Manish Gupta

MTS @magicailabs, Compiler Lover, Fortune Cookie Writer

[Manish Gupta @BigManniM9]

Magic: The Pledge, The Turn, The Prestige

[Magic @magicailabs]

Frontier pretraining is said to be a big-lab-only game. We don’t have 100k chips yet, so there’s only one way: algorithmic efficiency. Our new recipe matches DeepSeek V4 Pro’s pretrain using 50x less compute – that’s roughly half the FLOPs used for GPT3, or ~$0.5M on GB200.

https://t.co/uBs5o5Arsu

[Elon Musk @elonmusk]

I love you

[Dwarkesh Patel @dwarkesh_sp]

Super impressive. Congrats!

And maybe suggests that during RSI, the automated AI researchers will be less bottlenecked by compute than we might naively think.

[Eric Steinberger @EricSteinb]

The world will do its best to convince you that you can't do big things, but you can. I'm grateful for our team. Unmatched tenacity.

[Manish Gupta @BigManniM9]

@gaunernst Issues like this keeps me worried moving to the new version of CuTeDSL. These should be caught in basic MLIR LIT tests.

[GPU MODE @GPU_MODE]

Jack Carlisle & Jay Shah talk on the Fundamentals of CuTe Layout Algebra. Special guest Cris Cecka

See you soon! https://t.co/fg4OXeQYB2

[Manish Gupta @BigManniM9]

@marksaroufim Best of luck and wishes for your next adventure. More power. More GPUs. More kernels.

[Alex Zhurkevich @cudagdb]

Trtllmgen kernels are now open. Fastest prefill and decode kernels for our target workloads. We wrote these to win InferenceX, MLPerf, other benchmarks. Powering some of today’s top served models. Dive in, learn, use them, or level up your own. Enjoy.

https://t.co/2aQBwcdnZL

[Manish Gupta @BigManniM9]

@henrylhtsang Why do you need cutlass on cpu-only machine?

[Manish Gupta @BigManniM9]

@kohjingyu "1.45 years" - Training in FP8 and tenure in FP64.

[Manish Gupta @BigManniM9]

@jsuarez @ID_AA_Carmack Another example in the no reason category that we have faced bypassed https://t.co/nlYUpGam4U

[Manish Gupta @BigManniM9]

@gstsdn @metaai It is also impressive that if you replace any of the two "impressive"s in your tweet with "expensive", the sentence still holds.

[Manish Gupta @BigManniM9]

Sponsored ads: the AI colosseum is alive and well — Google’s still the gladiator ring where models battle for your clicks. The road to most tokens still runs through top clicks. ⚔️

#Sponsored #ads #OpenAl #Anthropic #Codex #Claude

[Manish Gupta @BigManniM9]

@_sanjoydas and @clattner_llvm PTX generation isn’t just emiting one PTX instruction. Yes, the backend (.td it or inline it) must cover legacy + new PTX. But the game is (a) sequencing, vectorizing, and interleaving of PTX within a warp, and (b) choreographing across warps in a warp-specialized kernel on Hopper and more on Blackwell. I read these results as : Mojo matches and/or beat alternatives on (a) and (b).

[Chris Lattner @clattner_llvm]

Thank you to folks at @metaai for publishing their independent perf analysis comparing CUDA and Mojo against Triton and TileLang DSLs, showing Mojo meeting and beating CUDA, and leaving DSLs in the dust.

[Modular @Modular]

We just launched Mojo🔥 GPU Puzzles Edition 1, a hands-on guide that teaches GPU programming through 34 progressive challenges, not lectures.
Learn by doing, from your first GPU threads to tensor cores.
Works on NVIDIA, AMD, and Apple GPUs.
https://t.co/nXy5cDSJCU

[Manish Gupta @BigManniM9]

Blackwell TMEM turns register spills from bad to a kernel design strategy. Need bigger tiles, larger faster memory. Spill/Fill RMEM to/from TMEM. Guess what the spilled data can be filled in other warps too :). #Blackwell #Kernels

[Manish Gupta @BigManniM9]

@clattner_llvm @samofort @kevbtu Congratulations @clattner_llvm

[Manish Gupta @BigManniM9]

@hyhieu226 @yacineMTB “Last 2% is the hardest to get, that’s why they leave it in the milk”