正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Two days ago, Deepseek surprised everyone with an "undefined-behavior" PTX optimization speeding up particular ML workloads on a Hopper NVIDIA GPU Kernel. Let's reverse engineer the hack, implement it ourselves, and benchmark the speedup on an H100.

LaurieWired

165,099 subscribers

228,854 次观看 • 1 年前 •via X (Twitter)

科学技术新闻政治教育

Anya Rossi• Live Now

Private livecam show

11 条评论

LaurieWired 的头像

LaurieWired1 年前

Full Video:

LaurieWired 的头像

LaurieWired1 年前

My test code:

NetMind.AI 的头像

NetMind.AI2 年前

Get access to a wide range of GPUs like H100, A100, 4090, 3090 and save over 90% at NetMind Power. Rent Now!

numanumabruh 的头像

numanumabruh1 年前

You'd never know she's 6'5"

Jason Ho 的头像

Jason Ho1 年前

laurie supremacy

Bob (Moderna #7) Kerns 的头像

Bob (Moderna #7) Kerns1 年前

Until recently, I'd only seen your tweets; the first video I encountered was the 2025 prediction ones. Assumptions violated: higher voice, younger. Always good to have one's assumptions flagged, but especially the age. I was struck by the maturity of your analysis!

Dave 🚀 的头像

Dave 🚀1 年前

LFG!

KnowledgeisMostValuable 的头像

KnowledgeisMostValuable1 年前

I'd fight off a bear for you

nisten - e/acc 的头像

nisten - e/acc1 年前

lfg

🥀shiVam🥀 的头像

🥀shiVam🥀1 年前

wait, did you film this at Google HQ? (must appreciate the audio recording and editing)

Calcs 的头像

Calcs1 年前

Fantastic video, more please, lol 😂

相关视频

Just built an MCP for Ghidra. Now basically any LLM (Claude, Gemini, local...) can Reverse Engineer malware for you. With the right prompting, it automates a *ton* of tedious tasks. One-shot markups of entire binaries with just a click. Open source, on Github now.

Just built an MCP for Ghidra. Now basically any LLM (Claude, Gemini, local...) can Reverse Engineer malware for you. With the right prompting, it automates a ton of tedious tasks. One-shot markups of entire binaries with just a click. Open source, on Github now.

LaurieWired

286,798 次观看 • 1 年前

🚨BREAKING NEWS🚨: Phala Shipped the First-Ever GPU TEE Benchmark! Benchmark Research Highlight: ✅ LLaMa 3, Microsoft Phi Models Tested ✅ Tests Performed on nVIDIA H100 ✅ With #TEE Mode On 📊 Key Results: • Up to 150 TPS for LLaMA-3-8B • Performance trade-off as low as 0.5%, up to 5% Read Full Report: <

🚨BREAKING NEWS🚨: Phala Shipped the First-Ever GPU TEE Benchmark! Benchmark Research Highlight: ✅ LLaMa 3, Microsoft Phi Models Tested ✅ Tests Performed on nVIDIA H100 ✅ With #TEE Mode On 📊 Key Results: • Up to 150 TPS for LLaMA-3-8B • Performance trade-off as low as 0.5%, up to 5% Read Full Report: <

Phala

69,043 次观看 • 2 年前

The exact NVIDIA optimization order I follow on every PC I optimize. Clean driver. GPU-Z check. NVIDIA App settings. Most people skip step two and wonder why their GPU is running slow. Full guide here.

The exact NVIDIA optimization order I follow on every PC I optimize. Clean driver. GPU-Z check. NVIDIA App settings. Most people skip step two and wonder why their GPU is running slow. Full guide here.

Xilly

33,887 次观看 • 3 个月前

a Tsinghua University lab just put a project on GitHub that replaces a $400,000 H100 rack with a single 24GB graphics card. it's called ktransformers, and the trick is almost stupidly simple: the experts you actually use stay on the gpu, the ones you don't sit on the cpu until they're called / deepseek-v3 and r1 with 139K context in 24gb of vram / up to 28x speedup over the standard setup / fine-tune deepseek-v3 across four rtx 4090s instead of a datacenter / built by tsinghua university's madsys lab, not a startup with a landing page apache 2.0, and already past 17,000 stars. -> bookmark it.

a Tsinghua University lab just put a project on GitHub that replaces a $400,000 H100 rack with a single 24GB graphics card. it's called ktransformers, and the trick is almost stupidly simple: the experts you actually use stay on the gpu, the ones you don't sit on the cpu until they're called / deepseek-v3 and r1 with 139K context in 24gb of vram / up to 28x speedup over the standard setup / fine-tune deepseek-v3 across four rtx 4090s instead of a datacenter / built by tsinghua university's madsys lab, not a startup with a landing page apache 2.0, and already past 17,000 stars. -> bookmark it.

dunik

314,378 次观看 • 3 个月前

Woow Nvidia has just released a 2.6B open-source world model 🔥 You can turn a single image, text prompt and trajectory into controllable worlds... And on a single GPU! - Code available on GitHub - Paper as well on arxiv You can use it for many things like embodied AI and robotics research, simulations, etc. Because it can run on a single GPU (like an RTX 5090 or H100) it makes world models accessible to basically everyone!

Woow Nvidia has just released a 2.6B open-source world model 🔥 You can turn a single image, text prompt and trajectory into controllable worlds... And on a single GPU! - Code available on GitHub - Paper as well on arxiv You can use it for many things like embodied AI and robotics research, simulations, etc. Because it can run on a single GPU (like an RTX 5090 or H100) it makes world models accessible to basically everyone!

Paul Couvert

174,537 次观看 • 5 个月前

You can run CUDA, on a Mac ARM GPU, in the browser. It sounds ridiculous but it actually works. HipScript chains CUDA, to OpenCL, to Vulkan, to Tint (Google’s shader translator), to a WASM WebGPU. I got a plasma simulation in running in just a few minutes, no NVIDIA GPU!

You can run CUDA, on a Mac ARM GPU, in the browser. It sounds ridiculous but it actually works. HipScript chains CUDA, to OpenCL, to Vulkan, to Tint (Google’s shader translator), to a WASM WebGPU. I got a plasma simulation in running in just a few minutes, no NVIDIA GPU!

LaurieWired

160,942 次观看 • 1 年前

Woow Nvidia has a 2.6B open-source world model 🔥 Turn an image, a text prompt, and a camera trajectory into a controllable world. All on a single GPU. - Code available on GitHub - Paper on arXiv - Useful for embodied AI, robotics research, and simulations Runs on GPUs like an RTX 5090 or H100, making world-model experiments possible without a multi-GPU setup.

Woow Nvidia has a 2.6B open-source world model 🔥 Turn an image, a text prompt, and a camera trajectory into a controllable world. All on a single GPU. - Code available on GitHub - Paper on arXiv - Useful for embodied AI, robotics research, and simulations Runs on GPUs like an RTX 5090 or H100, making world-model experiments possible without a multi-GPU setup.

Simplifying AI

21,151 次观看 • 24 天前

Speeding up Gemma! Gemini 4 Argon enabled a 2.85x speedup on an A100 via vLLM, jumping from ~37,900 to 108,000+ tokens/sec on gemma-4-E2B-it. ✅ 100% token parity ✅ Zero quality drop ✅ Done in days, not months See the exact changes across 8 PRs:

Speeding up Gemma! Gemini 4 Argon enabled a 2.85x speedup on an A100 via vLLM, jumping from ~37,900 to 108,000+ tokens/sec on gemma-4-E2B-it. ✅ 100% token parity ✅ Zero quality drop ✅ Done in days, not months See the exact changes across 8 PRs:

Google Gemma

222,754 次观看 • 11 天前

PyTorch core engineer at Meta turned CUDA kernel writing into a sport in 13 minutes - better than $1500 GPU programming bootcamps. profile the kernel -> find the bottleneck -> rewrite -> benchmark -> merge the winning code into PyTorch. That loop is how the open community now beats hand-tuned vendor kernels. GPU MODE community + KernelBot competition + winning kernel merged into the framework - that's the stack. Watch it, then steal the loop below.

h100envy

35,390 次观看 • 4 个月前

First ever melted power connector on an Nvidia GPU.

First ever melted power connector on an Nvidia GPU.

PowerGPU

13,477 次观看 • 5 个月前

DeepSeek V4 ships with two variants allowing it to hit 1M Context. Kimbo @ ICML26 breaks down the attention changes and the Mega MOE speeding up compute.

DeepSeek V4 ships with two variants allowing it to hit 1M Context. Kimbo @ ICML26 breaks down the attention changes and the Mega MOE speeding up compute.

SemiAnalysis

39,693 次观看 • 4 个月前

The XG Mobile combines the power of an external GPU with the expandability of a Thunderbolt Dock. Boasting up to an NVIDIA® GeForce RTX™ 5090 Laptop GPU, this portable powerhouse is designed to game at the highest levels.

The XG Mobile combines the power of an external GPU with the expandability of a Thunderbolt Dock. Boasting up to an NVIDIA® GeForce RTX™ 5090 Laptop GPU, this portable powerhouse is designed to game at the highest levels.

ROG Global

13,948 次观看 • 1 年前

Timelapse #177 (14 hrs) - Picked up an RTX Pro 6000 Blackwell and wired in another 1000W PSU - running two of them now - DeepSeek v4 Flash running locally on both GPUs - Distilling DeepSeek v4 Flash down into Nanbeige (still cooking) - Kernel bench runs for DeepSeek v4 Flash and Qwen 3.8 Max - Got it into chat but couldn't reliably get it into the OMP harness yet - Private contract work - days like this are stuff I actually want to build. and when it's the high-intensity grind, i still have to show up and get it done

Timelapse #177 (14 hrs) - Picked up an RTX Pro 6000 Blackwell and wired in another 1000W PSU - running two of them now - DeepSeek v4 Flash running locally on both GPUs - Distilling DeepSeek v4 Flash down into Nanbeige (still cooking) - Kernel bench runs for DeepSeek v4 Flash and Qwen 3.8 Max - Got it into chat but couldn't reliably get it into the OMP harness yet - Private contract work - days like this are stuff I actually want to build. and when it's the high-intensity grind, i still have to show up and get it done

Elliot Arledge

17,401 次观看 • 3 个月前

YOUR PARENTS PAID FOR THE CUDA MOAT! The #1 contributor to the CUDA MOAT isn't the the developers at NVIDIA, but it is the millions of developers outside of NVIDIA that invent new algorithms for CUDA like Flash Attention. For most of them, it started with an GeForce gaming GPU. NVIDIA is the only companies that has an reasonable good developer stack on consumer grade GPUs. As people grow up beyond playing CSGO & League of Legends & Minecraft, they either become anime weeaboos or they start programming on their existing computer with has an GeForce GPU

YOUR PARENTS PAID FOR THE CUDA MOAT! The #1 contributor to the CUDA MOAT isn't the the developers at NVIDIA, but it is the millions of developers outside of NVIDIA that invent new algorithms for CUDA like Flash Attention. For most of them, it started with an GeForce gaming GPU. NVIDIA is the only companies that has an reasonable good developer stack on consumer grade GPUs. As people grow up beyond playing CSGO & League of Legends & Minecraft, they either become anime weeaboos or they start programming on their existing computer with has an GeForce GPU

SemiAnalysis

25,230 次观看 • 7 个月前

Unboxing the ROG Zephyrus G14 🔥 Featuring an Intel® Core™ Ultra 9 386H processor and up to an NVIDIA® GeForce RTX™ 5080 Laptop GPU, it delivers stunning visuals on a 14-inch 3K ROG Nebula HDR OLED display with up to 1100 nits peak brightness.

Unboxing the ROG Zephyrus G14 🔥 Featuring an Intel® Core™ Ultra 9 386H processor and up to an NVIDIA® GeForce RTX™ 5080 Laptop GPU, it delivers stunning visuals on a 14-inch 3K ROG Nebula HDR OLED display with up to 1100 nits peak brightness.

ROG Global

20,645 次观看 • 7 个月前

I won an RTX 5090 GPU at Mistral AI worldwide hackathon in Paris, and I have absolutely no idea on how to use it. Here's what I actually built: an image generation model fine-tuned to do technical analysis on trading charts. I trained it on an NVIDIA cloud GPU then back-tested it against Binance historical data. The model was profitable. The jury didn't quite catch the vibe. But apparently someone did because two weeks later, they reached out to tell me I'd won the NVIDIA track. Today, a €4,000 RTX Founders Edition showed up at my door. Absolutely gorgeous piece of hardware. One small problem: I use a MacBook. So if anyone has creative suggestions for what to do with a top-tier GPU and no desktop to put it in… I'm listening Huge thanks to NVIDIA, Mistral AI and Iterate team for appreciating my weird idea and an attempt to push tech to its boundaries

I won an RTX 5090 GPU at Mistral AI worldwide hackathon in Paris, and I have absolutely no idea on how to use it. Here's what I actually built: an image generation model fine-tuned to do technical analysis on trading charts. I trained it on an NVIDIA cloud GPU then back-tested it against Binance historical data. The model was profitable. The jury didn't quite catch the vibe. But apparently someone did because two weeks later, they reached out to tell me I'd won the NVIDIA track. Today, a €4,000 RTX Founders Edition showed up at my door. Absolutely gorgeous piece of hardware. One small problem: I use a MacBook. So if anyone has creative suggestions for what to do with a top-tier GPU and no desktop to put it in… I'm listening Huge thanks to NVIDIA, Mistral AI and Iterate team for appreciating my weird idea and an attempt to push tech to its boundaries

Kirill

52,573 次观看 • 6 个月前

Excited to introduce PyRoki ("Python Robot Kinematics"): easier IK, trajectory optimization, motion retargeting... with an open-source toolkit on both CPU and GPU

Excited to introduce PyRoki ("Python Robot Kinematics"): easier IK, trajectory optimization, motion retargeting... with an open-source toolkit on both CPU and GPU

Chung Min Kim

117,384 次观看 • 1 年前

Dylan Patel on the importance of memory and storage Two key quotes: "An $NVDA GPU is faster than an $AMD GPU in most cases, but because AMD GPUs have more memory, they can outperform Nvidia in certain workloads." “It is a difficult, multivariable problem. Generally, you need the best GPU, such as a GB300, but you also need the best storage solutions. I will not spoil who comes out on top, but storage solutions matter a lot, memory solutions matter a lot, and frontend networking also matters significantly" Full Quote: “We have over $80 million of compute: GPUs from $NVDA and $AMD, TPUs from Google, and Trainium from Amazon. We constantly run this benchmark using the newest inference engines, drivers, PyTorch versions, and other software. It runs every day through automated CI across the latest Chinese models from GLM, Zhipu, Moonshot, Kimi, Alibaba, and others. Initially, when we were benchmarking the differences between these chips, inference engines, and parallelism schemes, we used fixed context lengths. But with Agent X, we have now analyzed more than $5 million worth of Claude Code traces. This is real production traffic that users have donated to us, combined with internally generated data, so we now understand what an actual agent workload looks like. When we implement those workloads and run the benchmarks, it turns out that the chip you are using is very important, but how you handle memory offload can be even more important. An Nvidia GPU is faster than an AMD GPU in most cases, but because AMD GPUs have more memory, they can outperform Nvidia in certain workloads. Similarly, you can use a less powerful GPU with a much better storage solution and outperform the best GPU when it lacks those solutions. Simply buying the newest GPU does not necessarily give you the best inference economics. You need to layer in other innovations, including storage and memory.” Interviewer: “Who is the top player on your chart? Can you tell us?” Dylan Patel: “It is a difficult, multivariable problem. Generally, you need the best GPU, such as a GB300, but you also need the best storage solutions. I will not spoil who comes out on top, but storage solutions matter a lot, memory solutions matter a lot, and frontend networking also matters significantly.”

Dylan Patel on the importance of memory and storage Two key quotes: "An $NVDA GPU is faster than an $AMD GPU in most cases, but because AMD GPUs have more memory, they can outperform Nvidia in certain workloads." “It is a difficult, multivariable problem. Generally, you need the best GPU, such as a GB300, but you also need the best storage solutions. I will not spoil who comes out on top, but storage solutions matter a lot, memory solutions matter a lot, and frontend networking also matters significantly" Full Quote: “We have over $80 million of compute: GPUs from $NVDA and $AMD, TPUs from Google, and Trainium from Amazon. We constantly run this benchmark using the newest inference engines, drivers, PyTorch versions, and other software. It runs every day through automated CI across the latest Chinese models from GLM, Zhipu, Moonshot, Kimi, Alibaba, and others. Initially, when we were benchmarking the differences between these chips, inference engines, and parallelism schemes, we used fixed context lengths. But with Agent X, we have now analyzed more than $5 million worth of Claude Code traces. This is real production traffic that users have donated to us, combined with internally generated data, so we now understand what an actual agent workload looks like. When we implement those workloads and run the benchmarks, it turns out that the chip you are using is very important, but how you handle memory offload can be even more important. An Nvidia GPU is faster than an AMD GPU in most cases, but because AMD GPUs have more memory, they can outperform Nvidia in certain workloads. Similarly, you can use a less powerful GPU with a much better storage solution and outperform the best GPU when it lacks those solutions. Simply buying the newest GPU does not necessarily give you the best inference economics. You need to layer in other innovations, including storage and memory.” Interviewer: “Who is the top player on your chart? Can you tell us?” Dylan Patel: “It is a difficult, multivariable problem. Generally, you need the best GPU, such as a GB300, but you also need the best storage solutions. I will not spoil who comes out on top, but storage solutions matter a lot, memory solutions matter a lot, and frontend networking also matters significantly.”

Daniel Romero

48,978 次观看 • 3 个月前