Zev Rekhter
@zevrekhter
Building AI Infrastructure
SGLang on AMD MI300X delivers 2x performance boost over VLLM on NVIDIA H100 for Deepseek-R1 inference. @lmsysorg @GenAI_is_real @zhyncs42 @deepseek_ai @AnushElangovan
Handwriting leads to widespread brain connectivity - typing does not.
Huge thanks to @AMD for donating an MI350 to SGLang! This advanced AI accelerator is making a meaningful difference—enabling us to move faster in developing scalable LLM systems and pushing the limits of inference optimization. Special thank to our awesome infra partner…
so @MarathonFusion figured out how to make infinite gold from mercury
Totally unmarked ads in the feed shown as just normal posts are the worst change ever pls roll this back @nikitabier it’s absolutely terrible
Broccoli is a flower that is harvested before it blooms. If you leave it alone it will bloom
can someone please explain why this happens
ADHD mfs spending the whole day in waiting mode because they have a thing at 5pm
Andrew Ng says the most extreme AI narratives are PR weapons. Extinction. Job loss. Nuclear-level energy needs. Narratives that make companies look powerful, not truthful. “A handful of companies got away with saying almost anything, without anyone fact checking them.”
hey jaguar. you're trying do picture 1, which will be a huge failure. when you should just put a battery in picture 2, which would kick ass and sell like crazy. literally don't even add a touch screen or anything. same knobs and dials. but with a battery. you're welcome
Agree!!!
AMD's AI software ROCm will inevitably outgrow Nvidia's CUDA due to it being open source and not proprietary. There are limitations for what can be done exclusively in-house. When you isolate yourself, you limit development potential, especially when larger communities join.
What happened to that 100M context window start up from like 2 years ago
So excited to share one of my white-whale projects: a fully physics-based, holographic foil shader. Each pixel simulates a ray of light diffracting into a rainbow of waves, which add and subtract to create these incredible patterns. Not a single gradient was used here!
In 2017, eight years ago, @AMD talked about the Project 47. With 30 EPYC 7601 and 80 Instinct GPUs they achieved 1 PFLOPs of FP32 in one Rack. Today you can have this with a single UBB8 board and eight Instinct MI355X.
Lisa Su: “I'm happy to announce that MI355 production shipments actually started earlier this month, and we have the initial wave of partners on track to launch platforms and public cloud instances here in the third quarter.” There we go…. $AMD
this never gets old @__tinygrad__ @realGeorgeHotz @comma_ai
imagine getting sued by sony for hacking their playstation and your first response is a diss track george hotz is the name
Got my swag bag from @AMD ...along with the $500,000 MI350X datacenter GPU server. grab em while you can @AIatAMD Advancing AI 2024 the future of AI is open ecosystem and real competition
