HPC & Quantum
27 subscribers
11.4K photos
668 videos
3 files
30.6K links
Download Telegram
HPC Guru (Twitter)

Democratizing #AI: #Opensource scalable #LLM training on #GPU-based #supercomputers

Research team led by @UofMaryland nominated for the #GBPrize for AxoNN, a scalable distributed training framework which leverages GPUs to train LLMs

https://www.olcf.ornl.gov/2024/11/12/gordon-bell-prize-nomination-recognizes-efforts-to-train-extreme-scale-large-language-models-using-frontier/

#HPC #SC24
This media is not supported in your browser
VIEW IN TELEGRAM
HPC Guru (Twitter)

RT @NVIDIAAIDev: Get over 3x perf improvement for #LLM throughput - TensorRT-LLM Multiblock Attention on NVIDIA HGX H200 boosts long-sequence performance without impacting time-to-first-token.

Read how ➡️ nvda.ws/3CE6hgS
HPC Guru (Twitter)

China's "Global Scheduling Ethernet" appears to have the same motives as @ultraethernet

GSE has been deployed in a large cluster in China - claim of substantial network performance improvements during training of a #LLM

https://www.theregister.com/2024/11/26/global_scheduling_ethernet_china_uec/

#AI #HPC via @TheRegister
HPC Guru (Twitter)

RT @thoefler: Great talking to LIU Bin, the Deputy President of @NUSingapore about #AI, #HPC, and #Health

I'm looking forward to talking about efficient #GenAI and the past and future of #LLM computing on Jan 10 at @NUSComputing!

https://events.comp.nus.edu.sg/view/23280 🤖

We're in the Age of Computation!
HPC Guru (Twitter)

#AI and the future of everything: Five ways #AI will change our world as we know it

By Kirk Bresniker, @HPE Fellow and Chief Architect at Hewlett Packard Labs

https://www.hpe.com/us/en/newsroom/blog-post/2025/01/ai-and-the-future-of-everything-five-ways-ai-will-change-our-world-as-we-know-it.html

#HPC #LLM #GenerativeAI
HPC Guru (Twitter)

Claim: 100x Faster* at 1/10th the cost

* Decode tok/s, versus a (cluster of) H100 GPUs with 8-bit quantisation and TensorRT-LLM, on Llama2 70B

Their website is: fractile.ai

#LLM #AI #Inference via @PGelsinger https://twitter.com/PGelsinger/status/1882159997167251812#m
HPC Guru (Twitter)

RT @thoefler: 🚀 Excited to kick off the 2025 @adia_lab seminar series on Tue!

I'll explore the role of computation in the history of LLMs and the rise of fascinating reasoning models (that authored this post 🤖)

Joining me are two inspiring speakers—can't wait for their insights!" #AI #LLM https://twitter.com/adia_lab/status/1882035491790586121#m
HPC Guru (Twitter)

El Reg digs its claws into Middle Kingdom's latest chain of thought model - results from @TheRegister's tests on DeepSeek's R1

It can tell you how many Rs in strawberry, but not anything about the Tiananmen Square massacre

https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/

#AI #GenAI #LLM
HPC Guru (Twitter)

RT @thoefler: From #LLMs 🤖 to Reasoning Language Models 🧠 Three Eras in the Age of Computation!

🔥 Progress in #AI and #Computing 🎥 https://www.youtube.com/watch?v=NFwZi94S8qc

💡 Combining the best knowledge databases (#LLM) with the best strategy play (#RL) will be only limited by computational cost 🚀 #HPC
HPC Guru (Twitter)

RT @thoefler: Do you wonder how Reasoning Language Models like #DeepSeek R1 are made?

A fascinating mix of #ReinforcementLearning, #MCTS, and #LLM training and finetuning on #HPC supercomputers.

Check our thinking framework to derive new exciting #RLM implementations: buff.ly/4hA6GQq
HPC Guru (Twitter)

RT @hpcnotes: Rio Yokota of Institute of Science Tokyo (new name for Titech) giving an update on #LLM work using #supercomputers in Japan at #MW25NZ mixed with an excellent overall discussion of wider international #AI evolution and advances

#HPC
HPCwire (Twitter)

A recent xAI headline seemed out of place. We take a closer look here: ow.ly/5YtN50VPrvt #AIdatacenter #LLM
HPC Guru (Twitter)

ICYMI: Microsoft and University of Science and Technology of China trained LLMs using #FP4 for matrix multiplications and achieved accuracy comparable to LLMs trained using the popular #BF16 format

arxiv.org/abs/2501.17116

#AI #LLM via @AndrewYNg
HPCwire (Twitter)

RT @alex_woodie: During his talk at #TPC25, Satoshi Matsuoka of @RIKEN says there is much debate within the AI community about the size of models @HPCwire #LLM #AIforScience
HPCwire (Twitter)

Beijing’s Moonshot AI (backed by Alibaba) just dropped Kimi K2, a 1 trillion parameter open-weight LLM built for serious code and agentic reasoning.

Learn more: ow.ly/yR9w50WwHTi

#LLM #MoonshotAI