Hacker News Reader

Comment by dragontamer | Hacker News Reader

I heard about this mysterious "Solid State GPU" around the internet, so I looked around for how to actually use it. Unfortunately, the $7000ish price tag is a bit too high for me to personally invest into, but the entire concept of making files on the GPU directly is intriguing. I'm curious if anyone has done anything with this hardware yet.

The manual is relatively straight forward, it shows 3 vendor-API extensions to OpenCL, OpenGL, and DirectX which allows the programmer to create files on the GPU. AMD markets this GPU as being able to play 8k video or perhaps do some kind of huge memory computation (without ever being bottlenecked by the PCIe bus).

EDIT: There's a second, shorter "device driver" manual which shows how the card works. Apparently, there are 4 NVMe SSDs on board the GPU.

http://pro.radeon.com/_downloads/SSG_Driver_Installer_Guide....

View on HN

Intel's Larrabee actually ran FreeBSD with a proper filesystem and everything. You could telnet into your GPU, install your own API implementation, pretty cool stuff: http://tomforsyth1000.github.io/blog.wiki.html#%5B%5BWhy%20d...

Interesting concept, downside is it is only available for Windows 10.

Well that's one way to destroy any chances of this gpu being useful for machine learning.

Why? One of my DL rigs runs on Win 10 reliably. Radeons anyway won't be a good candidate for DL, though for mining are best.

I'm not sure the 2TB of onboard NVMe would help machine learning tasks. Most machine learning tasks are already well designed for the ~16GB limit on standard GPUs.

The product page suggests that this sort of GPU can handle streaming 8k video. So maybe the card was designed for OpenCL accelerated video editing or some other 3d-task (maybe 3d rendering) with very large texture data.

You're absolutely correct on the video editing front.

The ability to work on full res 4k/8k video (in realtime) without generating down-scaled proxies will be a huge time-saver for editors in addition to saving money during post.

A lot of deep learning workloads aren't necessarily bounded by things like disk transfers, though, but stuff like bandwidth to the card itself. You probably have enough RAM to just pre-warm your disk cache anyway if you have exceedingly large training samples. This is more useful for things like massive CAD models, if I had to guess.

On Nvidia if you want this, I think you can basically achieve it using GPUDirect anyway if you really want to bypass CPU transfers and go straight from SSD to GPU, via PCIe. I guess it depends exactly on how this card is connected to the disk.

Having it on-card is pretty convenient, though. GPUDirect has practical limitations on what boards/systems it will work with.

Reply on news.ycombinator.com