> Imagine connecting a graphics card to a display and a power source and running diagnostics before installing it or running it standalone later when troubleshooting. Perhaps the card could have a compact Linux distro onboard, mused the iconic Doom developer.
You're quoting a blogger's interpretation - here's the tweet quote - "put a tiny linux system running busybox on your command processor". Command processor is the part of a GPU that schedules the hundreds of smaller cores.
Tangentially related, but GPUs with embedded m2 drives mentioned in the article (IIUC they effectively pass m2 drives through to the system by using unused PCI lanes) are a very interesting idea for a low volume transcoding machine.
It's a clever idea but unfortunately limited by the fact that it requires the motherboard to support PCIe slot bifurcation, which isn't very common on consumer boards. I assume the target market is OEM integrators who can go out of their way to pair it with a suitable motherboard.
My dirt cheap Asrock A320m-ITX and my mid-level Gigabyte Aorus Elite x570 support bifurcation, did the feature become much less common in the last couple years?
Some boards will only bifurcate the x16 slots into x8/x8 mode, which is sufficient for an x8 GPU with one SSD attached to the spare lanes, but that still leaves 4 lanes unutilized. Adding two SSDs to fill up the bus requires support for x8/x4/x4 mode, and an x16 carrier for four SSDs requires support for x4/x4/x4/x4 mode.
GPUs are slow for single threaded things. Often 20x+ slower than your CPU. They get all of their speed by massive parallelism and some specialized hardware (like samplers). If your problem can't be easily parallelized then they aren't a solution.
Honestly, what an absolutely minimal effort blog post. It would've been so much cooler if the author actually interviewed John and fleshed out or explored the idea a bit more.
I mean Larrabee did already back in 2008-ish, but with BSD. That was a real advantage of the reuse of some x86 cores.
On the flip side, until somewhat recently, NVIDIA GPUs didn't have the kind of hardware protections you'd want to build a modern multitenant OS (e.g., hardware page table and TLB support).
> I still think today’s GPUs should be able to operate without host CPUs if they have a private link. Chains of accelerators are a legitimate use, but it would just be fun if GPUs made their own video signal with diagnostic information when you apply power outside of a host system. You could go farther and put a tiny linux system running busybox on your command processor, and backchannel keyboard input through the display port if you don’t have a USB port.
Being able to run some graphics diagnostics without a host CPU would be useful.
However, if we're talking about putting a general-purpose processor/RAM onto a GPU that has enough power to run Linux, at that point, you've essentially created a game console.
May as well go all the way and add a display and battery too, for portability. I'm sure that would be a hit! /s
reply