the tiny corp
@__tinygrad__
Followers
57K
Following
460
Media
333
Statuses
3K
We make tinygrad and sell tinybox, the best perf/$ AI computer. Our mission is to commoditize the petaflop.
San Diego
Joined June 2023
In uncertain times, tinygrad always has bounties for you to solve. Over $50k paid out to date. And the only path to a job here, all hiring will look like this soon. bounties dot tinygrad dot org
10
29
467
This is tinygrad's description of the tensor cores of all the major GPUs. No per GPU dialects, just a spec for what they each are.
17
42
618
tinygrad is a bet against irreducible complexity. When you look at Triton and similar, so many things are GPU specific. The task in reverse engineering is ignoring the labels, what is that really? Forget the abstractions, search for the perfect program that runs on the hardware.
7
13
463
America has once again put up roadblocks to us getting a visa. And I'm sure it's not just us. Every decision made like this slowly chips away at America's lead in AI. Most Americans have never lived in a country that wasn't rich, but this is the road to that.
37
20
730
All of tinygrad, including non existent dependencies, fits in the context window. Seed AI type vibes.
6
8
383
RT @xingyu_liao: 之前在安装 tinygrad 的时候，选择了错误的 cuda 版本，导致一直 generated code ptx 版本有问题，不断反复安装都没有解决，让 claude-code 接手之后，他找了到 cache 成功解决了问题 https://….
0
2
0
RT @SharingPsyche: Got my company laptop and the first thing I did after convincing IT to give me WSL. ? Install cuda and ran beautiful_mni….
0
1
0
RT @UThree271828: tinygradのコードとかドキュメントとか参考にしながらやってるんだけど、これマジですごいな。読めば読むほど感動が生えてくる。.洗練された設計は読んでいて楽しいな。.
0
1
0
Our MI350X machines are here, thanks @AMD! They are just two racks down from their MI300X friends.
31
52
2K
end to end will win neural network frameworks just like it is winning self driving cars. if you have a Conv3D guy, you are going to lose. he's the new cone guy.
10
9
319
Wrote a blog post about if tinygrad can win or not. It's not a foregone conclusion, ChatGPT gives it a 25% chance. But I'll bet on it.
10
11
218
tinybox in action!.
10
10
279
. or maybe we should just paste the unoptimized kernel into ChatGPT and vibecode a drop-in fast replacement!
3
0
61
A key aspect of beating NVIDIA on MLPerf Llama 405B with AMD will be great flash attention. This is autogenerated basic flash attention. Now we're upgrading the optimizer to make it fast. The Halide PhD thesis is a great way to understand the search space.
8
21
290
