>>62286551

This is old news, we still don't know anything about the feature or if it will feature prominently in consumer Volta, or if it has any effect on non-compute workloads at all.

O boi

AMD's scheduler is apparently several generations behind nvidia's.

>Vega touts warp packing as a new exciting feature
>It doesn't even work
>Nvidia is already capable of scheduling individual threads

I don't understand this pic.

Why are curved arrows better than straight arrows?

Because NVidia Graphworks(tm).

Nice, any news on the new Tegras coming out? Is thw only rjibg kewping me from getting a Switch, underpowered garbage

The GTX2070 won't need external power connectors

well yeah the card itself its more efficient since they have shifted everything on the cpu LOL

their version of async the MPS is literally being emulated on the cpu along with the warp SC
i will love when async heavy games come to see the total system power draw LOL

Previously, you were bound to warp convergence. With volta, there are cooperative groups, meaning that you can schedule individual threads without warp divergence.

NOOOOOOOOOOOOOOO

DELETE THIS

IT'S NOT FAIR, AMD WAS SUPPOSED TO FUCKING WIN

>do a half-node shrink
>throw moar ALUs
>claim 50 more efficiency *per SM*
Whoa.

> mfw novideofags like getting anal from propietary technologies

Anyway, 17.320 batch for Vega when?

Libre hardware GPU when?

now, let Sup Forums design it, send it to god raja

Never since even basic shaders are patented.

Done

I can make the logo

That picture is extremely out of context.

I'll try to explain.

In earlier generations of GPUs, threads executed in lock step with each other. So each thread in a group (called a "warp") was always at the same place.

Issue with that was that some times you have Conditionals like so:
if (test) {
code block A;
} else {
code block B;
}


In those case, threads that passed the test would execute A and threads that failed would stop execution and wait for the other threads to finish. Then it flips, first group halts and the second groups executes. Once the condition block ends, all threads would continue executing.

Here's the problem, if your code was doing something that was communicating with other threads, and you were waiting on a reply from another thread that was halted until your thread completed, the threads would become deadlock unable to proceed.

This means algorithms that required fine grain communication with each other are not possible on older GPUs.

Volta's new approach allows for halted threads to continue execution by switching between executing threads and halted threads at regular intervals, and eventually forcing a convergence. This means deadlocks are not possible, and different more efficient algorithms can be used on the hardware.


This is more applicable to CUDA development instead for graphics though.

Thank you

>they have shifted everything on the cpu LOL
Is that bad?
With 6 and 8 core CPUs being affordable, it makes perfect sense to off load inefficient hardware scheduling to the CPU.

So this mean Volta won't even have a partial hardware scheduler?

Thaks sir.