Intel Says "Moore's Law isn't dead"

[aneil1998]

So we will have 'mad stacks' of processors then? 

 

Likes SSDs and whatnot. 

so it can be similar to HBM?

[Tomsen]

Ah yes it's transistors per inch per year. My bad. As for heat, yeah maybe, but it depends whether new technology can lead away the heat between the layers. HBM for instance has microbump dummys between the layers to dissipate heat. But as things are now we won't get it anytime soon.

Which stacked dies, you most likely would use some sort of integrated cooling solution (integrated in the chip itself) to combat critical heat-paths.

Also the chip-layout would most-likely also make sure you don't stack those critical heat-paths on top of eachother.

 

 

I personally understand why consumer computing is so far behind professional and enterprise, but damn it if Intel and Co. haven't done all they can to help push it along. If anyone's actually failed in this regard, it's AMD. AMD provided hardware that was in some ways revolutionary, but they never pushed the software side, and I think that sunk them on FX more than anything else. Intel made multithreading easy with OpenMP. You'd have to be braindead to fail at it. Nvidia made GPGPU programming fairly easy with CUDA, proprietary as it is. OpenCL just did not and doesn't yet compare. With 2.1 and native C++ being usable in kernel code (including parallel implementations of the Standard Library Algorithms), some of that may change, but the boilerplate is still ugly and cumbersome.

 

If AMD had backed OpenMP and helped developers with it, it would be in a much better place today.

It doesn't require a doctorate to figure out which computing segment Intel have been focusing on.

If you need a shovel, and I give you a fork, you wouldn't really think of me as helpful?

 

Why wouldn't AMD support openMP? Considering their position..

[Notional]

Which stacked dies, you most likely would use some sort of integrated cooling solution (integrated in the chip itself) to combat critical heat-paths.

Also the chip-layout would most-likely also make sure you don't stack those critical heat-paths on top of eachother.

 

Indeed. Some sort of graphene or nano tubes to remove heat directly from inside the chip. We've already seen prototypes of internal water cooling on FPGA's. Who knows what will become standard in just 5 years.

[patrickjp93]

It doesn't require a doctorate to figure out which computing segment Intel have been focusing on.

If you need a shovel, and I give you a fork, you wouldn't really think of me as helpful?

 

Why wouldn't AMD support openMP? Considering their position..

Intel's focused on enterprise and mobile, but every innovation they come up with for either one ends up implemented on desktop chips too (it's a proving ground for new instructions to hopefully find and eliminate bugs before the E5 and E7 Xeons ramp up).

 

Ask AMD. They didn't push software like they should have. For all of AMD's posturing over open standards, OpenMP is more open than OpenCL. There doesn't exist an open-source implementation of OpenCL last I checked because it's so damn overly complicated. OpenMP has 2 major ones under GNU and under the LLVM umbrella, and they've been around since 2007.

 

Say you want to manipulate the positions of every moving object on screen... Consider Visual Object to be the super class of every type of graphics object you could possibly have that isn't particle effects or the like. Like I said, parallelizing intelligently can be done by a braindead monkey under OpenMP. This even works in Visual Studio (still on OpenMP 2.0 instead of 4.1, but you can have this much). When I consulted at Epic Games, no one had ever heard of OpenMP. That's scary.

std::vector<VisualObject> movableObjects;//initialize everything, set starting coordinates and directions//other prep work//FUNCTIONSvoid update(float updateTimeMillis) {    //move view point, make decisions on color changes, texture loading, etc.    ...    //check for collisions in parallel, using in-lined function calls (no stack frame created)    //and loop-unrolling to make explicit use of branching AVX 256    //assume there is a mutex or semaphore to lock an object for analysis/deletion and a smart    //function to skip if locked and come back built into the calls to inline function     //collisionUpdate    #pragma omp parallel for    for(int i = movableObjects.size()-1; i >= 0; i -= 8) {        movableObjects[i].collisionUpdate(); //will run each type of object's unique function        movableObjects[i-1].collisionUpdate();        ...        movableObjects[i-7].collisionUpdate();    }    //collided objects now destroyed or had states updated to change physical effects routines,    //update all vertices of all objects in parallel,    //ensure dummies exist if not in multiples of 8 to take advantage of AVX 256 in loop unrolling    #pragma omp parallel for    for(int i = movableObjects.size()-1; i >= 0; i -= 8) {         movableObjects[i].update(updateTimeMillis);        movableObjects[i-1].update(updateTimeMillis);        ...        movableObjects[i-7].update(updateTimeMillis);    }        VisualObject.draw(); //draw the scene and all objects in it.} 

[No Nrg]

I think we have a long way before we max out Moore's Law. The manufacturing technology with continue to improve along with the chip tech being produced.

 

Computational power per $1,000 will continue to improve well into the future.

Credit: Ray Kurzweil

[Tomsen]

Intel's focused on enterprise and mobile, but every innovation they come up with for either one ends up implemented on desktop chips too (it's a proving ground for new instructions to hopefully find and eliminate bugs before the E5 and E7 Xeons ramp up).

 

Ask AMD. They didn't push software like they should have. For all of AMD's posturing over open standards, OpenMP is more open than OpenCL. There doesn't exist an open-source implementation of OpenCL last I checked because it's so damn overly complicated. OpenMP has 2 major ones under GNU and under the LLVM umbrella, and they've been around since 2007.

 

Say you want to manipulate the positions of every moving object on screen... Consider Visual Object to be the super class of every type of graphics object you could possibly have that isn't particle effects or the like. Like I said, parallelizing intelligently can be done by a braindead monkey under OpenMP. This even works in Visual Studio (still on OpenMP 2.0 instead of 4.1, but you can have this much). When I consulted at Epic Games, no one had ever heard of OpenMP. That's scary.

std::vector<VisualObject> movableObjects;//initialize everything, set starting coordinates and directions//other prep work//FUNCTIONSvoid update(float updateTimeMillis) {    //move view point, make decisions on color changes, texture loading, etc.    ...    //check for collisions in parallel, using in-lined function calls (no stack frame created)    //and loop-unrolling to make explicit use of branching AVX 256    //assume there is a mutex or semaphore to lock an object for analysis/deletion and a smart    //function to skip if locked and come back built into the calls to inline function     //collisionUpdate    #pragma omp parallel for    for(int i = movableObjects.size()-1; i >= 0; i -= 8) {        movableObjects[i].collisionUpdate(); //will run each type of object's unique function        movableObjects[i-1].collisionUpdate();        ...        movableObjects[i-7].collisionUpdate();    }    //collided objects now destroyed or had states updated to change physical effects routines,    //update all vertices of all objects in parallel,    //ensure dummies exist if not in multiples of 8 to take advantage of AVX 256 in loop unrolling    #pragma omp parallel for    for(int i = movableObjects.size()-1; i >= 0; i -= 8) {         movableObjects[i].update(updateTimeMillis);        movableObjects[i-1].update(updateTimeMillis);        ...        movableObjects[i-7].update(updateTimeMillis);    }        VisualObject.draw(); //draw the scene and all objects in it.} 

Back to my shovel -> fork example. Those innovations aren't as effective in other computing segments.

Enterprise processors require longer and more throughout validations. If it is a bug in the blueprint, then that might not even be fixed with their enterprise processors.

 

My point was that AMD have been pushng for openMP. AMDs software department is undermanned, sure.

But regarding AMDs influence in pushing software standards, I think you are overestimating them.

 

Now, I would imagine console developers  having more insight with with parallism/multi-threading and effeciency over PC developers.

[patrickjp93]

Back to my shovel -> fork example. Those innovations aren't as effective in other computing segments.

Enterprise processors require longer and more throughout validations. If it is a bug in the blueprint, then that might not even be fixed with their enterprise processors.

 

My point was that AMD have been pushng for openMP. AMDs software department is undermanned, sure.

But regarding AMDs influence in pushing software standards, I think you are overestimating them.

 

Now, I would imagine console developers  having more insight with with parallism/multi-threading and effeciency over PC developers.

Except, they are! Just b/c consumer computing nitwits focus way too much on legacy support and not at all on extensibility isn't evidence that Intel's innovations are ineffective for the consumer computing market. My senior and graduate work has been very much proof of that. You can turn most programs into something very similar to my code above. You can vectorize a lot of workloads. Rare is the one you can't, and that boils down to networking (still not entirely true though with the advent of link aggregation and session APIs) and fully dependent serial applications where every step in the program depends on the previous with no room for data parallelism or task parallelism. Those workloads are exceptionally rare. Even ray tracing is possible to do in highly parallel fashion since each pixel ray is independent of the others.

 

No, AMD has made no pushes at all until HSA, and frankly it's DOA compared to the other heterogeneous programming standards.

 

No, as they've simply done nothing. There's no way to overestimate that.

 

They really wouldn't. The "thread" models under the APIs for PS4 and XBOne are so far removed from PC standards it's a wonder anyone knows how to get them to work correctly.

[Tomsen]

Except, they are!

 

No, AMD has made no pushes at all until HSA, and frankly it's DOA compared to the other heterogeneous programming standards.

 

No, as they've simply done nothing. There's no way to overestimate that.

 

They really wouldn't. The "thread" models under the APIs for PS4 and XBOne are so far removed from PC standards it's a wonder anyone knows how to get them to work correctly.

Outside special case workloads, they don't bring any performance benefits. Newer SIMD instructions don't offer any substantially and benefically improvement over the previous.

 

Except they have. Why would you think that? Not only AMD is interested in HSA..

 

It was in regards to this statement:

If AMD had backed OpenMP and helped developers with it, it would be in a much better place today.

I would doubt that the API alone could make up the difference. 

[patrickjp93]

Outside special case workloads, they don't bring any performance benefits. Newer SIMD instructions don't offer any substantially and benefically improvement over the previous.

Except they have. Why would you think that? Not only AMD is interested in HSA..

It was in regards to this statement:

I would doubt that the API alone could make up the difference.

Completely untrue. The code I've written is an example of this. You have to think of every workload in terms of vectors. Rare is the workload that can't be vectorized. Video editting is easily vectorized. So is image and sound editting. So is CAD. So is almost everything in a video game other than networking stuff. SIMD can make 90+% of workloads run better if the coders have the willpower and the skill to design for it.

And further, there's nothing Intel or AMD can do about SISD. Its limits are upon us whether you like it or not. When AMD does catch up and they both putter along slowly afterward on consumer computing, I hope it wakes up the delusional dreamers on here who just think Intel is lazy. The reality is developers are behind and there's nothing anyone can do but increase clock speeds, because those few workloads that can't be vectorized are the same ones that can't be parallelized. More cores won't help with those either. Instruction level parallelism won't help either. I'm sorry but the age of SISD is done. The only programmers hanging on are the bottom of the barrel and the old guard that doesn't have the skill to keep up.

I'm sorry but AMD, had it backed OpenMP and pushed game developers to use it, would have been put in an incredibly better position than they ended up in when Bulldozer and Steamroller/Vishera launched under the FX 81 and 83 series. Having developers build code that scales from 1 core to 100 dynamically would have left AMD much stronger, and more cores would have suddenly mattered. That said Intel still would have run off with the prize since its IPC is so much stronger and they had competitive 6-core options against the 8350 and 9370/9590, but AMD would not have been left clear in the dust as we all witnessed.

That 1 API, OpenMP, is the bread and butter of the HPC world. Everything gets built in MPI for the cluster communication and OpenMP for the individual CPUs. OpenMP further gets used for Xeon Phi offloading, OpenCL for FPGA/GSP offloading, and CUDA for Nvidia Tesla accelerators. The core of the code is all OpenMP. MPI is a thin but robust skeleton that abstracts away the interconnect networking topology for the supercomputer in question, but the muscle is all OpenMP. Sometimes developers go back after initial deployment and rework the threads natively, but OpenMP is already stupidly efficient for a high-level abstraction solution, so the gains from doing the native version are tiny, and the time to do it is significant. It's literally as easy as throwing parallel pragmas around. AMD should have backed it with Intel and helped game studios adapt. They didn't. That was a complete failure of leadership when they actually had the leverage to pull it off (before the Bulldozer fallout).

There's no excuse for consumer developers several years into their craft to be behind college sophomore students. I'm sorry but there just isn't. You, the consumers have been fighting the wrong battle going after Intel. Intel did all the work on both sides, filled up the oasis with fresh water (hardware) and led the horse right up to it (open, easy programming standards). AMD gets the praise for HSA which is 15 years late (wrt OpenMP) and which no one is interested in until AMD provides a working proof of concept at the enterprise level, and even then the programming model is so disgusting the HPC crowd is already turned off to it. It took the worst parts of C++ 99 and OpenCL 1.x and threw it together into a squirrely mess. Yes, HSAF has backers in Samsung and Imagination Technologies, but do you see them implementing anything? I don't. Intel and Nvidia have both said it's doomed to fail. The Foundation members are waiting for AMD to prove them wrong. It will not get off the ground unless AMD pulls a serious 2.0 revision that completely overhauls HSAIL.

In summation, AMD had failures of intelligence and leadership that stagnated the entire market. Yes, Intel paid OEMs to not use AMD's products, back in 2007... However, you want to know why we're stuck right now? Developers are behind and AMD is in the way floundering with a concept that has no hope of ever working in its current form when already there is a perfectly good, widely-used alternative that can be adapted to other parts of the market if AMD helps Intel make that push. Don't doubt the API until you try it. It's very difficult to do better multithreading than OpenMP, both in terms of raw performance and in efficiency.

[patrickjp93]

I think we have a long way before we max out Moore's Law. The manufacturing technology with continue to improve along with the chip tech being produced.

 

Computational power per $1,000 will continue to improve well into the future.

 

Credit: Ray Kurzweil

Moore's Law is purely about transistor density, not their clock speed or performance.

[mr.skyrim_pc]

Is it really a law if it isn't a law in a decade?

[Tomsen]

Completely untrue. The code I've written is an example of this. You have to think of every workload in terms of vectors. Rare is the workload that can't be vectorized. Video editting is easily vectorized. So is image and sound editting. So is CAD. So is almost everything in a video game other than networking stuff. SIMD can make 90+% of workloads run better if the coders have the willpower and the skill to design for it.

It is completely true. The code snippet you have written are great examples for few certain workloads. Still does not hold true for much else.

SIMD can't make 90% of workloads run better. Don't foul yourself.

 

And further, there's nothing Intel or AMD can do about SISD. Its limits are upon us whether you like it or not. When AMD does catch up and they both putter along slowly afterward on consumer computing, I hope it wakes up the delusional dreamers on here who just think Intel is lazy. The reality is developers are behind and there's nothing anyone can do but increase clock speeds, because those few workloads that can't be vectorized are the same ones that can't be parallelized. More cores won't help with those either. Instruction level parallelism won't help either. I'm sorry but the age of SISD is done. The only programmers hanging on are the bottom of the barrel and the old guard that doesn't have the skill to keep up.

Aaaargh, *sigh*..

First thing; It can also be extended to general servers aswell.

Fortunately, there are things that can be done to speed up single-thread performance.

Intel most certainly are near their current implementations limit. Only minor gains are had optimizing existing architecture.

No, I'm not saying Intel is been lazy. You are right, that there are no **free** performance. However, it does have compromises in other areas which Intel do think of as important. Also, there are other variables that needs to scale accordingly, which simply hasn't been available. Don't you think the smart people at Intel/AMD cannot build something performing better. Don't you think there are Software devs sure are half the problem. The other half is ineffective tools been provided. It it about having the right tool for the right workload, which is an area where a LOT of research have been done.

Clockspeed most likely wont become much higher, unless we start seen asynchronized clockspeed.

**We've tried nothing and we're all out of ideas**

 

I'm sorry but AMD, had it backed OpenMP and pushed game developers to use it, would have been put in an incredibly better position than they ended up in when Bulldozer and Steamroller/Vishera launched under the FX 81 and 83 series. Having developers build code that scales from 1 core to 100 dynamically would have left AMD much stronger, and more cores would have suddenly mattered. That said Intel still would have run off with the prize since its IPC is so much stronger and they had competitive 6-core options against the 8350 and 9370/9590, but AMD would not have been left clear in the dust as we all witnessed.

Great example of how things dont work. Push game developers to use openMP? Really, those are the ones you expected for AMD to council about openMP?

AMD aren't getting shit across them, unless AMD solely are the ones investing in it. Wasteful ressources spend.

How about something like AMD council their actual customers/partners, where performance would actually be critical and software been tailored around the hardware.

With the current pricing? No, Intels 6-core would be laughed at, however, I doubt we would see AMD selling their processors for that low if that was the case.

 

That 1 API, OpenMP, is the bread and butter of the HPC world. Everything gets built in MPI for the cluster communication and OpenMP for the individual CPUs. OpenMP further gets used for Xeon Phi offloading, OpenCL for FPGA/GSP offloading, and CUDA for Nvidia Tesla accelerators. The core of the code is all OpenMP. MPI is a thin but robust skeleton that abstracts away the interconnect networking topology for the supercomputer in question, but the muscle is all OpenMP. Sometimes developers go back after initial deployment and rework the threads natively, but OpenMP is already stupidly efficient for a high-level abstraction solution, so the gains from doing the native version are tiny, and the time to do it is significant. It's literally as easy as throwing parallel pragmas around. AMD should have backed it with Intel and helped game studios adapt. They didn't. That was a complete failure of leadership when they actually had the leverage to pull it off (before the Bulldozer fallout).

Developers usually do more customizations in software to increase performance efficiency. A lot of software analyses will be done, to optimize areas in which openMP default implementation will slow down in performance (which could be due to any number of reasons).

So, how much did Intel back openMP for consumers? Are intel pushing game developers to use openMP?

 

There's no excuse for consumer developers several years into their craft to be behind college sophomore students. I'm sorry but there just isn't. You, the consumers have been fighting the wrong battle going after Intel. Intel did all the work on both sides, filled up the oasis with fresh water (hardware) and led the horse right up to it (open, easy programming standards). AMD gets the praise for HSA which is 15 years late (wrt OpenMP) and which no one is interested in until AMD provides a working proof of concept at the enterprise level, and even then the programming model is so disgusting the HPC crowd is already turned off to it. It took the worst parts of C++ 99 and OpenCL 1.x and threw it together into a squirrely mess. Yes, HSAF has backers in Samsung and Imagination Technologies, but do you see them implementing anything? I don't. Intel and Nvidia have both said it's doomed to fail. The Foundation members are waiting for AMD to prove them wrong. It will not get off the ground unless AMD pulls a serious 2.0 revision that completely overhauls HSAIL.

They are not "behind" you. What you don't seems to understand, is that these developers, will only have limited ressources and have to follow certain directions.

Any actual application been critical for performance, will most likely already be well optimized.

You're so apolegetic towards Intel.

HSA isn't 15 years late. What about other comparable technologies? Are they even later??

Again, you are speaking out of your arse. HSA seems to bring interest forth on partners. More partners are comming in.

Now, HSA aren't only targeting HPC. Lot of interest are comming from mobile market too. Semi-custom even more.

You have to consider time-to-market. Implementations will come. At the right time.

Intel and Nvidia don't really got any better solution, do they?

 

In summation, AMD had failures of intelligence and leadership that stagnated the entire market. Yes, Intel paid OEMs to not use AMD's products, back in 2007... However, you want to know why we're stuck right now? Developers are behind and AMD is in the way floundering with a concept that has no hope of ever working in its current form when already there is a perfectly good, widely-used alternative that can be adapted to other parts of the market if AMD helps Intel make that push. Don't doubt the API until you try it. It's very difficult to do better multithreading than OpenMP, both in terms of raw performance and in efficiency.

AMD stagnated the entire market? Wow, I didn't realize AMD had such an influence.

You would think, which such an influence in software, AMD were able to gain on it?

Intel payed OEMs up to 2007 atleast.

We are stuck because Intel didn't advance their processors in any useful way. Intel keeps undermining the market (due to low x86 competition), which is one reason to why the PC market have stagnated. What exactly isn't working in its current form?

There is a need for this new technology, nothing current technologies can fullfil.

You are sadly just over-praising it. You are overestimating the pros and underestimating the cons.

[patrickjp93]

***Garbage rant not based in reality***

 

Great example of how things dont work. Push game developers to use openMP? Really, those are the ones you expected for AMD to council about openMP?

AMD aren't getting shit across them, unless AMD solely are the ones investing in it. Wasteful ressources spend.

How about something like AMD council their actual customers/partners, where performance would actually be critical and software been tailored around the hardware.

With the current pricing? No, Intels 6-core would be laughed at, however, I doubt we would see AMD selling their processors for that low if that was the case.

AMD doesn't need to push their other big customers. They're already doing it. Workstation and HPC picked it right up nearly a decade ago.

 

Intel's 6-core Sandy Bridge still smokes the FX 9590 in heavily threaded workloads. It wouldn't have been laughed at, especially since the 9590 launched for $1000.

 

 

Developers usually do more customizations in software to increase performance efficiency. A lot of software analyses will be done, to optimize areas in which openMP default implementation will slow down in performance (which could be due to any number of reasons).

So, how much did Intel back openMP for consumers? Are intel pushing game developers to use openMP?

 

They are not "behind" you. What you don't seems to understand, is that these developers, will only have limited ressources and have to follow certain directions.

Any actual application been critical for performance, will most likely already be well optimized.

You're so apolegetic towards Intel.

HSA isn't 15 years late. What about other comparable technologies? Are they even later??

Again, you are speaking out of your arse. HSA seems to bring interest forth on partners. More partners are comming in.

Now, HSA aren't only targeting HPC. Lot of interest are comming from mobile market too. Semi-custom even more.

You have to consider time-to-market. Implementations will come. At the right time.

Intel and Nvidia don't really got any better solution, do they?

 

AMD stagnated the entire market? Wow, I didn't realize AMD had such an influence.

You would think, which such an influence in software, AMD were able to gain on it?

Intel payed OEMs up to 2007 atleast.

We are stuck because Intel didn't advance their processors in any useful way. Intel keeps undermining the market (due to low x86 competition), which is one reason to why the PC market have stagnated. What exactly isn't working in its current form?

There is a need for this new technology, nothing current technologies can fullfil.

You are sadly just over-praising it. You are overestimating the pros and underestimating the cons.

The only way OpenMP slows you down is if you're an idiot.

 

Intel invented OpenMP and then made it an open standard. Second, the entire Havok CPU physics engine was built using it.

 

No, they leave optimization to the last minute instead of designing for it in the first place (speaking from experience when consulting for Epic Games this past summer). The optimizers get called in last after all the code's been written (poorly, with little documentation). And they are behind me. They've never even heard of half the techniques I discussed at length with senior developers.

 

OpenMP beat HSA to the punch, in 2000, and it has been keeping up more than anyone gives it credit for. Intel has kept up with every technology the HSAF implements (such as unified memory between the CPU and iGPU), but the programming model of HSA is one it doesn't believe in (and has publicly stated so), and frankly with neither Intel nor Nvidia nor IBM backing it, it's going nowhere fast. 

 

No, HSAF has fewer members today than it did even 3 years ago: http://www.hsafoundation.com/ Most notably IBM and Citrix who both dropped around the same time.

 

Yes, they do, between OpenMP and OpenACC the solutions exist today and are easy to program for while having every performance benefit HSA ever could hope to achieve. 

 

Yes, AMD went off on a tangent and didn't do anything realistic or helpful for the industry for 6 years.

 

Intel's done plenty to improve its processors in useful ways and has more than doubled the compute power of its processors in the last 3 years. Developers are not keeping up. Give me a common workload and I'll vectorize and parallelize it. You will be hard-pressed to come up with something that truly cannot be usefully parallelized. They are rare and niche. How is Intel undermining the market when it's the only processor maker doing the work?

 

You haven't pointed out any cons. No one has.

[Tomsen]

AMD doesn't need to push their other big customers. They're already doing it. Workstation and HPC picked it right up nearly a decade ago.

 

Intel's 6-core Sandy Bridge still smokes the FX 9590 in heavily threaded workloads. It wouldn't have been laughed at, especially since the 9590 launched for $1000.

I would heavily doubt all their costumers already are using openMP. How much influence does AMD have with workstation and HPC CPUs? Not much. There is more than consumer, workstation and HPC. Much more. Where do you think AMD finds it business?

I'm not saying it didn't. I'm saying, considering todays prices, a 8320/8350 would make any Intel +6 core pro/consumer processor laughable. However, that is *IF* prices remain the same as of today (as said, prices wouldn't be the same if that was the situation).

Good thing there aren't the any real difference between 8320 to 9590. And that $1000 on the 9590 is long gone, due to lack of competition.

 

The only way OpenMP slows you down is if you're an idiot.

 

Intel invented OpenMP and then made it an open standard. Second, the entire Havok CPU physics engine was built using it.

Yes, it was worded inproperly. Areas in which openMP don't offer any substansial performance benefits.

Now, I haven't heard that Intel invented openMP.

 

No, they leave optimization to the last minute instead of designing for it in the first place (speaking from experience when consulting for Epic Games this past summer). The optimizers get called in last after all the code's been written (poorly, with little documentation). And they are behind me. They've never even heard of half the techniques I discussed at length with senior developers.

From my previous statement:

Any actual application been critical for performance, will most likely already be well optimized.

 

OpenMP beat HSA to the punch, in 2000, and it has been keeping up more than anyone gives it credit for. Intel has kept up with every technology the HSAF implements (such as unified memory between the CPU and iGPU), but the programming model of HSA is one it doesn't believe in (and has publicly stated so), and frankly with neither Intel nor Nvidia nor IBM backing it, it's going nowhere fast.

You would think I wouldn't take openMP as a comparable technology.. Maybe Intels whole "neo-heterogeneous" computing with its software stack could be considered a comparable technology. Intel did take a fundamentally different approach. Both have pros and cons. Also, how is that implementation back in 2000 looking compared to todays? I bet there have been some major changes.

Now, I can't remember Intels hUMA/hQueuing implementation on top of my head.

Is it a *full* hUMA implementation or only partially, and have they even implemented hQueuing?

Well, AMD needs to put together a software stack for HSA, which they have been building up.

Being so flexible, makes it more attractive for semi-custom for a wide range of applications.

However, that have little effect on PC/general purpose server to do.

Why would you think that not having Intel, Nvidia or IBM would slow it down?

It is not such a small world, the sea is filled with interested people.

ARM and MIPS both seems to be working on some HSA compliance. These two, will have more affection on the industry, due to their business strategy of core IP licensing.

 

No, HSAF has fewer members today than it did even 3 years ago: http://www.hsafoundation.com/ Most notably IBM and Citrix who both dropped around the same time.

Now, I haven't been keeping a count, but have you?

However, have been seen new parterns showing up recently.

 

Yes, they do, between OpenMP and OpenACC the solutions exist today and are easy to program for while having every performance benefit HSA ever could hope to achieve.

That is some claim. Maybe in your fantasy land.

Neither of those solution offer the flexible as HSA in hardware.

Customers can in theory make IP blocks compliant for HSA giving improved performance, efficiency and ease of programmability.

 

Intel's done plenty to improve its processors in useful ways and has more than doubled the compute power of its processors in the last 3 years. Developers are not keeping up. Give me a common workload and I'll vectorize and parallelize it. You will be hard-pressed to come up with something that truly cannot be usefully parallelized. They are rare and niche. How is Intel undermining the market when it's the only processor maker doing the work?

The theoretical computing power has doubled, if the workflow follows a certain paradigm. Aka certain workloads see these gains, the rest lies within our usual 0-10%.

Most complex workflows dont stand to gains any substantial benefit. I wouldn't exactly call them rare..

Can we both agree, that the lack of competition, have made Intel change certain priorities. Not in favor of PC.

Yes, 0-5% change per generation sure is doing work. No bigger changes to their micro-architecture in a high-level.

Making a unified core, for multiple seperate markets, all favoring different power-, performance- and cooling-characteristics, does have its disadvantages.