[wangmianzhi]

Intel just announced its socket-based Xeon Phi: the Knight Landing (KNL).
It seems that the heterogeneous computing models: CUDA, OpenCL, and OpenACC, would quickly phase-out from the CFD software development domain.
What's your opinion?

[wyldckat]

I wanted to give my 2 cents on this, since it's a favourite subject of mine. And sorry for the near rambling I go on below...


As with most new technologies, they are very expensive, either because the hardware is very expensive, or because the cost of re-adapting a working infrastructure takes a lot of time. This makes adoption take at least 5 to 10 years for things to really pick-up, therefore the current technologies can still make their money's worth in the meantime.

Then there is the other factor: the lessons learned with the current technologies will open way to the new hardware technologies. In other words, at least for OpenCL and OpenACC, they can still be reused for the new hardware.

Either way, here's a simple/similar recent situation: AMD released the hardware and HSA standard a few years ago. I have yet to see any work done in Computational Fluid Dynamics that takes advantage of this hardware, even though the HSA enabled processors have essentially large vector processors embedded along with the conventional x86_64 architecture. Essentially they sport oversized Vector Floating Processor Units and I can't find any work done with it, not even data stating if there is a memory bottleneck or not for this kind of calculations.
The "4 x86 + 8 GPU stream cores = 12 cores" that the top tier AMD CPUs provide and advertise, should be making for some serious damage on Intel's and NVidia's business model by now. But it's not, at least I haven't seen it make any damage.

Then there is the other detail: CFD and GPUs only do their best work together, when most of the workload is on the GPU's side... but most CFD isn't just vector-like operations.

Knight's Landing has the home advantage, namely being able to use nearly everything that is meant for x86_64, but the multi-threaded side of it won't be worth much for CFD. The über fast memory access it has near the cores can help a lot, but this is where things start to get more similar/identical to the GPU implementations: the best performances will only hold up if you use only the nearby memory accesses, at least the ones that use multi-threaded access. As soon as you need to swap out a chunk of memory between the cache and the main memory, things will start to slow down by a considerable amount. This is where the current OpenCL and OpenACC technologies can help, because this is the exactly the same kind of use scenarios that the similar implementations on GPUs face every day.

Another favourite example of mine on this kind of topic: the Thunderbolt technology has the potential to replace almost all other existing data interface technologies: 5Gbps USB, 6Gbps SATA, 1/10Gbps Ethernet and potentially even 100Gbps Infiniband. Will it ever replace any of them: I doubt it. The Thunderbolt technology is expensive and so far implementations for it have to be dedicated. In the meantime, it seems that now whenever Thunderbolt bumps 2x the speed, so do the other technologies.


edit: What I mean by "multi-threaded side of Knight's Landing" is the hyperthread feature that each core has got.

[AvtechScientific]

An important advantage of OpenCL is its portability. Once written your code can be deployed on different accelerators - GPU/FPGA/DSP/etc.. I doubt Intel will make it possible.

We have recently open-sourced our accelerated multiphysics software - Advanced Simulation Library. It implements i.a. Lattice Boltzmann Method, provides easy to learn C++ API, has an OpenCL-based internal engine (i.e. runs on CPU/GPU/FPGA/DSP/etc.), and demonstrates remarkable performance even on a single GPU, e.g.:

Multicomponent flow:
computed on NVIDIA Tesla K80 12 GB GDDR5 with host CPU Intel Ivy Bridge E5-2690 v2 3GHz
10’517’227 points - 390sec.
75’293’400 points - 2486sec.

more multicomponent flow benchmarks...