Do any of yall niggas fw cuda?

do any of yall niggas fw cuda?

what kinds of projects do you make with it?

Attached: cuda.jpg (406x246, 14K)

Other urls found in this thread:

github.com/enfiskutensykkel/ssd-gpu-dma
twitter.com/NSFWRedditGif

>what kinds of projects do you make with it?

I made this: github.com/enfiskutensykkel/ssd-gpu-dma

No one on Jow Forums knows any CUDA, m8.

Well, it's very good for fakeapp

>still using FakeApp
Why?

waifu 2x

waifu2x and machine learning

>... relies on hardware polling rather than being interrupt driven ...
Stopped reading there.

Extension that polls a server I have running locally that detects pornographic images and hides them

Interrupts are slow, man.

And burning CPU time thousands of times per second is fast and totally not a waste of power, right? I'm sure laptop users would LOVE it if every developer had that mindset. Who cares about alertable waits or interrupts, just fucking poll it over and over again until the user does what you want them to. Very efficient.

>And burning CPU time thousands of times per second is fast and totally not a waste of power, right?
1) You don't have to burn CPU, the point of NVMe is that you can poll for completions when you are ready.

2) The whole point of polling rather than being interrupt driven is that you can remove CPU from the data path entirely.

>I'm sure laptop users would LOVE it if every developer had that mindset.
It's not intended for laptops, seeing how you need a Quadro or Tesla workstation GPU to make it work and that I wrote it for sharing NVMe drives among multiple hosts in a PCIe cluster.

>Who cares about alertable waits or interrupts, just fucking poll it over and over again until the user does what you want them to. Very efficient.
It actually is, if you look at the benchmarking numbers. But you didn't, you stopped reading because you're opinionated and fail to understand what you are looking at.

This is legit.

>It actually is, if you look at the benchmarking numbers. But you didn't, you stopped reading because you're opinionated and fail to understand what you are looking at.
Subtle insult, I like it.

I did not intend it as an insult, just as a suggestion that maybe he should read more than a paragraph and a half before jumping to conclusions next time.

>do any of yall niggas fw kooda?

YOU CAN TALK HOT ON DA INNANET BOY, DASS DAT GOOFY SHIT WE AINT INNUH DAT BOY

SCUUUUUUUUUUUUUUUUUUUUUUM GAAAAAAAAAAAAAAAAAAAAAAAAAANG

Attached: si6ni9e.png (813x850, 399K)

>You

Attached: dunning-kreuger2.jpg (1110x816, 166K)

that chart only applies to brainlets

>that chart only applies to brainlets
Because their head is shaped like the curve?

Attached: brainlet2.jpg (645x729, 26K)

Didn't even notice before now, lol

Efficiency and performance is not the same thing. Something that takes up 50% CPU can not be said to be more efficient than something that takes up 25%, even if it performs better.

And again, the point of polling instead of interrupts is so you can move code to userspace and even run everything on the GPU, meaning that you don't have to involve the CPU at all. The point is not to be as energy efficient as possible, it is to achieve optimal performance (bandwidth and latency) when moving data from disk to GPU or vice versa.

Also, I doubt that invoking an interrupt routine is more efficient than simply checking a bit in memory, especially when you can move that memory around to be closest as possible to the compute node processing completions.

Attached: figure_3.png (1200x600, 81K)

Not the same guy that's being a dick to you. I'm genuinely curious.
Why not use a built in processor DMA engine to move the data and have that interrupt when done?

>Why not use a built in processor DMA engine move the data
I am, the SSD is doing DMA. I'm simply setting up command structures in memory and then triggering DMA with a simple register write.

When the NVMe disk is done, it posts a completion to memory as well.

>have that interrupt when done?
Interrupts are slow and impractical, because they trigger a context switch and forces you to write a driver in kernel space in order to register an interrupt handler. By simply polling on a memory address, I can detect very efficiently whether or not the disk is done.