Noob QUestion:
If AVX2 really is such a killer, why hasn't it been implemented like 10 years ago?
Same reason we didnt have SB/IB 10 years ago.
Where's the official HSA and OpenCl threads?
GPU competition?However if AVX2 increases performance 3x or more (without using 3x more die space) one would think it should have been implemented a very long time ago?
And that reason is?
The reason for not having an SB 10 years ago is process technology. 4 cores with integrated GPU ain't really great on like 130 nm.
However if AVX2 increases performance 3x or more (without using 3x more die space) one would think it should have been implemented a very long time ago?
Only one of these technologies will prevail. There is no room for both since both attempt to cover the need for general purpose throughput computing. History shows that incompatible competing technologies cannot coexist. Think about AMD64 versus IA64: Itanium is practically dead. Think about 3DNow! versus SSE: Bulldozer no longer supports 3DNow!.
So the question now is which is the superior throughput computing technology: homogeneous AVX2+ or heterogeneous GPGPU? And yes, both companies will support both for a while, but they have a different idea of what to focus on. There's a lot at stake for AMD since it's sacrificing CPU performance to make the GPU more powerful, in an attempt to make GPGPU more attractive. Not just that, it's also sacrificing graphics performance. As illustrated by NVIDIA's Fermi and Kepler, graphics and GPGPU require different architectures. HSA leans very much toward GPGPU, which compromises graphics.
Intel doesn't make any sacrifices. It already has a superior CPU architecture and it will be the first to add high throughput performance to it using AVX2. Even when AMD implements AVX2, there will still be a big difference in computing density because of Bulldozer's shared SIMD cluster architecture. There's also no sign of Intel sacrificing graphics performance for the sake of GPGPU. And last but definitely not least, AVX2 is much easier to adopt by developers than GPGPU, and will offer more consistent performance across system configurations.
Easier, yes, but it will never be easy. In fact heterogeneous computing becomes harder when things scale up. So they're fighting an uphill battle. The only way to guarantee that it doesn't suffer from bad latency and bandwidth scaling, is to fully merge the GPU technology into the CPU. And that's what AVX2 already does!
It's no coincidence that Intel's Knights Corner chip, which is pretty much a GPU architecture (minus the graphics components), uses an instruction set that has a very close resemblance to AVX2.
So it's inevitable that things will converge into a single architecture. All general purpose computing will happen on the CPU. The GPU either has to become fully focused on graphics, or the programmable shaders too get processed on the CPU and the GPU decays into some fixed-function units that act as peripheral components which assist the CPU in graphics processing.
AMD desperately wants the CPU and GPU to remain heterogeneous, but in doing so it ironically converges them closer together, making the case for AVX2 and its successors.
I guess the ability of designing DLP software could also be an issue..I thought it was because we didnt have the design knowledge at the time.
Maybe you should call all the other companies that makes different products too. They obviously also wasted alot of time on middle designs.
AVX represents a simple vector way by applying it to normal x86 CPU to enhance theoretical performance. OpenCL can use it but mostly OpenCL refer to the way the nonhomogeneous way AMD use.I completely disagree. The two are very different and aren't competing at all. They even work together. I just don't understand the AVX2 or Open CL talks, since Open CL is a method for many different processors to use the same code, while AVX2 is just an Intel construct that Open CL can use. Maybe someone can explain why they feel the two things are competing?
OpenCL isn't even mentioned in my post. And that's because the discussion is not about AVX2 against OpenCL at all. It's about homogeneous versus heterogeneous throughput computing. That said, AVX2 is homogeneous and OpenCL is aimed at heterogeneous computing.I completely disagree. The two are very different and aren't competing at all. They even work together. I just don't understand the AVX2 or Open CL talks, since Open CL is a method for many different processors to use the same code, while AVX2 is just an Intel construct that Open CL can use. Maybe someone can explain why they feel the two things are competing?
The past doesn't support that expectation. For instance AMD's 3DNow! capable CPUs were released more than a year before Intel added support for SSE. But SSE became the dominant SIMD technology, simply because it was superior. AMD ended up also implementing SSE, and with Bulldozer the 3DNow! support was removed entirely.If it is an either or thing, which it is not, then I would expect Open CL to take greater hold than AVX2. Open CL can be run on millions of current processors and devices while AVX2 isnt even available yet and will only run on future devices. We have seen how such things have fared in the past.
The past doesn't support that expectation. For instance AMD's 3DNow! capable CPUs were released more than a year before Intel added support for SSE. But SSE became the dominant SIMD technology, simply because it was superior. AMD ended up also implementing SSE, and with Bulldozer the 3DNow! support was removed entirely.
Likewise we're still in the early days of throughput computing technology. In fact AMD's HSA roadmap won't even be finished before 2014. They need that technology to make GPGPU computing more widely applicable. AVX2 will arrive in 2013, and is more versatile than HSA will ever be.
But OpenCL will likely become a victim of AVX2. Homogeneous computing devices can run heterogeneous code, but heterogeneous ones can't run all homogeneous code.
Once again you're making this about AVX2 versus OpenCL, which it isn't. But even if it was, "supporting" OpenCL doesn't mean a thing. The GTX 680 supports OpenCL but it loses against a quad-core CPU, even before AVX2. And the average GPU is much weaker.3D Now was a AMD only thing, and only a small percentage of processors supported it. OpenCL is supported by nearly every modern processor and GPU, so a very large percentage of systems support it. In this situation Intel is in the position of weakness like AMD was with 3D Now.
MMX was used extensively in multimedia codecs and audio drivers. It was a monumental success.Also MMX was released before 3D Now. Neither of those extensions actually gained any traction though.
Once again you're making this about AVX2 versus OpenCL, which it isn't. But even if it was, "supporting" OpenCL doesn't mean a thing. The GTX 680 supports OpenCL but it loses against a quad-core CPU, even before AVX2. And the average GPU is much weaker.
It's really about whether homogeneous or heterogeneous throughput computing is superior. And for heterogeneous computing to have any chance at all, it needs affordable GPUs to have higher efficiency. This requires support for things like a uniform address space and preemption. These features won't be available until 2014. CPUs with AVX2 support will arrive a year earlier.
Last but not least, Intel has an 80% market share, so they're definitely not in a position of weakness. In the not too distant future, 80% of all new x86 systems will have great homogeneous throughput computing performance. Sooner or later AMD will have to implement AVX2 as well and it will be 100% of all systems. Meanwhile how many systems will have worthy heterogeneous computing support? The desktop Haswell chip will be limited to a GT2, which means the CPU with AVX2 will have more raw computing power than the iGPU. And that's without taking into account that the CPU has far fewer bottlenecks, and is much easier to develop for.
MMX was used extensively in multimedia codecs and audio drivers. It was a monumental success.
A CPU can execute absolutely anything. Including advanced graphics, e.g. through SwiftShader. It even runs Crysis if you want, and variants of it power WebGL in Google Chrome and Adobe's Stage3D for Flash when your GPU isn't adequate.Wouldn't it be the opposite?
Homogeneous computing devices can only run the code designed for them.
For example, as far as I know, an x86-CPU cannot perform, in Windows, hardware accelerated graphics. An APU (heterogeneous device) can.
Wouldn't it be the opposite?
Homogeneous computing devices can only run the code designed for them.
For example, as far as I know, an x86-CPU cannot perform, in Windows, hardware accelerated graphics. An APU (heterogeneous device) can.
Great. Like I said before, OpenCL is not the enemy of AVX2.I have a computer that can run Open CL right now. I am pretty sure you will too. I wont have one that will run AVX2 for who knows how long. My agency sure won't write code for AVX2 if it means they need to upgrade every single computer in 2013 to make use of it. The fact that practically everything built within the last few years will bet able to run Open CL code and make use of it would make companies far more likely to code for it. 80% of new computers is one thing, but that is still a rather small percentage of total userbase.
My agency sure won't write code for AVX2 if it means they need to upgrade every single computer in 2013 to make use of it.
The GTX 680 supports OpenCL but it loses against a quad-core CPU, even before AVX2.
Else why is it that a 3000 GFLOPS GPU loses against a 230 GFLOPS CPU? And that's still a CPU without AVX2, and the average system will have a much weaker GPU!
But you have to look longer term. If homogeneous computing prevails over heterogeneous computing, then sooner or later there will be no point in using OpenCL. It imposes limitations, and homogeneous throughput computing technology can be used with any programming language.
What counts is that it does prove the point. A high-end GPU is losing, in an OpenCL benchmark, against a CPU that doesn't even have AVX2 yet. That's no less than three reasons why GPGPU will have a really hard time going mainstream:It's a bit hypocritical to knowingly cherry pick to prove a point.
AVX2 won't be the only homogeneous throughput computing technology. ARM's NEON extension already has FMA, and adding gather support should be relatively straightforward. Also, from a performance/Watt perspective it makes more sense to widen the vectors than to add more cores. The end result would be the ARM equivalent of AVX2. Let's call this NEON2. They'd be stupid not to already started working on this.Limitations on hardware support, for instance? That's a limitation. In fact, that's quite a big limitation, wouldn't you say? x86-only proprietary ISA means it's only limited to Intel/AMD chips.
No it doesn't. Some high-end GPUs running OpenCL are losing against CPUs, prior to AVX2. You can't ignore this. Mainstream GPUs are much weaker too. And while HSA indeed aims to improve GPGPU efficiency, it will not be widely supported.HSA is meant to allow openCL to prosper and openCL is ubiquitous and has and will always outpace AVX2...