interesting article on AMD fusion.

[happysmiles]

http://www.tomshardware.com/reviews/fusion-hsa-opencl-history,3262.html

It paints a more positive picture than what it actually was like but it's interesting to see a more behind the scenes look at fusion.

It's long but it gives me hope for AMD's future.

 

[Idontcare]

AMD has pretty much always been to Intel what Linux has been to Microsoft.

Dwarfed in every respect, from economics to manpower to business connections.

Only at least with Linux there is the option of crowd-sourcing and leveraging the inherently lower cost structure that comes with free/volunteer labor. AMD not so lucky there as to convince plucky and smart design engineers to work for free from their own homes.

AMD's management never had the luxury of competing with Intel on Intel's terms, they had to keep doing things different just to have a chance at upstaging the 800lb gorilla.

Fusion was exactly that. It may not have played out as intended or hoped, but had they done nothing then their fate would have been sealed nevertheless. I applaud their courage in attempting to forge their own destiny, to get out from Intel's shadow, but no one ever said it was going to be easy

 

[pelov]

AMD has pretty much always been to Intel what Linux has been to Microsoft.

That's a rather apt analogy if you consider what has to go right for AMD to pull this off. Much like Linux, the crux of whole argument of what Linux lacks is usually centered around applications, but because of their low market share the developers don't really see a need to port over their software to such a low percentage of users. (Bear in mind, I don't think any average user really lacks any applications on Linux but that's another story altogether)

AMD faces the same dilemma. They have a small market share and their success too hinges on the developers and their willingness to adapt a new programming model that targets a small minority of the market... in the desktop. In mobile and HPC the story changes dramatically.

The addition of ARM joining HSA, though, and on the HPC end GPGPU has been a staple as far as compute goes improves the outlook. If you're looking for AMD's HSA to make a stamp on the desktop then you're looking in the wrong segment of the market.

If AMD doesn't pull through then it would be a strange twist of irony to see HSA, their method of clawing back into the market, as being their nail in the coffin, as the ARM manufacturers and Intel beat them to the punch. I have no doubt that GPGPU across all platforms will become absolutely huge, I'm just not sure whether AMD will be there to spearhead that movement. They've been taking far too long and the low end segment of the market is improving at a blistering pace with cutthroat competition that we haven't seen in x86 for years. A couple of back-to-back delays and they'll have postponed themselves out of contention.

 

[Nemesis 1]

I found the article to be more hype than anything else sorta PH1 and BD type hype. xtreme has a topic on same article. I love how AMD tries to go back to 2002 . As the foundation when AMD didn't buy ATI till 06. Intel had already choozen their road in 2004 after they bought Elbrus and choose between 3 test projects what we know as larrabbee was choosen in 2004 . That morphed into nights ferry and finnally into nights corner. AMd is just saying we came up with fusion 1st when in fact intel had already made its choice in 2004. The GPU part of the project failed for the time being . What the future holds is anyones guess. But AMDs assertion that Intel is going in wrong direction are just empty words until we see the fat lady sing.

 

[frozentundra123456]

I dont really know that much about HSA, and whether it has a chance to take off or not. I do know that at one time I was a big fan of the ATI purchase and the Fusion platform. Unfortunately, they had a delayed launch and mediocre performance, especially in the CPU area. Now Intel has almost caught up, or at least has decent integrated graphics.

The whole HSA hype seems eerily like Bulldozer. Basically "we have a great idea, it is just that the software has not caught up yet." What AMD needs to do IMO is to make a great product that utilizes the current software. Especially since they have such a small share of the market, I dont think their initiatives will be widely adapted unless Intel jumps on the bandwagon, and then Intel can probably beat them at their own game.

 

[Dresdenboy]

I found the article to be more hype than anything else sorta PH1 and BD type hype. xtreme has a topic on same article. I love how AMD tries to go back to 2002 . As the foundation when AMD didn't buy ATI till 06. Intel had already choozen their road in 2004 after they bought Elbrus and choose between 3 test projects what we know as larrabbee was choosen in 2004 . That morphed into nights ferry and finnally into nights corner. AMd is just saying we came up with fusion 1st when in fact intel had already made its choice in 2004. The GPU part of the project failed for the time being . What the future holds is anyones guess. But AMDs assertion that Intel is going in wrong direction are just empty words until we see the fat lady sing.

Elbrus is more related to Itanium type architectures...

Gesendet von meinem GT-I9300 mit Tapatalk 2

 

[Dribble]

As far as driving the market and adoption goes in many ways its a similar proposition to CUDA in that it's a smallish company trying to drive adoption of it's new standard.

So looking at the nvidia example - they have been driving this for years, and invested plenty of money. This was helped by their competitors (AMD + Intel) failing to compete in this area for a long time (e.g. Intel have no HW at all after larrabee failed, AMD only just got competing hw in 2012).

Does AMD have the time? Do they have the money to invest in driving adoption (e.g. software engineers)? Will they get free reign due to competitors standards failing as they couldn't make the hardware work?

Answer is pretty well no on all fronts. imo idea is fine, but like many things in AMD it'll only be successful in the marketing slides.

 

[zlatan]

I dont really know that much about HSA, and whether it has a chance to take off or not. I do know that at one time I was a big fan of the ATI purchase and the Fusion platform. Unfortunately, they had a delayed launch and mediocre performance, especially in the CPU area. Now Intel has almost caught up, or at least has decent integrated graphics.

The whole HSA hype seems eerily like Bulldozer. Basically "we have a great idea, it is just that the software has not caught up yet." What AMD needs to do IMO is to make a great product that utilizes the current software. Especially since they have such a small share of the market, I dont think their initiatives will be widely adapted unless Intel jumps on the bandwagon, and then Intel can probably beat them at their own game.

It's very clear that we need a virtual ISA for easier heterogeneous programing. HSA with HSAIL is a possible solution for the problems. It's open for all partners, and it has a very good functional design.

 

[zlatan]

As far as driving the market and adoption goes in many ways its a similar proposition to CUDA in that it's a smallish company trying to drive adoption of it's new standard.

So looking at the nvidia example - they have been driving this for years, and invested plenty of money. This was helped by their competitors (AMD + Intel) failing to compete in this area for a long time (e.g. Intel have no HW at all after larrabee failed, AMD only just got competing hw in 2012).

Does AMD have the time? Do they have the money to invest in driving adoption (e.g. software engineers)? Will they get free reign due to competitors standards failing as they couldn't make the hardware work?

Answer is pretty well no on all fronts. imo idea is fine, but like many things in AMD it'll only be successful in the marketing slides.

CUDA is different. Even if a company licenc it, they don't get the rights to reserve engineering. With the adoption of OpenCL (and later C++ AMP) CUDA will dead at the consumer market.
HSA infrastructure will automaticly take advantage of OpenCL and C++AMP applications.

 

[Cerb]

Answer is pretty well no on all fronts. imo idea is fine, but like many things in AMD it'll only be successful in the marketing slides.

The thing RR needs to do to keep AMD around, and possibly even make them truly successful, is to get on the ball and take control of their markets.

NV didn't wait for people to have an infrastructure ready: they created it, and told customers why they needed it.

AMD has been behind on basically everything lately. A good example is workstation GPUs. As soon as developers started opting for D3D and GLSL, they should have seen the writing on the wall and had ECC ASAP. Instead, they let NV beat them to it. Yeah, they have it on a few, but that's a feather in the cap for AMD, compared to NV's headdress made of an half an eagle.

They got Fusion, sure, but Intel beat them to good integration by a couple generations, and NV already started selling mobile SoCs that can support desktop-derived software development tools.

They implemented CMT very well, saving them space and power, but with a core that is otherwise targeting the opposite of the direction the industry has been headed since ~2001, negating most of the benefits, even in the target server market.

More than anything, AMD needs to decide on a few futures, and make them, wholly, not merely going on with building the hardware and hoping devs and users will come. That method works, but only so long as Intel isn't pressed into making their big money on low ASP parts; a change I suspect will occur one of these years. It also prevents AMD from having many high-volume high-ASP parts.

 

[zebrax2]

^ This
Plus they need to get someone really good for in their marketing department.

 

[Ancalagon44]

^ This
Plus they need to get someone really good for in their marketing department.

They dont need more marketing, they need better management. I dont need AMD to tell me how good their stuff is - I need it to actually be good!

Look at Bulldozer for example. Classic marketing failure example - they hyped it up to be the best gaming CPU ever - the return of FX, etc etc. And it was a complete failure.

You know how they could have handled that one better (besides fixing BD)? By simply not marketing it at us! Market it to the server crowd, maybe even the workstation crowd. But dont insult us by calling it FX, when sucks that badly. Did they think nobody would notice?

They have competent engineers, apparently they have no competent managers.

 

[zebrax2]

Well the post above mine was already talking about the management aspect. Also they do have a pretty competitive GPU division that could use a little bit more marketing

 

[Ferzerp]

They need accurate marketing. Their marketing department is one of the shadiest I know of. Staged events in GPU limited scenarios to sell CPUs, flat out lies (JFAMD), etc. Pretty much we know that if AMD marketing says one thing, the truth is the opposite at this point.

If you don't have the performance (and they don't on the CPU side. GPU side is fine), then don't try to market performance. Market price or something else.

 

[Cerb]

^ This
Plus they need to get someone really good for in their marketing department.

To market oddly-designed products without enough of the right support? No. Marketing is far from unimportant, but they need products that really make sense, and are well-supported. If the industry as a whole isn't doing the support, they need to get more programmers to do it.

Even with BD, they've put out some good products. It's not necessarily all downhill. If the design is flexible enough, they might even be able to rework it enough (caches more like what desktop/mobile users need, fix those latencies, and either make the front-end a little wider, or get the perf/W high enough that we won't care), BD's and Bobcat's children and might not be bad at all for markets where they've maintained some success, lately.

But, they can't keep being late and half-finished, and I don't think there's a silver bullet for that. RR wil need to make good decisions, some of which won't be easy, and be able to figure out the right group of people to work under him. Not only that, but really, none of us have nearly enough information to even know what all of those decisions are, much less be able to effectively affect change. We just see the outcomes, and usually 1-3 years later, at that.

 

[blastingcap]

http://www.tomshardware.com/reviews/fusion-hsa-opencl-history,3262.html

It paints a more positive picture than what it actually was like but it's interesting to see a more behind the scenes look at fusion.

It's long but it gives me hope for AMD's future.

Thanks for the link

 

[BenchPress]

Unfortunately that looks like yet another article which illustrates AMD's desperation to have developers adopt HSA. Without that, it's going nowhere. But the ROI just isn't interesting. It's a lot of effort to develop for various heterogeneous configurations, with uncertain results, and a low install base. In contrast, AVX2 and its successors will offer a steady speedup for minimal effort and the install base will grow more rapidly.

So AMD's only chance at a comeback is to beat Intel at fully homogeneous computing. That is, combining the strengths of CPUs and GPUs into a single unified architecture. Every computer scientist worth his/her salt knows that arithmetic throughput increases faster than bandwidth. Hence heterogeneous computing does not scale. It's why the GPU's shading cores became unified, and the next necessary step will be to unify them with the CPU cores.

Intel is leaps ahead though. Their Xeon Phi product is already a kind of hybrid between a CPU and a GPU. All that's lacking is high single-threaded performance, which can be achieved by using an out-of-order execution architecture, and splitting long vector operations such as AVX-1024 into multi-cycle operations to lower front-end and scheduling overhead (effectively getting the benefits of in-order execution). Intel has already declared to intend to consolidate AVX2+ and the MIC's instruction set to get closer to this goal.

So if AMD wants to have any relevance left in the near future, they'd better act fast and figure out how to optimally unify the CPU and GPU into one and create a seamless synergy to revolutionize the computing landscape.

 

[Soulkeeper]

I find myself still wondering what happened to dirk meyer, no new information there (i've only seen speculation from various people).
I guess the board of directors just didn't agree/like him ?
I was hoping he'd be the guy to pickup the mess after ruiz

 

[zlatan]

So AMD's only chance at a comeback is to beat Intel at fully homogeneous computing. That is, combining the strengths of CPUs and GPUs into a single unified architecture. Every computer scientist worth his/her salt knows that arithmetic throughput increases faster than bandwidth. Hence heterogeneous computing does not scale. It's why the GPU's shading cores became unified, and the next necessary step will be to unify them with the CPU cores.

There are no chance to optimize a core for both low latency and high throughput. For good throughput performance you need dozens of threads for hiding latency.
There is a chance to integrate a throughput optimized core in a latency optimized, but this is also heterogeneous computing and not homogeneous.

 

[BenchPress]

There are no chance to optimize a core for both low latency and high throughput.

I'm sorry but that's complete nonsense. Haswell will double the throughput with AVX2, with no impact on latency.

For good throughput performance you need dozens of threads for hiding latency.

That is not true either. AVX can be extended to 1024-bit instructions, and by executing these on 256-bit units in four cycles the throughput stays the same but it would allow to hide a lot of latency. Also note that Xeon Phi is an in-order architecture but it only needs 4-way threading to hide latency, not dozens. Having fewer threads helps cache coherence and thus hit rate, which in turn reduces the average latency that needs to be hidden.

 

[happysmiles]

I think their marketing need to make AMD a household name, not megahype their products.

 

[Olikan]

i always remember ibm's Cell when reading anyhitng about future apus, HSA or something similar

the gpu at the end will work like cell's SPEs

 

[ShintaiDK]

i always remember ibm's Cell when reading anyhitng about future apus, HSA or something similar

the gpu at the end will work like cell's SPEs

SPEs was RISC CPUs. They got more in common with Xeon Phi and CPUs than they ever will with a GPU.

Cell essentially tried a Larabee. Hence why Sony had to get a GPU in the last minute.

 

[sefsefsefsef]

I'm sorry but that's complete nonsense. Haswell will double the throughput with AVX2, with no impact on latency.

And how does "double throughput" of AVX2 compare to the throughput of a GPU? What are we doubling? Is doubling really good enough? GPUs still have far higher throughput by orders of magnitude than AVX2 will offer in Haswell. There's a reason why Haswell dedicates a large number of its transistors to a GPU and doesn't just use another Haswell core for graphics.

 

[BenchPress]

GPUs still have far higher throughput by orders of magnitude than AVX2 will offer in Haswell.

That is not correct.

There's nowhere near even a single order of magnitude difference, and CPUs continue to catch up. In fact the mainstream desktop Haswell chip will be capable of close to 500 GFLOPS on the CPU cores and only 400 GFLOPS on the iGPU! So it starts to make a lot of sense to fully unify them and have twice the number of homogeneous cores.