Intel® HD Graphics

[tuxutku]

CPU:Intel® Core™ i3-380M Processor
GPU:Intel® HD Graphics
(GPU is inside CPU (İntegrated))
OS:Ubuntu 16.10 64-bit
Output of $ sh -c "cd /usr/bin ; FAHClient --lspci"
VendorID:DeviceID:Vendor Name:Description
0x8086:0x0044:Intel Corporation:
0x8086:0x0046:Intel Corporation:
0x8086:0x3b64:Intel Corporation:
0x8086:0x3b3c:Intel Corporation:
0x8086:0x3b56:Intel Corporation:
0x8086:0x3b42:Intel Corporation:
0x8086:0x3b44:Intel Corporation:
0x8086:0x3b4a:Intel Corporation:
0x8086:0x3b34:Intel Corporation:
0x8086:0x2448:Intel Corporation:
0x8086:0x3b09:Intel Corporation:
0x8086:0x3b29:Intel Corporation:
0x8086:0x3b30:Intel Corporation:
0x8086:0x3b32:Intel Corporation:
0x1969:0x2060:Qualcomm Atheros:
0x168c:0x002b:Qualcomm Atheros:
0x8086:0x2c62:Intel Corporation:
0x8086:0x2d01:Intel Corporation:
0x8086:0x2d10:Intel Corporation:
0x8086:0x2d11:Intel Corporation:
0x8086:0x2d12:Intel Corporation:
0x8086:0x2d13:Intel Corporation:

[Joe_H]

Welcome to the folding support forum.

If you are requesting that your Intel iGPU be added to the whitelist, that is not currently going to happen. The only GPU's currently supported are from AMD and nVidia, the integrated GPU 's from Intel are not supported. That may change at some time in the future if the processing power increases and the OpenCL support provided by the drivers supplied by Intel becomes sufficient to support folding.

[darkbasic]

What about Beignet? It provides decent OpenCL 2 support with decent performance. Iris PRO cards with eDRAM should be fast enough to be worth it and it the future it can only improve.
If the Windows driver is not good enough then fuck Windows and simply support Linux.

[bruce]

Integrated GPUs, whether from Intel or AMD or NVidia or others have consistently produced marginal GFLOPs (mostly due to a limited number of parallel GPU threads) while discrete GPUs continue to evolve into more powerful resources with hundreds/thousands as many parallel compute threads. Adding support may seem like a good idea but when the results are marginal, such a project represents a significant developmental expense that has never reached the level of a net benefit for FAH.

Recently, FAH support was added for AVX which does a good job of utilzing the parallel compute threads of the CPU. I'd encourage you to utilize that capability instead (and potentially with the concurrent use of a discrete GPU, if applicable).

[darkbasic]

I fold with an Ivy Bridge 3770K@5Ghz and an AMD HD7950@1200/6000 (~200K PPD total). Sometimes I even wonder if my system is still worth the current it consumes considering a single 1080 Ti folds 1500K PPD (7x faster), so I have no personal interest for Intel GPUs.
Still, do you have any OpenCL number to share for this CPU? I've an hard time believing that its integrated GPU will be slower than the CPU. Releasing an experimental Intel OpenCL client would be very useful to start optimizing the graphic stack, especially Beignet.

[bruce]

I've an hard time believing that its integrated GPU will be slower than the CPU. Releasing an experimental Intel OpenCL client would be very useful to start optimizing the graphic stack, especially Beignet.

You're asking the wrong question. It's about both (A) the development costs of a new capability and (B) having a class of projects to assign to that hardware class.

CPU have a range of capabilities and the deadline for projects assigned to them needs to be long enough for the slowest member of that hardware class to successfully complete their assignment. The same statement applies to GPUs but with different numbers.

There are no active GPU projects with deadlines long enough for the typical iGPU so those projects are assigned to AVX or SSE based hardware. Allowing longer deadlines for slow GPUs is detrimental to FAH's overall performance.

[darkbasic]

(A) the development costs of a new capability

Which development? You mean the whitelisting, the testing and some minor modifications (if needed)?

(B) having a class of projects to assign to that hardware class.

There are no active GPU projects with deadlines long enough for the typical iGPU

Who said you have to whitelist all Intel GPUs ever made? Start by whitelisting the Iris Plus 650 (which is the fastest one available today) and then whitelist every new iGPU released from now on. Iris Plus 650 is fast enough to finish before the deadline.

Also, don't you have an Android OpenCL client? Those GPUs are so damn slow that even an ancient Sandy Bridge iGPU will eat them for breakfast. In the worst case scenario you could assing iGPUs to the same project, but it wouldn't be a good idea IMHO. It makes much more sense to start whitelisting a modern Iris Plus iGPU, assign it to normal GPU projects and then whitelist further hardware when it gets released.

[Nathan_P]

The android client is not being developed by Pande group, so its not their resource being used.

Development may entail just some testing and minor tweaks, it may require more than that. either way the costs associated with getting the programming team to do the work may not be justified, spending x to achieve y is no good if you can spend x and achieve more than y. Current focus is on a new gpu core and a new cpu core utilising avx. In addition the same team also has to keep updating the work server and assignment server code, and its a small team trying to do all of this.

[Joe_H]

Starting from the last, no, there is not an Android OpenCL client. The Android client created by Sony in collaboration with PG runs on the ARM CPU, not the GPU.

As for just whitelisting the current Intel GPU's and later more powerful ones, that is not enough. A separate GPU folding core based on OpenMM using OpenCL would also need to be coded, tested and then released. That is just like there being separate AMD and nVidia GPU folding cores.

As for the open source implementation of OpenCL for the Intel GPU - Beignet, has anyone even got it far enough that the FAHBench program will run on it? If so, what are the performance figures.

[darkbasic]

As for just whitelisting the current Intel GPU's and later more powerful ones, that is not enough. A separate GPU folding core based on OpenMM using OpenCL would also need to be coded, tested and then released. That is just like there being separate AMD and nVidia GPU folding cores.

Doesn't the Nvidia core run on top of CUDA instead of OpenCL? I may be a bit outdated in that regard

As for the open source implementation of OpenCL for the Intel GPU - Beignet, has anyone even got it far enough that the FAHBench program will run on it? If so, what are the performance figures.

I could test it on my Broadwell "HD 5500" iGPU if you want, but I already know it's not fast enough to be worth it because I can barely reach the same performance of the CPU when using OpenCL with Darktable. Fortunately Iris PRO/Plus (especially modern ones, like on Kaby Lake) are *way* faster than mine.

[Joe_H]

No, the current nVidia cores have been running on OpenCL since Core_17 was developed and released 4 years ago. At the time the the nVidia CUDA tools did not include JIT which was needed for folding.

The reason I ask if Beignet worked to run FAHBench is that is where many prior attempts to test the Intel iGPU's has failed. Trouble folding and running FAHBench AMD GPU's has also been reported in connection with the open source OpenCL implementation as compared to using the AMD provided one.

[darkbasic]

Trouble folding and running FAHBench AMD GPU's has also been reported in connection with the open source OpenCL implementation as compared to using the AMD provided one.

AMD's current OpenCL implementation is a joke, unfortunately
Lately there is one guy actively working on it, so there could be some chances to get it working now (there are also several out of tree patches floating on the mailing list which fixed several applications, especially the ones requiring image support which was notably missing), but it's probably better to just wait for AMD to release the source code of the new OpenCL stack.

Intel is another story: Beignet was even in a worse shape, but lately it improved alot and recently we even got OpenCL2 support. I cannot assure you that it will work because I still didn't test it, but I will

[darkbasic]

Here it is:

Code: Select all

$ ./Built/bin/FAHBench-cmd
FAHBench Simulation
-------------------
Plugin directory: "/home/niko/devel/fahbench/Built/lib/openmm"
Work unit: dhfr
WU Name: Dihydrofolate reductase
WU Description: A common system for benchmarking molecular dynamics
System XML: /home/niko/devel/fahbench/Built/share/fahbench/workunits/dhfr/system.xml
Integrator XML: /home/niko/devel/fahbench/Built/share/fahbench/workunits/dhfr/integrator.xml
State XML: /home/niko/devel/fahbench/Built/share/fahbench/workunits/dhfr/state.xml
Step chunk: 40
Device ID 0; Platform OpenCL; Platform ID 0
Run length: 60s

Loading plugins from plugin directory
Number of registered plugins: 3
Deserializing input files: system
Deserializing input files: state
Deserializing input files: integrator
Creating context (may take several minutes)
Checking accuracy against reference code
Creating reference context (may take several minutes)
Comparing forces and energy
Starting Benchmark
                                                                               
Benchmarking finished
Final score:    3.8875
Scaled score:   3.8875 (23558 atoms)

It works flawlessly, but as I said my Broadwell HD 5500 iGPU sucks. Try Iris Plus 650 and you will be surprised.

Edit: I also did the CPU test, I got 0.292496 on the same system.

[Joe_H]

The Iris Plus 650 is only about 2 1/2 times as powerful as your HD 5500 when measured in GFLOPS, that tends to be indicative of folding power. So a scaled score of about 10, compared to a scaled score of around 110 for a GTX 1080.

[darkbasic]

So a scaled score of about 10

I agree on your estimate, that's what I would expect too. Don't you think it is enough to reach the deadline? Even an AMD HD7950 like mine (which is a modern, high end GCN 1.0 card) should score about 30 points at stock frequencies and it's far from being the slowest one: there are plenty of low and mid range cards. Also we are talking about an OpenCL stack which can still improve a lot, performance optimizations only started recently in Beignet because they worked on conformance first: a score of 10 could easily become a score of 15.
Which kind of modifications do you need in order to make the core suitable for Intel? OpenCL should be suited to run on every kind of GPU, you can't possibly rewrite everything from scratch.