HD Graphics 4000: Native Compute Support
Ivy Bridge also includes the groundwork necessary to support OpenCL and DirectCompute 5.0.
But wait, didn’t Intel already release a driver for Sandy Bridge that enabled compute functionality? It did. HD Graphics 3000/2000 doesn’t actually support these APIs, though. They’re emulated and executed on the host processor, which is why compute workloads peg Sandy Bridge-based chips at 100% utilization.
HD Graphics 4000, on the other hand, supports FP32/FP64 under DirectCompute and FP32 in OpenCL. Intel currently lacks Khronos certification for ARB_gpu_shader_fp64, so it’s not enabled.
It’s immediately apparent how much faster native FP32 runs on HD Graphics 4000 compared to Sandy Bridge’s IA cores emulating OpenCL support. Because Sandra has to emulate FP64 through FP32, native double-precision performance looks a lot lower. However, that’s still GPU-only.
With Nvidia’s Kepler architecture capping FP64 at 1/24 single-precision performance, it’ll be interesting to see how HD Graphics 4000 compares to derivative discrete GPUs from Nvidia (particularly if Intel adds its own OpenCL FP64 extension).
Again, HD Graphics 3000 lacks native OpenCL support, so it’s missing from this chart. However, we see that Intel’s quad-core processors emulate OpenCL very well (albeit at full processor load and higher power consumption).
The discrete Radeon HD 6570 trails both Ivy and Sandy Bridge architectures, and is turn followed by HD Graphics 4000 with native OpenCL support. But instead of the IA cores completely tied up, CPU utilization sits at 0%, while power hovers 50 W lower. As far as performance per watt goes, that’s pretty darned impressive.