Now that smartphones get more powerful and internet makes it possible to have all functionality and documents with you anywhere, the computer needs to be reinvented. You see all big IT-companies searching for how that can be, from Windows Metro to complete docking stations to replace the desktop by your phone. A turbulent market.
One of the new products are USB-stick sized computers. Stick them into a TV or monitor, zap in your code and you have your personal working environment. You never need to carry laptops to your hotel-room or conference, as long as a screen is available – any screen.
There are several USB-computers entering the market, but I wanted to introduce you to two. Both of these see a future in a strong processor in a portable device, and both do not have a real product with these strong processors. But you can expect that in 2013 you can have a device that can do very fast parallel processing to have a smooth Photoshop experience… at your key-ring.
Steve Streeting tweeted a few weeks ago: “Remember, experts are always wrong about disruptive tech, because it disrupts what they’re experts in.”. I’m happy I evangelise and work with such a disruptive technology and it will take time until it is bypassed by other technologies. And that other technologies will be probably be source-to-OpenCL-source compilers. At StreamComputing we therefore keep track of all these pre-compilers continuously.
Steve’s tweet got me triggered, since the stability-vs-progression-balance make changes quite hard (we see it all around us). Another reason was heard during the opening-speech of engineering world 2011 about “the cloud”, with a statement which went something like: “80% of today’s IT will be replaced by standardised cloud-solutions”. Most probably true; today any manager could and should click his/her “data from A to B”-report instead of buying a “oh, that’s very specialised and difficult” solution. But at the other side companies try to let their business live as long as possible. It’s therefore an intriguing balance.
So I came up with the idea to play my own devil’s advocate and try to disrupt GPGPU. I think it’s important to see what can disrupt the current parallel-kernel-execution model of OpenCL, CUDA and the others.
Both AMD and Intel have added parallel instruction-sets for their CPUs to accelerate in media-operations. Each time a new instruction-set comes out, code needs to be recompiled to make use of it. But what about support for older processors, without penalties? Intel had some troubles with how to get support for their AVX-instructions, and choose for both their own Array Building Blocks and OpenCL. What I want to discuss here are the possibilities available to make these things easier. Also I want to focus on if a general solution “OpenCL for any future extensions” could hold. I make an assumption that most extensions target mostly parallelisation with media in mind, most notable embedded GPUs on upcoming hybrid processors. I talked about this subject before in “The rise of the GPGPU compiler“.
Java started in 1996 with the idea that end-point optimisation should be done by compiling intermediate code to the target-platform. The idea still holds and there are many possibilities to optimise intermediate code for SSE4/5, AVX, FMA, XOP, CLMUL and any other extension. Same is of course for dotNET.
Disadvantage is the device-models that are embedded in such compilers, which have not really take specialised instructions into account. So if I have a normal loop, I’m not sure it will work great on processors launched this year. C has pragmas for message-protocols, Java needs extensions. See Neal Gafter’s discussion about concurrent loops from 2006 for a nice discussion.
With for instance LLVM and Intel’s fast compilers, a lot can be done to get code optimised for all current processors. A real danger is that too many specialised processors will arrive the coming years; how to get maximum speed at all processors? We already have 32 and 64 bit; 128 bit is really not the only direction there is. Multi-target compilers can be something we should be getting used to, for which no standard is created for yet – only Apple has packed 32 and 64 bits together.
Years ago when CPUs started to have support for the multiply-add operation, a part of the compiled code had to be specially for this type of processor – giving a bigger binary. With any new type of extension, the binary gets bigger. It has to, else the potential of your processor will not be used and sales will drop in favour of cheaper chips. To sell software with support for each new extension, it takes time – in most cases reserved only for major releases.
Because not everybody has Gentoo (A Linux-distribution which compiles each piece of software targeting the user’s computer for maximum optimisation), it takes at least a year to get full use of the processor for most software.
So where does OpenCL fit in this picture? Virtual machines are optimised for threads and platform-targeting compilers are slow in distribution. Since drivers for CPUs are part of the OS-updating system, OpenCL-support in those drivers can get the new extensions utilised soon after market-introduction. The coming year more will be done for automatic optimisation for a broad range of processor-types – more about that later. This focus from the compiler to an OpenCL-library for handling optimal kernel-launching will get an optimum somewhere in between.
The coming time we will see OpenCL is indeed a more stable solution than for instance Intel’s Array Building Blocks, seen from the light of recompiling. If OpenCL can target all kinds of parallel extensions, it will offer the demanded flexibility the market demands in this diversifying processor-market. I used the word ‘demand’, because the consumer (being it an individual or company) who buys a new computer, wants his software to be faster, not potentially faster. What do you think?
If you read The Disasters of Visual Designer Tools you’ll find a common thought about easy programming: many programmers don’t learn the hard-to-grasp backgrounds any more, but twiddle around and click a program together. In the early days of BASIC, you could add Assembly-code to speed up calculations; you only needed tot understand registers, cache and other details of the CPU. The people who did that and learnt about hardware, can actually be considered better programmers than the C++ programmer who completely relies on the compiler’s intelligence. So never be happy if the control is taken out of your hands, because it only speeds up the easy parts. An important side-note is that recompiling easy readable code with a future compiler might give faster code than your optimised well-engineered code; it’s a careful trade-off.
Okay, let’s be honest: OpenCL is not easy fun. It is more a kind of readable Assembly than click-and-play programming. But, oh boy, you learn a lot from it! You learn architectures, capabilities of GPUs, special purpose processors and much more. As blogged before, OpenCL probably is going to be the hidden power for non-CPUs wrapped in something like OpenMP.
The System-on-a-chip (SoC) for X86 will be a revolution for GPGPU. Why? Because currently a big problem is transferring data from CPU-memory to GPU-memory and back, which will be solved with SoCs. Below you can read this architecture-target is very possible.
With AMD+ATI, Intel and its future high-end GPUs, and NVidia with the rumours around its X86-chips, we will certainly get changes in the field. If it is the way to go, what is probable?
- Get both CPU and high-end GPU on 1 chip, separated memory
- Techniques for sharing memory
- Translating OpenCL from and to C on the fly
ARM-processors are combined with GPUs a lot of times, but they don’t have current support for a common shader-languages (read: OpenCL) to make GPGPU in reach. We’ve asked ourselves many times why ARM & friends are involved in OpenCL since the beginning, but still don’t have any public and promoted driver-support. More on ARM, once there is more news on multi-core ARM-CPUs or OpenCL drivers.
1: One chip for everything
The biggest problem with split CPU/GPU-functionality is the bus-speed between the two is limited. The higher this speed, the more useful GPGPU can be. The highest speeds are possible when the signal does not have to leave the chip and there are no concessions made to the architecture of the graphics-card, in other words: glueing CPU and GPU together, but leave the memory-buses the same.
Currently there is Intel’s Nehalem and AMD’s Fusion, but they use DDR3 for both GPU and CPU; this will not really unlock the GPGPU-possibilities of high-end GPUs. It seems these products were designed with lower costs in mind.
But the chances high-end GPUs will be integrated on the CPU is rising. Going to 32nm gives room for more functionality, such as GPUs. Other choices can be smaller chips, more cores and integrating functionality of the north/south-bridge of the motherboard. If GPU-cores can be turned off when not working optimally when being tested in the factory (just like they do with mult-core CPUs), integrating high-end GPU-cores will even become a save choice.
Another way it could go is using optical buses between the GPU and CPU. It’s unknown if it will really see mainstream markets soon enough.
2: Shared memory – new style
Some levels of cache and all memory should be easy accessible by both types of cores. Why? Because eventually you want to switch between CPU- and GPU-instructions continuously. CUDA has a nice feature already, which keeps objects synchronised between CPU and GPU; one step further is leaving out the need of synchronising.
The problem is that video-memory is accessed more parallel to provide higher data-speeds (GDDR5), so we don’t want to limit the GPU by attaching them to slower (=lower bandwidth) DDR3. Doing it the other way would then be the best solution: giving CPUs direct access to GDDR. There is always a probable option that a new type of (replaceable) memory will be used, which has a dual-bus by design.
The hard part is memory-protection; since now more devices get control to memory, the overhead of controlling/arranging the spots can increase enormously and might need a separate core for it – just like the Cell-processor. This need-for-control is a reason I don’t expect access to each other memory before there will be a fast bus between GPU and CPU, since then the access to GDDR via the GPU’s memory-manager will be much faster and maybe even fast enough.
3: Grown up software
If software would be able to easily select devices and use the same code for each device, then we’ve made a giant step forwards. Software has always been one step behind hardware; so when you do not develop such techniques, you just have to wait a while.
Translating OpenCL into normal C and back will be possible in all kinds of ways, once there is more acceptance of (and thus demand for) GPGPU. AMD’s OpenCL-implementation for CPUs is also a way to merge the fields of CPU and GPU. It’s hard to tell how these techniques will merge, but it will certainly happen. Think of situations that some instructions will be sent to the GPU by the OS even when the (OpenCL) programmer did not think of it. Or do you expect to be an ARM-processor integrated in a near-future CPU, when you write an OpenCL-kernel now?
See our article on the bright future of GPGPU to read more about it.
In case this is the way it goes, there will be a lot possible for both OpenCL and CUDA – depending on market demands. Some possibilities will be discussed in an upcoming article about FPGAs, but also let me hear what you think about X86-SoCs. Comment or send an e-mail.