skavi

11 ms·

121.
by skavi 1y ago
I don’t see any B2s flying over Israel.
122.
by skavi 1y ago
Ah, that makes a lot of sense. Is this fine grained task scheduling related to CUDA Dynamic Parallelism at all? If not would you have a pointer on where to look? I suppose I could look through the code of this project, but I’d hate to have
123.
by skavi 1y ago
Fair enough. I should have clarified that “approximately the cost of a single kernel launch” is pretty much what I meant by “tiny”. What I was getting at was that a “megakernel” and a captured graph should have similar launch costs.
124.
by skavi 1y ago
Does anyone have an intuition on why this offers significant gains over CUDA Graphs?. The CPU launch cost of a graph is tiny which implies most of the work has been offloaded to the GPU's own scheduler. I'd expect that some I&#x2F
125.
by skavi 1y ago
My bad. My comment then should be seen as directed to the authors of the paper rather than the user I replied to. I am less concerned with the “stolen valor” than I am with the metaphor being somewhat nonsensical due to its reference to the
126.
by skavi 1y ago
No issue with your analysis, but it’s a bit odd to say the “idea of Performance and Efficiency cores” is from the Core i5. First, the i5 is a tier in Intel’s product stack and doesn’t refer to any specific generation. Core iX series process
127.
by skavi 1y ago
I would not expect that to be simpler. Working within that was designed as an integrated system sounds more annoying than putting together components that are intended to be used in DIY quadcopters.
128.
by skavi 1y ago
It’s fairly simple to build your own quadcopter.
129.
by skavi 1y ago
It doesn’t appear that Tartakovsky was affiliated at all with Pantheon. I would have been surprised if he was affiliated given the show’s fairly mundane animation. So what about Pantheon? It doesn’t seem relevant to this discussion apart fr
130.
by skavi 1y ago
I think the last sentence of the comment you’re replying to implies an awareness of that fact.
131.
by skavi 1y ago
I’d say it’s their schtick, but typically well worth reading through.
132.
by skavi 1y ago
> As shown in Figure 1, our megakernel outperforms vLLM and SGLang baselines (which use CUDA graphs and torch compilation) I’m surprised the reduction in overhead for graphs vs streams alone was so little. I feel I’ve observed larger gai
133.
by skavi 1y ago
are there stats which suggest "couchsurfer" and non "couchsurfer" homeless people have significantly different origins?
134.
by skavi 1y ago
D has a “BetterC” subset that depends only on a C runtime. Presumably there is some overlap with the “SafeD” subset. Curious if there are any D users here who try to stay within that intersection?
135.
by skavi 1y ago
i think stable coroutines [0] would be huge for rust. they would enable writing pure state machines in the form of straight line imperative code. currently they’re used in the implementation of async/await, but aren’t themselves expose
136.
by skavi 1y ago
allows trivially mocking effects for testing
137.
by skavi 2y ago
i have never used a lisp, but i’d assume due to its focus on macros, you are alternately the developer of a compiler and the user of that compiler. so making it easy on the compiler dev makes it easy on you.
138.
by skavi 2y ago
if i were to lean into cynicism, i might suggest this choice was meant to increase the effort required to reimplement cuda for other cards.
139.
by skavi 2y ago
i’m curious what advantage is derived from this existing independently of the PTX stack? i.e. why doesn’t cuTile produce PTX via a bundled compiler like Triton or (iirc) Warp? Even if there is some impedance mismatch, could PTX itself not h
140.
by skavi 2y ago
in certain situations, latency is an aspect of correctness. HFT and robotics come to mind.
141.
by skavi 2y ago
it is a bit more annoying: https://nix.dev/manual/nix/2.22/command-ref/new-cli/nix3-she...
142.
by skavi 2y ago
if i had to hazard a guess, that’s probably the exact response the design system is hoping for
143.
by skavi 2y ago
Yup, and you can apply the same technique to any language. The obvious example is bash with all the dependencies specified, but I’ve also hacked up quick single file rust scripts using nix shebangs. https://nixos.wiki/wiki&#
144.
by skavi 2y ago
Fittingly, I’ve heard [0] a similar transformation happened to Silicon Valley itself. Apparently it used to be a bunch of orchards. [0]: https://www.reddit.com/r/bayarea/comments/10672gn/til_before...
145.
by skavi 2y ago
they probably mean it’s going to be difficult early on to find standards compliant implementations. and then always very annoying to make sure all involved components support the correct feature set.
146.
by skavi 2y ago
This work eschews practicality. The API will be super annoying and no one will want to use it but me. For example, the IO god object is passed to main as a reference rather than by value. I don't think there's any risk of splittin
147.
by skavi 2y ago
> Do you also count heap allocations as side effects to be externalized? I’ve gone back and forth on heap. I admire Zig’s handling of allocation failures, but I’d also like to use Rust’s alloc lib. I’m currently leaning towards allocatio
148.
by skavi 2y ago
Really love the principles set out here. Love the concessions made in order to keep the language simpler. I've been idly considering building an alternative Rust std lib with a shape that reminds me of the capabilities system laid out
149.
by skavi 2y ago
linux and llvm contributions are where I would look.
150.
by skavi 2y ago
are the transforms one applies with this always correct (not necessarily optimal)? or is it possible to introduce a bug which didn’t exist in the original kernel.

More ›