
Dylan Patel
@dylan522p
Followers
88K
Following
35K
Media
1K
Statuses
12K
SemiAnalysis Boutique AI & Semiconductor Research and Consulting DMs are open for consulting, quotes, or to talk shop
Joined April 2018
RT @asianometry: In today's insane (and not in a good way) episode of Transistor Radio, we talk but can't talk about Meta's hiring spree, p….
0
4
0
The weakness of the US grid isn't just the lack of power.It's also the lack of mechanisms to keep it stable.I expect a training run to plunge hundreds of thousands of people into a blackout.This will make normies anti-AI infrastructure.There's solutions we can implement though!.
AI Training Load Fluctuations at Gigawatt-scale.Risk of Power Grid Blackout?.108GW Large Load Queue.Tesla Megapacks.Supercapacitors.Gigawatt-scale batteries.PyTorch No Power Plant Blow Up.
22
22
313
IDK who Kevin Rippey is but this is the funniest thing I've ever seen written about me!.Thanks for the compliment Kevin!.We had a sick core research note today on Arista. Slides below I think @IanCutress got like 50% of them but we'll post them with full explanation soon. $ANET
Going fast. Thisbnight be a slide show thread with minimal notes FYI. Replacing copper with other stuff.
14
1
103
The Nvidia Tensor Core is the most important evolution of computer architecture in the last decade.We explain why / how it's evolved.Shout out to collaborators @bfspector @tri_dao @colfaxintl @charles_irl @ia_buck Neil Movva Jonah Alben.esp @simonguozirui for the cutest cover pic.
NVIDIA Tensor Core Evolution.From Volta To Blackwell.Amdahl’s Law, Strong Scaling.Asynchronous Execution.Blackwell, Hopper, Ampere, Turing, Volta.
8
22
324
Morris Chang started TSMC at 55. That's the most important and irreplaceable company in the world. This young boy culture in SF is weird. I'm still in my 20s, but I know there's even bigger things I'ma do in my 30s and 40s and 50s and 60s and 70s.
steve jobs was 21 when he made apple.kalomaze was 19 when he joined prime intellect .it’s too late, Give up
46
43
739
Selective metal printers are cool as fuck.Met a company that's printing missle chassis at 72 an hour the other week. Thread quoted about it.
Nikon lost the DUV race to ASML, but their laser sources never quite disappeared. Meet the Nikon SLM NXG, which resembles a DUV machine surprisingly closely. Buckle up and hear about Nikon's NXG printer, which could kick off a generative manufacturing revolution, especially for
35
2
101
AMD is making big moves in winning over the neocloud ecosystem through their balance sheet. At the same time Nvidia is alienating many with DGX Lepton.MI355 is great perf per TCO againist HGX B200 but it is not rack scale unlike GB200 nvl72.MI355 is rack scale from temu dot com.
AMD Advancing AI: MI350X, MI400 UALoE72, MI500 UAL256.Rapid Software Improvement.Marketing RDF.AMD Fostering Neocloud.MI355 is not Rack Scale, MI400 is UALoE, Not UALink.
29
15
326
RT @rob_lh: @dylan522p on Transistor Radio: "I can be a professional, this is just a character I play for the podcast.". Dylan in his day j….
0
5
0
Meta Scale AI deal is wild.Lotta folks are criticizing it.Multiple labs now backing away from Scale data.Meta fell behind despite lotta spend on compute + team + data.Snagging @alexandr_wang + crazy salaries for talent.Is it desperation or leadership?.What should they have done?.
67
14
246
RT @AnushElangovan: docker run --gpus now works on AMD. Ease of use is key. Thanks to Semi-Analysis team for the….
0
35
0
Mistral got hit by export restrictions again!.They couldn’t evaluate the latest DeepSeek and Qwen.I am risking being detained by going around export restrictions and checking the numbers on Qwen.TL;DR Qwen 4B is ~close to their model and the small 30B MoE is better, let’s not
Announcing Magistral, our first reasoning model designed to excel in domain-specific, transparent, and multilingual reasoning.
39
22
535
RT @dylan522p: @cHHillee @willccbb If ~= O(MlogM + E), M = mantissa, E = exponent, then e5m2 and e4m3 should not save much power vs e3m2 an….
0
2
0