I think we should do what we did in the 1970s, which worked much better than what we are doing today.
-
-
Replying to @Jonathan_Blow
But that's not an option if I want to program GPUs today. I want a CPU-style toolchain straight to machine code for major GPUs cross-platform too. But that's not an option unless I form a decade-long conspiracy of friends to become execs at Apple, MSFT, AMD, Intel, Nvidia...
2 replies 0 retweets 18 likes -
Replying to @trishume
Somebody has to be the adult in the room or things will continue to get worse.
1 reply 0 retweets 3 likes -
Replying to @Jonathan_Blow
Like which concrete people do you want to do what concrete action? Everyone should stop working on GPU stuff until execs pay attention to the strike? Just everyone complain on Twitter? Anyone who wants to use a GPU should try to become a MSFT GPU exec instead?
2 replies 0 retweets 11 likes -
Replying to @trishume
Know what’s extremely off-putting? Learned helplessness. Don’t be helpless.
1 reply 0 retweets 9 likes -
Replying to @Jonathan_Blow
I already mentioned one of the two plans to a new driver equilibrium I think have a smidge of hope, the other being try to piggy-back off of WebGPU momentum and then work down unifying drivers from there. I'm interested in how you think we might get there. What's your plan?
2 replies 0 retweets 6 likes -
Replying to @trishume
There are many possibilities. One is, build the thing. Build a pretend instruction set that is at a low level such that it would be targetable by any high-level language. Do a pass for LLVM that outputs to this. Then do all the gross garbage that up-translates this to
3 replies 0 retweets 7 likes -
Replying to @Jonathan_Blow @trishume
all the various shading languages. In the short term this is just another sucky shader language, but at least now it's at the right level of abstraction, so someone can do driver-level support in Linux, for example, then continue from there.
3 replies 0 retweets 6 likes -
Replying to @Jonathan_Blow @trishume
Intel are doing close to this. It's hard to see through the fog of uncertainty, but OneAPI is groping toward that goal. It's a bunch of tools and languages and so on with interop as the focus. Of course it's Yet More Standards in the short term. We'll see where it leads.
2 replies 0 retweets 3 likes -
An obvious tensions is between what Jon wants which is a super low level "assembly" that works on every GPU, and Intel's 50 years of experience doing this on CPUs where it's a gigantic pain in the arse.
4 replies 0 retweets 4 likes
At least for me, I don't want low level assembly that works on every GPU. I want one ISA per vendor. So AMD, Intel, and nVidia can all have their own, so long as they accept the same binaries across all platforms if the card is the same. This would be _less work_ than now.
-
-
Less work for who?
2 replies 0 retweets 3 likes -
Replying to @tom_forsyth @cmuratori and
You know, you're both right. But it's funny how the exactly same people who complain about not having simple isas are those who complain about not having hyper-complex isas.
1 reply 0 retweets 0 likes - Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.