An option to enable/disable generic opencl apps

[ahorek]

BOINC currently categorizes GPUs by manufacturer, with generic OpenCL devices added later as an additional layer on top of this structure.
Each new group requires updates to BOINC to support an additional vendor, this is how Intel and Apple GPUs were incorporated.

However, ARM GPUs like Mali or Adreno don’t fit into these categories, and continually adding new categories doesn’t seem like a sustainable approach. It also becomes difficult to separate these applications from standard CPU apps without server-side workarounds, similar to what PrimeGrid had to do for Apple GPUs before they were officially supported.

This mainly affects Windows on ARM and Android, though some Linux SoCs also include these GPUs.

Ideally, BOINC shouldn't depend on vendor detection at all, but would instead assess GPU capabilities directly. For instance, if a device supports OpenCL, the server could simply assign an OpenCL application without needing to check the vendor name. I understand that implementing this would require a significant redesign.

What do you think?

https://www.primegrid.com/forum_thread.php?id=14297&nowrap=true#188290
https://discord.com/channels/357493752434130944/357550381431259136/1498047947873714237

[brucemoreg]

From an interested user that runs BOINC: With the move to AI support in chip sets, is it still assumed that OpenCL support will be viable in the near future, e.g., 2-5 yrs? It sounds a lot more doable to broadly support OpenCL for BOINC, but what else is bubbling up for GPU/CPU vendor specific features in the near future, where BOINC, and project software developers would be forced to support these vendor specific hardware/drivers/libraries in order to get computation done? If this is the route, it seems like choices will have to be made as to what vendors will be supported, but what will that be based on, e.g., number of consumer units shipped, computation speed, ease of programming for vendor specific features, etc..?

[AenBleidd]

AI chips don't replace OpenCL: they are just made for different purposes (e.g. on the NPU devices that are now shipped with the new Intel chips are good for inference (and basically inference only), and can't be used for AI training.

[ahorek]

Most GPU projects rely on OpenCL for compatibility reasons - you only need to maintain a single application that can run across AMD, Intel, Nvidia, Adreno... without requiring significant extra effort.

• CUDA for Nvidia

CAL (ATI) - is dead
other APIs like HIP, Vulkan, OpenAPI, or Metal.... are rarely used.

[AenBleidd]

If you want your application to be performant and use as much capabilities of the every particular GPU - you have to create a separate application for every GPU type.

[ahorek]

An NPU can only run AI workloads, nothing else. Its main goal is improving power efficiency on laptops and mobile devices rather than maximizing raw performance. Just like video-playback accelerators. On desktop PCs, your CPU or GPU is your AI chip. You don't need an NPU.

[ahorek]

Fair point about optimal performance, but in practice, apps are rarely optimized for a single vendor (CUDA + NVIDIA being an exception)

[AenBleidd]

Generic OpenCL application are usually way too slow because their code doesn't incorporate the features of every particular GPU type.
So adding one more 'generic' specification will doubtly give any benefits.

[ahorek]

Adreno X2-90 on Snapdragon X2 Elite laptops has 7.58 TFLOPS 228.6GB/s @15w. This offers performance competitive with Apple GPUs or the GeForce GTX 1650

Adding support for both CPU and OpenCL applications (with the option to disable them in project preferences) would necessitate changes to the server-side code - something the project admins are usually not willing to do for maintenance reasons.

[AenBleidd]

New GPU vendors appear very rare: adding one new vendor once in several years is not a big problem (Adreno and Mali GPUs are quite long on the market).

[ahorek]

There's also https://en.wikipedia.org/wiki/Moore_Threads :)

The application can adjust its performance for a specific device model at runtime, but BOINC still needs to identify the GPU as OpenCL-capable and dispatch a GPU/OpenCL app.

That’s why a general OpenCL application concept makes more sense to me than adding vendor-specific checkboxes

[AenBleidd]

Project can create a single OpenCL application and just add it to different plan classes.