Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Need faster LLM inference without sacrificing accuracy? Speculative decoding can help. Choosing the right draft length and drafting method depends on your model, workload and hardware. We break down five practical guidelines for balancing throughput and latency.

NVIDIA AI

338,344 subscribers

35,445 views • 12 days ago •via X (Twitter)

Science & Technology Education Comedy

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Everyone talks about our hardware @Cerebras. Few notice the software. Ryan Loney breaks down the hidden optimizations powering 20× faster LLM inference than GPUs, speculative decoding, token reuse, and why we’re just getting started. Watch the full story here

Everyone talks about our hardware @Cerebras. Few notice the software. Ryan Loney breaks down the hidden optimizations powering 20× faster LLM inference than GPUs, speculative decoding, token reuse, and why we’re just getting started. Watch the full story here

Cerebras

222,816 views • 8 months ago

Production-ready RAG systems need observability. From tracking latency and throughput to evaluating response quality with human feedback or LLM-as-a-judge, robust observability gives you visibility into both system performance and output quality, on both a component and system-wide level. This lesson from our Retrieval Augmented Generation course breaks down the core components of an effective eval system and how to balance cost, automation, and accuracy when choosing your metrics. 📚 Learn more in the full course:

Production-ready RAG systems need observability. From tracking latency and throughput to evaluating response quality with human feedback or LLM-as-a-judge, robust observability gives you visibility into both system performance and output quality, on both a component and system-wide level. This lesson from our Retrieval Augmented Generation course breaks down the core components of an effective eval system and how to balance cost, automation, and accuracy when choosing your metrics. 📚 Learn more in the full course:

DeepLearning.AI

18,793 views • 8 months ago

Just launched #CES2026, the new open-source NVIDIA Nemotron Speech ASR model is here to solve latency drift and redundant compute. Its cache-aware streaming architecture eliminates the need for buffered inference, giving you stable, sub-100ms latency (24ms median T-T-F) and up to 3x more throughput on your GPU. 🤗 Read the technical blog with real-world results from Daily and Modal on Hugging Face:

Just launched #CES2026, the new open-source NVIDIA Nemotron Speech ASR model is here to solve latency drift and redundant compute. Its cache-aware streaming architecture eliminates the need for buffered inference, giving you stable, sub-100ms latency (24ms median T-T-F) and up to 3x more throughput on your GPU. 🤗 Read the technical blog with real-world results from Daily and Modal on Hugging Face:

NVIDIA AI Developer

138,518 views • 8 months ago

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Researchers found a way to make LLMs 8.5x faster! (without compromising accuracy) Speculative decoding is quite an effective way to address the single-token bottleneck in traditional LLM inference. A small "draft" model first generates the next several tokens, then the large model verifies all of them at once in a single forward pass. If a token at any position is wrong, you keep everything before it and restart from there. This never does worse than normal decoding. But current drafters in Speculative decoding still guess one token at a time. That makes the drafting step itself a bottleneck, capping real-world speedups at 2-3x. DFlash is a new technique that swaps the autoregressive drafter with a lightweight block diffusion model that guesses all tokens in one parallel shot. Drafting cost stays flat no matter how many tokens you speculate. On top of that, the drafter is conditioned on hidden features pulled from multiple layers of the target model and injected into every draft layer, so it makes significantly better guesses than a drafter working from scratch. In the side-by-side demo below, vanilla decoding runs at 48.5 tokens/sec. DFlash hits 415 tokens/sec on the same model, with zero quality loss. It's already integrated with vLLM, SGLang, and Transformers, with draft models on HuggingFace for several models like Qwen3, Qwen3.5, Llama 3.1, Kimi-K2.5, gpt-oss, and many more. I have shared the GitHub repo in the replies! KV caching is another must-know technique to boost LLM inference. I recently wrote an article about it. Read it below. 👉 Over to you: What use case are you working on that can benefit from this new technique?

Avi Chawla

157,390 views • 4 months ago

Introducing EVO Router - a smart router that continuously optimizes and hill-climbs on your AI inference workloads. It learns from your code, prompts, production traffic, use cases, and SLAs, then searches for the best setup for every workload. That can mean more than choosing just a single model or routing through a classifier. EVO (YC F26) can optimize across model × provider combinations, fusion systems, cascades, routing policies, and prompts — whatever the workload allows, while staying within your quality, latency, and reliability constraints. As open-source models move closer to frontier-model performance, we believe this kind of workload-specific hill-climbing will become an important part of how AI systems are deployed and operated. Our early users have seen 30–60% lower inference costs across use cases ranging from complex multi-turn agents to asynchronous batch workloads. EVO Router has also reached the Pareto frontier on several benchmarks, outperforming existing setups on cost, quality, or both.

Introducing EVO Router - a smart router that continuously optimizes and hill-climbs on your AI inference workloads. It learns from your code, prompts, production traffic, use cases, and SLAs, then searches for the best setup for every workload. That can mean more than choosing just a single model or routing through a classifier. EVO (YC F26) can optimize across model × provider combinations, fusion systems, cascades, routing policies, and prompts — whatever the workload allows, while staying within your quality, latency, and reliability constraints. As open-source models move closer to frontier-model performance, we believe this kind of workload-specific hill-climbing will become an important part of how AI systems are deployed and operated. Our early users have seen 30–60% lower inference costs across use cases ranging from complex multi-turn agents to asynchronous batch workloads. EVO Router has also reached the Pareto frontier on several benchmarks, outperforming existing setups on cost, quality, or both.

Alok Bishoyi

16,014 views • 1 month ago

The Inference Engineering Masterclass: 10x faster models, quantization, speculative decoding, Rubin, & self-optimizing AI Baseten Philip Kiely and ali explain what actually happens after a model is trained, why turning weights into a fast and reliable product creates an entirely new optimization problem, how quantization errors can cancel out to unlock more throughput, why inference teams are still finding 20–200% performance gains, how video generation runs into a quadratic attention wall, and how GLM-5.2 helped rewrite and optimize the GPU kernels serving GLM-5.2 itself.

The Inference Engineering Masterclass: 10x faster models, quantization, speculative decoding, Rubin, & self-optimizing AI Baseten Philip Kiely and ali explain what actually happens after a model is trained, why turning weights into a fast and reliable product creates an entirely new optimization problem, how quantization errors can cancel out to unlock more throughput, why inference teams are still finding 20–200% performance gains, how video generation runs into a quadratic attention wall, and how GLM-5.2 helped rewrite and optimize the GPU kernels serving GLM-5.2 itself.

Latent.Space

148,203 views • 1 month ago

SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as batch size scales. We heavily optimized variable-length verification in SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇

SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as batch size scales. We heavily optimized variable-length verification in SGLang. Across batch sizes 1 to 256, DSpark gives the best throughput/latency tradeoff on DeepSeek-V4-Flash, ahead of both MTP and non-spec. At high concurrency, dynamic scheduling provides up to ~20% higher throughput compared to a fixed budget, while maintaining high verification quality across workloads. With fused kernels and zero-overhead scheduling, DeepSeek-V4-Pro reaches 383.7 tok/s at B=1 on B300. DSpark is now available in SGLang with support for Qwen3 and DeepSeek-V4. Thanks DeepSeek for open-sourcing! Blog with full technical details and commands to run below 👇

LMSYS Org

170,431 views • 2 months ago

Why are there so many ways to connect to Postgres? Because each method has tradeoffs. The right choice depends on your architecture. Let's break down 5 connection methods including direct, PgBouncer, and Hyperdrive.

Why are there so many ways to connect to Postgres? Because each method has tradeoffs. The right choice depends on your architecture. Let's break down 5 connection methods including direct, PgBouncer, and Hyperdrive.

PlanetScale

16,681 views • 8 months ago

Qwen3.6-35B-A3B-NVFP4 Benchmark: MoE vs. Dense Performance Report This report details a performance comparisonbetween two artificial intelligence models, focusing on the Qwen3.6-35B-A3B-NVFP4 and its predecessor. While both models achieved flawless accuracy across a comprehensive 65-test benchmark covering vision, video, and reasoning, the 35B version demonstrated significant speed advantages. Utilizing a Mixture of Experts architecture, the 35B model achieved nearly four times higher throughput and drastically reduced latency for initial responses. It also showed superior efficiency in speculative decoding and successfully managed massive context lengths up to 128,000 tokens. Ultimately, the source recommends the 35B model for high-speed production environments, while noting the 27B model remains more stable for specific complex reasoning modes.

Qwen3.6-35B-A3B-NVFP4 Benchmark: MoE vs. Dense Performance Report This report details a performance comparisonbetween two artificial intelligence models, focusing on the Qwen3.6-35B-A3B-NVFP4 and its predecessor. While both models achieved flawless accuracy across a comprehensive 65-test benchmark covering vision, video, and reasoning, the 35B version demonstrated significant speed advantages. Utilizing a Mixture of Experts architecture, the 35B model achieved nearly four times higher throughput and drastically reduced latency for initial responses. It also showed superior efficiency in speculative decoding and successfully managed massive context lengths up to 128,000 tokens. Ultimately, the source recommends the 35B model for high-speed production environments, while noting the 27B model remains more stable for specific complex reasoning modes.

Mike Gannotti

10,457 views • 2 months ago

You can't buffer an LLM. Unlike video streaming, where you can render lower-resolution frames while data loads, an LLM requires 100% of its KV cache to generate the next token. This means your slowest chunk dictates your entire inference speed. If you are managing LLM infrastructure, optimizing for average latency isn't enough; your architecture is entirely bound by P99 performance. Samuel Shen breaks down how dropping variable-sized metadata and enforcing fixed-size chunking doubled LMCache's throughput from 2GB/s to 5GB/s. Watch the full technical breakdown here: #LMCache #KVCache #LLM #AIInfrastructure #OpenSource

You can't buffer an LLM. Unlike video streaming, where you can render lower-resolution frames while data loads, an LLM requires 100% of its KV cache to generate the next token. This means your slowest chunk dictates your entire inference speed. If you are managing LLM infrastructure, optimizing for average latency isn't enough; your architecture is entirely bound by P99 performance. Samuel Shen breaks down how dropping variable-sized metadata and enforcing fixed-size chunking doubled LMCache's throughput from 2GB/s to 5GB/s. Watch the full technical breakdown here: #LMCache #KVCache #LLM #AIInfrastructure #OpenSource

LMCache Lab

10,360 views • 2 months ago

Watch Nvidia CEO Jensen Huang reveal what he described as the state of the art in AI hardware: the new Rubin computing architecture. “Vera Rubin is designed to address this fundamental challenge that we have: The amount of computation necessary for AI is skyrocketing.” Huang said. “Today, I can tell you that Vera Rubin is in full production.” According to Nvidia’s tests, Rubin architecture operates three and a half times faster than the previous Blackwell architecture on model-training tasks and five times faster on inference tasks, and that's just the start.

Watch Nvidia CEO Jensen Huang reveal what he described as the state of the art in AI hardware: the new Rubin computing architecture. “Vera Rubin is designed to address this fundamental challenge that we have: The amount of computation necessary for AI is skyrocketing.” Huang said. “Today, I can tell you that Vera Rubin is in full production.” According to Nvidia’s tests, Rubin architecture operates three and a half times faster than the previous Blackwell architecture on model-training tasks and five times faster on inference tasks, and that's just the start.

TechCrunch

76,412 views • 8 months ago

Laguna S 2.1 is built to run locally. We heard the ask for a serious open model in the 80–120B range. Laguna S 2.1 is our answer, and we want it to redefine what this size class can do. Weights ship in BF16, FP8, INT4 and NVFP4, with official GGUF and MLX conversions, plus official DFlash draft models for faster inference. Day-one support across llama.cpp, ollama, Valerie Lee Make Up, SGLang, NVIDIA TRT-LLM, Hugging Face transformers and ZML. Watch it run on an NVIDIA AI DGX Spark.

Laguna S 2.1 is built to run locally. We heard the ask for a serious open model in the 80–120B range. Laguna S 2.1 is our answer, and we want it to redefine what this size class can do. Weights ship in BF16, FP8, INT4 and NVFP4, with official GGUF and MLX conversions, plus official DFlash draft models for faster inference. Day-one support across llama.cpp, ollama, Valerie Lee Make Up, SGLang, NVIDIA TRT-LLM, Hugging Face transformers and ZML. Watch it run on an NVIDIA AI DGX Spark.

Poolside

31,242 views • 1 month ago

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

🚨 YOUR GPU IS PROBABLY WASTING MORE THAN YOU THINK. vLLM is built to squeeze far more useful work out of your GPU when serving LLMs. Running an LLM at scale isn’t just about having a powerful GPU. The real problem is how efficiently you use its memory and compute. That’s where vLLM comes in. → High-throughput LLM inference and serving → PagedAttention for smarter KV-cache memory management → Continuous batching to keep GPUs busy → Prefix caching + chunked prefill → OpenAI-compatible API out of the box → Supports a huge range of modern LLM architectures → Quantization support for running models more efficiently And the crazy part? You can start an OpenAI-compatible inference server with: `vllm serve ` So your application can talk to your own model almost like it’s talking to OpenAI. The bigger idea: Don’t just buy more GPUs. Make the GPUs you already have work harder. That’s why vLLM has become such a major project in LLM inference. 🔥 #vLLM #AI #LLM #Inference #GPU #MachineLearning #AIInfrastructure #OpenSource #AIAgents #Developers

Vikas gupta

14,138 views • 5 days ago

Realtime-VLA FLASH tackles one of the biggest deployment bottlenecks for diffusion-based VLAs: inference latency. The key idea is speculative inference for flow-matching VLAs. A lightweight draft model predicts an action chunk, while the main model’s Action Expert verifies it in parallel using flow-consistency checks instead of running full denoising every replanning round. This lets the system replace many expensive 58 ms full inference rounds with speculative rounds as fast as 7.8 ms, reducing average latency to 19.1 ms and achieving a 3.04× speedup on LIBERO while largely preserving success rate. Interesting systems insight: they profile π0 and show VLM prefill is compute-bound, while Action Denoise is memory-bound. FLASH exploits this by reusing KV cache and parallelizing verification instead of repeatedly running sequential denoising.

Robots Digest 🤖

14,576 views • 4 months ago

If you’re over 50 and training the same way you did at 30, we need to talk. I sat down with Jim Stoppani to break down what actually happens to muscle as we age, why leucine matters, when I’d take amino acids, how creatine fits in, and how to keep building muscle without destroying your joints. This one is packed with practical takeaways you can use immediately. Watch the full episode right here...

If you’re over 50 and training the same way you did at 30, we need to talk. I sat down with Jim Stoppani to break down what actually happens to muscle as we age, why leucine matters, when I’d take amino acids, how creatine fits in, and how to keep building muscle without destroying your joints. This one is packed with practical takeaways you can use immediately. Watch the full episode right here...

Gary Brecka

43,805 views • 15 days ago

$NBIS cofounder Roman Chernin describes how their recent acquisitions of Eigen AI and Clarifai were all about speed, incredible talent, and acceleration: "The philosophy is very simple. We need to build so many things, and we need to move so fast, that we're always looking for people who can accelerate us. It should be exceptional talent, and/or something that has a great adoption." "Our two recent acquisitions [were] two teams that work on inference optimization. A big part of our business is how efficiently we convert GPUs into tokens. And these two teams — Eigen AI and Clarifai — one is focused on model optimization, the engine of inference. How you run specific models and all the techniques around spec decoding, quantization, and so on." "And the other is system optimization. All the routing, KV caching, and orchestration across the big cluster of compute and so on." "We have a very strong internal team working on inference. But we felt that we needed to move faster, bring more capabilities. Because the market is so fast."

$NBIS cofounder Roman Chernin describes how their recent acquisitions of Eigen AI and Clarifai were all about speed, incredible talent, and acceleration: "The philosophy is very simple. We need to build so many things, and we need to move so fast, that we're always looking for people who can accelerate us. It should be exceptional talent, and/or something that has a great adoption." "Our two recent acquisitions [were] two teams that work on inference optimization. A big part of our business is how efficiently we convert GPUs into tokens. And these two teams — Eigen AI and Clarifai — one is focused on model optimization, the engine of inference. How you run specific models and all the techniques around spec decoding, quantization, and so on." "And the other is system optimization. All the routing, KV caching, and orchestration across the big cluster of compute and so on." "We have a very strong internal team working on inference. But we felt that we needed to move faster, bring more capabilities. Because the market is so fast."

TBPN

30,608 views • 4 months ago

Agents let us build systems for different workloads and requirements. But… can we trust what they build? We release 🌟SkySynth🌟: an engine for synthesizing high-performance, just-in-time (JIT) systems we can trust, by co-evolving formal proofs and tests alongside the code. Results: 💿 KV stores up to 2.3× faster than Redis and FASTER + formally verified stores with 2.9× Claude Code's pass rate 🚏 Model routers up to 48% cheaper than a general router ⚡ Specialized inference engine with 2.2× the throughput of vLLM/SGLang 🧵👇

Agents let us build systems for different workloads and requirements. But… can we trust what they build? We release 🌟SkySynth🌟: an engine for synthesizing high-performance, just-in-time (JIT) systems we can trust, by co-evolving formal proofs and tests alongside the code. Results: 💿 KV stores up to 2.3× faster than Redis and FASTER + formally verified stores with 2.9× Claude Code's pass rate 🚏 Model routers up to 48% cheaper than a general router ⚡ Specialized inference engine with 2.2× the throughput of vLLM/SGLang 🧵👇

Shu Lynn Liu

56,366 views • 5 days ago

This Jingle Jam, WE NEED YOUR HELP! We're not the only ones that can fundraise for Jingle Jam anymore - anyone can! All you need to do is sign up. Takes five minutes. What you waiting for? Join us, stream and fundraise alongside us, and help us make this Jingle Jam the best yet!

This Jingle Jam, WE NEED YOUR HELP! We're not the only ones that can fundraise for Jingle Jam anymore - anyone can! All you need to do is sign up. Takes five minutes. What you waiting for? Join us, stream and fundraise alongside us, and help us make this Jingle Jam the best yet!

The Yogscast

37,311 views • 10 months ago

we made distributed inference verifiable with <1% overhead. verification is critical for any distributed system. in a trustless network, actors may swap your 70B model for a cheaper 8B one to cut costs. until now, maintaining inference integrity meant either doubling your cost (redundancy) or exploding your latency (zkp). we created veri: an on-chain verification layer light enough for high-throughput frameworks like Parallax. it hits the economic sweet spot through architectural elegance: 1. commit-sample-verify we don't prove every step; we check a random slice using game theory. workers commit to their work before the audit. cheating becomes statistically irrational, allowing a 1% sample to secure the entire sequence. 2. simultaneous execution inference and verification happen simultaneously on the same worker pool. we don't need a separate "verifier set", so compute utilization stays high. find out more about the architecture and benchmarks: paper: blog:

we made distributed inference verifiable with <1% overhead. verification is critical for any distributed system. in a trustless network, actors may swap your 70B model for a cheaper 8B one to cut costs. until now, maintaining inference integrity meant either doubling your cost (redundancy) or exploding your latency (zkp). we created veri: an on-chain verification layer light enough for high-throughput frameworks like Parallax. it hits the economic sweet spot through architectural elegance: 1. commit-sample-verify we don't prove every step; we check a random slice using game theory. workers commit to their work before the audit. cheating becomes statistically irrational, allowing a 1% sample to secure the entire sequence. 2. simultaneous execution inference and verification happen simultaneously on the same worker pool. we don't need a separate "verifier set", so compute utilization stays high. find out more about the architecture and benchmarks: paper: blog:

Parallax

28,496 views • 8 months ago

How do you go five years without running a basic unit test on hardware random number generation? In this clip from Roundtable_023, we break down the bizarre logic behind a recent firmware fix and why fundamental testing seems to have been overlooked. Hear the full context below.

How do you go five years without running a basic unit test on hardware random number generation? In this clip from Roundtable_023, we break down the bizarre logic behind a recent firmware fix and why fundamental testing seems to have been overlooked. Hear the full context below.

Guy Swann

16,046 views • 1 month ago