The Linux Kernel Likely To See A Hardware Accelerator Subsystem

[phoronix]

Phoronix: The Linux Kernel Likely To See A Hardware Accelerator Subsystem

Given the increasing rise of hardware accelerators for compute offloading of particular tasks especially now around deep/machine learning with more chips coming to market, the Linux kernel will likely soon see the introduction of a formal subsystem for these different accelerator drivers...

http://www.phoronix.com/scan.php?pag...ccel-Subsystem

[Pickup]

This question probably will make me look very dumb:
What exactly does this Goya thing do?
Does it speed up PCs by just plugging it in and run a kernel module?
Will it make GPUs unnecessary/obsolete and replace CPUs?

[DrYak]

What exactly does this Goya thing do?

It's dedicated special hardware that is good at running those specific operations that happens when running a neural net.
(in practice : good at repeating tons of low-precision floating point arithmetic over and over)

Does it speed up PCs by just plugging it in and run a kernel module?

Nope. Not at a all. It's not a general purpose CPU.
When you plug in the card and load the driver you'll get a device to which you can send tons of AI operations.

Just like a GPU plugged and its driver loaded, gives a "graphic card" device to which OpenCL can talk and you can use openCL to run some scientific computations.
This card and its drivers will give you an interface to which, e.g. Tensorflow can talk to and you can use Tensorflow to run some neural nets.

Will it make GPUs unnecessary/obsolete and replace CPUs?

In big data centers, for the specific purpose of running tensorflow for neural net? Probably yes (at least that's the hope of the company).
You get dedicated hardware that's a bit faster at running neural nets than GPU.
and any form of acceleration (GPGPU, AI accelerators, etc.) is better than classic CPU for running neural nets.

In everyday life on your workstation ? probably nope.
(it's going to be a bit more expensive than a consumer GPU, and it's single purpose only - e.g. lacks 32 and 64 bit precision floats and doesn't have the extra ciruictry to do graphics)

[coder]

This question probably will make me look very dumb:
What exactly does this Goya thing do?
Does it speed up PCs by just plugging it in and run a kernel module?
Will it make GPUs unnecessary/obsolete and replace CPUs?

He did link to this, in the article.

https://www.phoronix.com/scan.php?pa...ux-Habana-Goya

It's probably not useful for much beyond AI, due to these sorts of things only supporting rather low-precision arithmetic.

For a more general-purpose accelerator, see FPGA cards.

[coder]

...and I can't hear the name Goya without thinking of this:

https://en.wikipedia.org/wiki/Saturn_Devouring_His_Son


And that's just one of his more uplifting works:

https://en.wikipedia.org/wiki/Black_Paintings


It'd be cute if they referred to these in the names of architectural components of the chip/board.