Using graphics processing units (GPUs) to perform general purpose computing (GPGPU) has been all the rage in the past few years. In fact, the second fastest high performance computer (HPC) in the world today uses NVidia GPUs. But for all their benefits, there is a problem with using GPUs to solve problems. The problem is that the bandwidth between the GPU and the rest of the system is limited to what the PCI Express bus can offer. Again, this can be seen in the second fastest HPC system in the world. That machine can only reach 46% of theoretical peak performance, while the fastest HPC machine can reach 80% of theoretical peak performance. The solution is to run as much of your code on the GPU as possible. If a researcher can do that, then they can potentially run their codes hundreds of times faster than on a CPU.
That is what makes the latest paper from Intel so interesting. In the paper, Intel showed that you could only achieve a 14x speedup. That is a nice, backhanded compliment, Intel.
The paper is called “Debunking the 100x GPU vs CPU Myth” and it is indeed true that not *all* applications can see this kind of speed up, some just have to make do with an order of magnitude performance increase. But, 100X speed ups, and beyond, have been seen by hundreds of developers.
via The NVIDIA Blog – “GPUs Are Only Up To 14 Times Faster than CPUs” says Intel.
>”The problem is that the bandwidth between the GPU
>and the rest of the system is limited to what the
>PCI Express bus can offer.”
It is quite a misleading statement; even the PCI-Ex bus is indeed may be a bottle neck, it is not the main weakness of GPU for HPC; the major weakness is SIMD architecture of GPU. Even modern GPU may have several SIMD modules it does not make it an effective MIMD machine like Intel/AMD multi-core CPU. The bottom line is that SIMD runs _SINGLE_ thread of instructions upon very wide array of numbers – it executes the _SINGLE_ instruction upon wide array of numbers. MIMD architecture runs _MULTIPLE_ threads of instructions and may execute concurrently/simultaneously _MULTIPLE_ instructions upon different array of numbers. There is SIMD-friendly class of algorithms however it is a quite small class comparatively with class of MIMD-friendly algorithms which may run SIMD algorithms equally well while opposite is not correct. CUDA is a very good example of misleading terminology; CUDA API talks about thousands of threads but in fact it is very few threads of instructions executing the same-single instruction upon very wide array of data.
Stefan