have the tech and ability to do it. Intel is moving in that field also.

 

[opethfan]

This

The mobile segment is very lucrative and AMD needs to get into it asap, they have the tech and ability to do it. Intel is moving in that field also.

Are they allowed to? One can only assume the sale to Qualcomm included some non-compete clauses.

 

[Idontcare]

Getting back on topic I think that AMD should develop (on their own) a low-power optimized GPU and license it to third parties. This has the benefit of allowing AMD to profit the growth of ARM while producing a GPU that allows AMD to scale down Bobcat derived SOC.

That's what I am thinking too. Part of that licensing model could include licensing to ARM to enable all the easier cross-licensing to ARM's large customer base.

It just seems like an easy opportunity for AMD versus them finding the capital to do it all themselves in the design space.

 

[ViRGE]

As things currently stand I really don't think ARM needs AMD. All of AMD's desktop/laptop GPU experience isn't worth much in the mobile space until they can prove otherwise.

 

[Dribble]

There's no need for AMD to partner with ARM, if AMD are so good they should just make a gpu component and sell it like everyone else does. Why half the already minimal profit you get by working with ARM. If they're not good enough to do it themselves then why would ARM want to partner with them?

ARM on the other hand should be on the lookout to pick up a few good ex-AMD/ATI gpu devs. I hear there's a few looking for work right now.

 

[BD231]

You honestly expect ARM to have AMD's best interests in mind? That's laughable.

 

[cbn]

That's what I am thinking too. Part of that licensing model could include licensing to ARM to enable all the easier cross-licensing to ARM's large customer base.

It just seems like an easy opportunity for AMD versus them finding the capital to do it all themselves in the design space.

I can see AMD working with ARM for CPUs (in some fashion), but I just hope another company (besides ARM) has better plans for the Radeon GPU IP.

Maybe a company with a focus (or future plans) for super computing? (This to me seems to be the highest use for all this big machine GPU hardware the company makes).

 

[cbn]

It certainly seems ARM is trying to create a market for large Smartphone GPUs:

I just wonder what kind of die size and power consumption for that eight core T658? Surely AMD could field something equivalent to that?

 

[cbn]

Ah but while were on the topic of basically meaningless specs I will note that Bobcat is capable of 280 to 500 million triangles, while the T604 is only be capable of 133 million.

I wonder why the Mali is so strong on Compute, but weak on triangles?

Maybe it has to do with some ARM initiative towards heterogenous computing? ("big.LITTLE" for the cpu cores and a high compute spec for the Mali GPU seem to point that direction.)

 

[ViRGE]

I wonder why the Mali is so strong on Compute, but weak on triangles?

Maybe it has to do with some ARM initiative towards heterogenous computing? ("big.LITTLE" for the cpu cores and a high compute spec for the Mali GPU seem to point that direction.)

Nah. It's because you don't need a lot of triangle throughput on smartphones due to the lower resolutions and the tighter pixel densities. All of the mobile GPUs are weak in that regard. On the PC side it's really only with DX11 that change came about.

 

[cbn]

Nah. It's because you don't need a lot of triangle throughput on smartphones due to the lower resolutions and the tighter pixel densities. All of the mobile GPUs are weak in that regard. On the PC side it's really only with DX11 that change came about.

I see your point about the need for less triangles @ lower resolutions.

When I looked up the following resolutions I calculated a significant difference in total pixel amounts.

1. iphone 4s resolution = 960 x 640 = 614,400 pixels
2. ipad 2 resolution= 1024 x 768= 768,432 pixels
3. 1080p resolution=1920x1080= 2,073,600 pixels

A 1080p screen has approximately ~3 1/3 more pixels than the iphone.

Maybe AMD can make their lowest end GPU work if they aim it at future smartphones that use 1080p docking stations?

 

[cbn]

In this diagram, there are three arrows pointing at the 80 SP GPU (built on 40nm)

I'm estimating that amounts to anywhere from 15 to 20 mm2 die space.

With a die shrink to Low power 22nm or Low power 28nmm, wouldn't this be considered somewhat reasonable for a large smartphone SOC built with Cortex A15 or ARMv8 cores?

Or is the situation not really so much about die size as it is power conserving features?

 

[cbn]

Regarding AMD's Graphic vs Compute balance:

It looks like AMD is definitely changing this with the new GCN "Graphics core Next" GPU:

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute/1

AMD has stretched their traditional VLIW architecture as far as they reasonably can for computing purposes, and as more developers get on board for GPU computing a clean break is needed in order to build a better performing GPU to meet their needs. This is in essence AMD’s Fermi: a new architecture and a radical overhaul to make a GPU that is as monstrous at computing as it is at graphics.

This is great news!

 

[ViRGE]

In this diagram, there are three arrows pointing at the 80 SP GPU (built on 40nm)

I'm estimating that amounts to anywhere from 15 to 20 mm2 die space.

With a die shrink to Low power 22nm or Low power 28nmm, wouldn't this be considered somewhat reasonable for a large smartphone SOC built with Cortex A15 or ARMv8 cores?

Or is the situation not really so much about die size as it is power conserving features?

It used to be that SoC manufacturers were very conservative on die space. Apple changed that, particularly with the (relatively) massive A4. Now SoC makers for high-end devices are willing to accept a larger die. The trade-off as you note is that power consumption still needs to stay low. So a large GPU isn't necessarily a problem depending on its power consumption.

 

[cbn]

As a way of illustrating how impressive the 272 Gflops is for the eight core T658 and the 68 Glops is for the T604 I thought I would find some comparison data from current GPUs:

1. http://www.anandtech.com/show/5072/nvidias-tegra-3-launched-architecture-revealed/2

^^^^ In the above Anandtech link, we can see the current Mobile GPUs span 3.2 Gflops to 19.2 Gflops. (Tegra 2 coming in a 4.8 to 7.2 Gflops)

2. The AMD supplied spec for a 151 watt TDP HD6870 is 2 Tflops (ie, 2000 Gflops).

 

[SickBeast]

Strapping one of AMD's GPUs to an ARM CPU is like putting a Ford Fiesta engine into a Corvette. It just doesn't make sense. You would be CPU limited in all games at all resolutions.

That said, I think AMD could make an interesting computer with all of their know how. Perhaps they could become the new Apple.

If AMD creates a "Fusion" core based on an ARM CPU, what is its use going to be? nVidia at least has a GPGPU platform that is useful to many people. If AMD does it, it's going to be dead on arrival with no software and no use for 99.9% of people.

 

[cbn]

Strapping one of AMD's GPUs to an ARM CPU is like putting a Ford Fiesta engine into a Corvette. It just doesn't make sense. You would be CPU limited in all games at all resolutions.

It would be great to see an ARM CPU connected to a little AMD GPU for dockable smartphone.

X86 on the high power end (competing against Nvidia Denver HPC or an Intel CPU + separate Nvidia Tesla GPU)

ARM + AMD GPU on the low end (competing against Nvidia SOC).

EDIT:

Speaking of x86 on the high power end:

One thing that concerns me about the Bulldozer vs ARM war:

If the ARM partners find a way to justify greater R&D expenditures, we may begin to see more Custom ARM cores with hand optimizations.

How could this happen?

The one big advantage I see with ARM vs Bulldozer is that ARM partners are nearing the point where a custom ARM core could be used across both Server SOC platforms and high end Smartphones SOC platforms. That is a lot of chip volume and could potentially lead to justification for greater CPU R&D expenditures.

Potential ARM Server SOC Volume + Smartphone SOC volume vs Bulldozer's lower volume, but higher profit margin server sales.

Who will win the tug of war for R&D money? Or will Fusion and AMD's future GPGPU developer tools be able to re-center this and result in a better spec for the Bulldozer CPU core? Maybe even allowing Bulldozer and "Custom ARM CPU core" to exist together at different server market segments? If so, I think it would be a big mistake for AMD to license their GPU portfolio to ARM.