FP64 support on GPU of 4th Gen Core (Haswell)

[rahul_garg]

Now that the Haswell NDA has been lifted, I wanted to know

1. Do any of the GPUs (Iris Pro 5200, Iris 5100, HD 5000, HD 4600) support cl_khr_fp64 (or perhaps some custom fp64 extension) in OpenCL?

2. If they do, what speed can we expect?

3. If they don't support fp64 in OpenCL, do they support fp64 in some other GPU compute API? Ivy Bridge used to provide limited fp64 support (add and mul, but no divide or FMA) in DirectCompute.

[ARNON_P_Intel]

Hi Rahul,

Intel Iris Graphics and Intel HD Graphics do not support the cl_khr_fp64 extension. 

Generally for fp64 usage, our majority of users are using Intel Xeon processors or Intel Xeon Phi coprocessors (vs low power devices like the intel processor graphics), Have you try to use fp64 OpenCL extension with Xeon or Xeon Phi?

Regards,

Arnon 

[rahul_garg]

Thanks for the info Aaron!

I am a PhD student working on compilers targeting multi and many-core systems.  For GPUs, I generate OpenCL. For CPUs, I have a different backend. While I would love to have access to a Xeon Phi, I am limited by the lab's hardware budget and so don't have one.  Regular desktop processors are obviously much cheaper to buy. Also, most people who are interested in using my compiler are also mostly using regular desktops and notebooks.

As for HD graphics, I do think they can be used as nice co-processors along with the Haswell or Ivy Bridge CPU cores.  With Haswell GT3e, I think the HD graphics GPU now has sufficient processing power and bandwidth. They are very suitable for the case where both the CPU and the GPU cores work on different parts of the problem, rather than a pure offload model. They benefit from not having to do data transfers in many cases.  However, the majority of the use-cases of my compiler are in scientific computing where people tend to use fp64 a lot. Thus, it would have been great if we could use fp64 on HD graphics through OpenCL.

I have a feature request that perhaps Intel can release limited fp64 support through a vendor extensions (something like cl_intel_fp64) where a subset of cl_khr_fp64 can be supported. AMD used to do something similar through their extension cl_amd_fp64.

[MSimm2]

The A10-5800K does support double precision. An A10-5800K's GPU does about 40Gfps double precision.
http://forum.beyond3d.com/showthread.php?t=63628 (Find "A10-5800K GPU")
The Kaveri APU will use gcn which supports 1/4 fp32 rate, so fingers crossed we might see ~250 Gflops double precision by the end of the year.

[rahul_garg]

Relatedly, AMD's Kabini supports cl_khr_fp64 on the GCN based GPU though at a fairly slow rate. I have verified this with reliable sources, and some OpenCL info listings on CLBenchmark also confirm this.

[MSimm2]

moozoo wrote:
The Kaveri APU will use gcn which supports 1/4 fp32 rate, so fingers crossed we might see ~250 Gflops double precision by the end of the year.

It's more likely the rate will be 1/16 like Kabini. Although they could make one that was 1/4.

so Kaveri will probably be about 62.5 Gflops double precision.