Intel Looking To Finally Upstream Linux Driver For Their Gaussian & Neural Accelerator

[phoronix]

Phoronix: Intel Looking To Finally Upstream Linux Driver For Their Gaussian & Neural Accelerator

Found with mobile Intel CPUs across Tiger Lake, Ice Lake, and even Cannon Lake has been the Intel GNA accelerator. This Gaussian and Neural Accelerator is also found with Intel Gemini Lake processors and various development kits. The Intel GNA has been backed by an out-of-tree Linux driver while now the company is finally working to upstream their GNA support in the Linux kernel...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-GNA-Upstream-V1

[chuckula]

On a standard TGL SoC you have three different neural network acceleration architectures. There's VNNI, the GPU, and GNA. Incidentally, Tiger Lake has GNA 2.0 that is an upgrade from IceLake but it's the same basic device.

[kcrudup]

That's all fine and good Intel, but what about the freakin' IPU4 driver so laptops like mine (XPS 13 2-in-1) can get a working camera?!

[RedEyed]

Didn't get it. Why do we need kernel driver for that?
I guess, it a special hardware device, isn't it?

[RedEyed]

If so, just wonder, will MKL use it?

[kcrudup]

Why do we need kernel driver for that?

It's likely got DMA and IRQs and likely to be a client of the multimedia subsystem.

[Guest]

On a standard TGL SoC you have three different neural network acceleration architectures. There's VNNI, the GPU, and GNA. Incidentally, Tiger Lake has GNA 2.0 that is an upgrade from IceLake but it's the same basic device.

Well, VNNI is a CPU instruction. GPU is just a catch all for highly parallel work. Neither of these are necessarily the best or fastest for neural network workloads, they're just more generic and flexible.

GNA is more specific, and can guarantee high performance in real time..