The Intel Haswell chip contains an integrated GPU that delivers significantly better OpenCL performance than an NVIDIA GeForce GT 650M – exceeding 800 GF/s peak performance. Allan MacKinnon at PixelIO has been investigating the OpenCL performance of this device and has been finding a plethora of on-gpu registers but also that the GPU appears to be both power and thermally capped. (He can watched the IGP clock bounce up and down along with the wattage as reported by GPU-Z.) Allan also found that the Intel OpenCL SDK does not support GPU debugging. His remark is that he is “stunned at the size of the GRF (general register file) per EU thread and the amount of SLM (shared local memory)”. The Intel Iris Pro 5200 has 4 “slices” that each have 70K 32-bit registers and 64KB of shared memory.
CPUbench has published the Iris 5200 OpenCL specifications. Note that the floating-point is not IEEE compliant.
The only gotchas Allan is seeing so far are:
- No debugger.
- There is no free profiler, but one can use VTune. The programmer can get some free utilization graphs.
- Documentation is pretty bad, examples are few.
- Intel has plenty of low-level opensource docs for their Gen architecture. There are chapters that describe the instruction set and SIMD architecture. It appears to be a very compact set of instructions.
- Be careful of your install asl the new 2014 Beta seems to conflict w/the NVIDIA driver. Try reinstalling the Intel HD Graphics driver if you get a compilation problems due to a DLL issue.
Thanks to Hiroshige Goto for the lovely diagram of the Iris 5200. (Modified to be the featured image for this article by techEnablement.com)