Anoop Saha
@asyncanoop
I correlate; therefore, I cause! Building and deploying RL post training Infrastructure @awscloud. #hyperpod #100kGPU
Was dir gefallen könnte
When in doubt, write CUDA kernels.
Meta just dropped this paper that spills the secret sauce of reinforcement learning (RL) on LLMs. It lays out an RL recipe, uses 400,000 GPU hrs and posits a scaling law for performance with more compute in RL, like the classic pretraining scaling laws. Must read for AI nerds.
Don’t do that. I REPEAT. Do not work on an FPGA.
openai is building its own chips and youre still trying to learn CUDA pick up an fpga and learn
👀👀👀
SCOOP: TSMC IS LOSING MONEY ON NEW 2NM CHIP RAMP!!!!!! THERE ARE NEGATIVE GROSS MARGINS ON A NEW PRODUCT, THERE ARE NO CLEAR PLANS ON HOW TO IMPROVE MARGIN ACCORDING TO DOCUMENTS ACQUIRED BY ME (SEC FILINGS)
Huge news for both AMD and OpenAI! - 6 GW of AMD MI450, Helios, and next gen deployment for OpenAI (NVIDIA deal with OpenAI is at 10GW) - 160 million AMD shares to OpenAI in tranches - Incremental revenue of over $100B over the next few years wsj.com/tech/ai/openai…
There are two main problems in AI that’s worth solving for. How to build an effective model cost effectively that can learn continuously based on user interactions. And how to serve these models cost effectively - maximizing the tokens per $$. And we are still a year or two out…
In between all the AI for science news, the coolest effort is @CAIAorg - the cancer AI alliance - bringing folks together across the industry. If it can make meaningful progress, world would be a great place.
🚀 This is a call to tackle the hard-tier problems where RLVR pipelines stall with pass@k=0 (no reward, no gradient). To push further, we need to train on hard subsets and leverage signals from tough data — that’s where discovery happens. 👉 github.com/sunblaze-ucb/a… The grand…
My entire timeline is @periodiclabs content. Great job to the team.
Can't understate the talent level I'm surrounded by at @PeriodicLabs The team is largely a mix of: - researchers from OpenAI, GDM, FAIR/MSL - PhDs from Stanford, MIT, Harvard, etc - leaders in open-source ML - world-class physics professors Humbled to learn from them every day.
New in-depth blog post time: "Inside NVIDIA GPUs: Anatomy of high performance matmul kernels". If you want to deeply understand how one writes state of the art matmul kernels in CUDA read along. (Remember matmul is the single most important operation that transformers execute…
“The reports of my demise are highly exaggerated “ - GRPO, circa 2025.
In this new blog, we show how you can leverage the instance topology with EKS to reduce latency of distributed inference aws.amazon.com/blogs/machine-…
aws.amazon.com
Schedule topology-aware workloads using Amazon SageMaker HyperPod task governance | Amazon Web...
In this post, we introduce topology-aware scheduling with SageMaker HyperPod task governance by submitting jobs that represent hierarchical network information. We provide details about how to use...
This take is accurate - open weight models is not open source. Nobody is releasing their training code or even the training data. However the true assets to the ecosystem are open source technology - from DeepEP by DeepSeek, to Verl by Bytedance, PyTorch and Ray, sglang and…
Anthropic CEO Dario谈开源模型： - 大模型开放权重不同于软件开源，不存在开发者社区的反向贡献。 - 开源只是吸引注意力的幌子，用户只关心这个模型是否好用。Deepseek开源与否都无所谓，作为一个超大模型，推理起来很困难。 - 开源并不等于免费，推理服务器运行，是有成本的。
The cruelty is the point….
Those on an H1B cannot return to the US from tomorrow (Sunday) unless paying $100K. This is an out-of-the blue presidential action. We’ll see software engineers stranded abroad. One easy to predict outcome: those on US visas will travel less… for work, for conferences etc.
This is huge and can alone bring down the cost and duration of RL significantly.
Introducing checkpoint-engine: our open-source, lightweight middleware for efficient, in-place weight updates in LLM inference engines, especially effective for RL. ✅ Update a 1T model on thousands of GPUs in ~20s ✅ Supports both broadcast (sync) & P2P (dynamic) updates ✅…
My Google colleagues Norm Jouppi & Sridhar Lakshmanamurthy gave a talk today at Hot Chips on TPUv7 ("Ironwood"). The TPUv7 system offers 9216 chips / pod (42.5 exaflops of fp8), but we can scale across many of these pods to provide multiple zettaflops.
Google presents for the first time ever their TPUv7 block diagram at hot chips conference. TPUv7 (formerly known as TPUv6p, internally called ghostfish) has 8 stacks of HBM3e memory, 4 medium size systolic arrays and be connected in a 3D torus with a scale up world size of up to…
This is an excellent overview. I am surprised despite all the growth in AI, how little material there is in core GPU architecture. The JAX team changes it for good.
Today we're putting out an update to the JAX TPU book, this time on GPUs. How do GPUs work, especially compared to TPUs? How are they networked? And how does this affect LLM training? 1/n
Amazon’s chip division is working with Taiwan’s Alchip on Trainium3 and Trainium4 to bring the chips to mass production on TSMC 3nm and 2nm processes, respectively, media report, citing unnamed industry sources. Trainium3 will be in mass production in the 1st quarter of 2026.…
United States Trends
- 1. #เพียงเธอตอนจบ 1.27M posts
- 2. LINGORM ONLY YOU FINAL EP 1.28M posts
- 3. #FanCashDropPromotion 1,141 posts
- 4. Apple TV 8,952 posts
- 5. #FridayVibes 6,680 posts
- 6. trisha paytas 1,092 posts
- 7. Good Friday 59.4K posts
- 8. No Kings 210K posts
- 9. zendaya 3,739 posts
- 10. #SlideToMe 12.6K posts
- 11. #Yunho 22.8K posts
- 12. Shabbat Shalom 4,207 posts
- 13. Mamdani 274K posts
- 14. Cuomo 118K posts
- 15. F1 TV 2,718 posts
- 16. Ayla 168K posts
- 17. Justice 330K posts
- 18. Happy Friyay 1,485 posts
- 19. Arc Raiders 4,436 posts
- 20. My President 61.1K posts
Was dir gefallen könnte
-
Joe Burden
@slowbuddens -
Dylan Patel
@dylan522p -
Vikas Khanna
@TheVikasKhanna -
GaonConnection
@GaonConnection -
Cabot Wealth Network
@CabotAnalysts -
𝐷𝑟. 𝐼𝑎𝑛 𝐶𝑢𝑡𝑟𝑒𝑠𝑠
@IanCutress -
Samar Halarnkar
@samar11 -
Dr. Sanjukta Basu, M.A., LLB., PhD
@sanjukta -
bored
@boredbh4r -
Sabbah Haji Done With This Place
@imsabbah -
Lee Weissman
@JihadiJew -
Oxblood Ruffin
@OxbloodRuffin -
Vipul Ramtekkar
@VipulRamtekkar -
sonia_manchanda
@sonia_manchanda -
Raju Yadav(P D A parivar)
@JlnYadav
Something went wrong.
Something went wrong.