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.

Great visual of NVIDIA’s Inference Context Memory architecture. KV cache offload to NVMe via BlueField-4 DPU this is how Rubin handles massive context windows at scale. Similar to Amazon’s approach but with full-stack hardware integration. H/T 程工

Ben Pouladian

14,121 subscribers

54,335 views • 6 months ago •via X (Twitter)

News & Politics Education Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

I added KV caching and INT8 KV quantization to our transformer inference, improving throughput by 35x. All of this was done from scratch in Rust + CUDA, on top of a homemade ML framework. On a 4-token prompt with 252 generated tokens: - Original: 0.76 tok/s - KV cache fp32: 27.21 tok/s - KV cache int8 (quantized): 27.29 tok/s Try it out yourself here: In practice: - KV caching gave us about a 35x end-to-end speedup - INT8 KV cache kept roughly the same speed as fp32 but cut KV cache memory by 3.78x FP32 cache used 4.5 MB in this run while the INT8 cache used only 1.19 MB This simple change to inference created a huge impact on performance. To learn more about the KV cache and other optimizations like this, check out the blog at

Reese Chong

52,588 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

76,394 views • 6 months ago

✨ The #NVIDIARubin platform is engineered to master multi-step problem-solving and massive long-context reasoning workflows at scale. Built using extreme co-design, Rubin delivers more AI tokens per watt and lower cost per token compared to the previous architecture. ✔️ 50 PFLOPS of NVFP4 inference ✔️ 22 TB/s of HBM4 bandwidth ✔️ 3.6 TB/s of NVIDIA NVLink bandwidth per GPU Learn more about NVIDIA Rubin ➡️

✨ The #NVIDIARubin platform is engineered to master multi-step problem-solving and massive long-context reasoning workflows at scale. Built using extreme co-design, Rubin delivers more AI tokens per watt and lower cost per token compared to the previous architecture. ✔️ 50 PFLOPS of NVFP4 inference ✔️ 22 TB/s of HBM4 bandwidth ✔️ 3.6 TB/s of NVIDIA NVLink bandwidth per GPU Learn more about NVIDIA Rubin ➡️

NVIDIA AI Infrastructure

10,475 views • 5 months ago

Introducing HydraDB. The graph native context infrastructure for agents. Purpose built to deliver precise context & observability into why agents act the way they do. We've always believed graphs are the best way to manage AI context, but they've been too expensive to scale or impractical for storing full context. Until now. HydraDB combines in memory, NVMe, and object storage into a single graph layer, making context delivery faster, cheaper, and more precise. We want context delivery to be extremely fast, 1000x cheap, and highly precise. Give your agents a brain.

Introducing HydraDB. The graph native context infrastructure for agents. Purpose built to deliver precise context & observability into why agents act the way they do. We've always believed graphs are the best way to manage AI context, but they've been too expensive to scale or impractical for storing full context. Until now. HydraDB combines in memory, NVMe, and object storage into a single graph layer, making context delivery faster, cheaper, and more precise. We want context delivery to be extremely fast, 1000x cheap, and highly precise. Give your agents a brain.

Nishkarsh

2,292,880 views • 1 month ago

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

SITUATION EXPLAINED: Cerebras raised $5.55 billion in their IPO and closing their first day of trading valued at $66 billion, making it the biggest US tech IPO since Snowflake in 2020. Cerebras makes Wafer-Scale Engine chips built for AI inference. We asked Sarah Fong the main difference between wafer-scale chips and traditional GPUs: - GPUs are great at parallel work (graphics, training) - AI inference is sequential, AKA one token at a time This causes the "memory wall" problem: - Every GPU core needs model weights, KV cache, and activations to do its math - On a GPU, that data lives in off-chip memory (HBM) - Cores constantly load and offload from off-chip memory, which is a huge bottleneck; hardware accounts for ~70% of inference latency Cerebras' chips: -Dinner-plate sized (vs. GPUs which are palm-sized) with tens of thousands of cores -Memory sits directly on top of the cores as distributed SRAM -Weights and KV cache can be accessed at on-chip speeds in the PB/s range, compared with off-chip speeds in the TB/s range achieved by GPUs with HBM.

MTS

44,057 views • 1 month ago

DeepSeek-V4 dropped. 1M context. 10x smaller KV cache. First open model where the context window and the agentic post-training meet.

DeepSeek-V4 dropped. 1M context. 10x smaller KV cache. First open model where the context window and the agentic post-training meet.

Ben Burtenshaw

49,900 views • 2 months ago

Introducing DuoAttention: Our new framework slashes both memory and latency for long-context LLMs without sacrificing performance! By applying full KV cache only to critical heads, we achieve: ⚡ 2.55x memory reduction ⚡ 2.18x decoding speedup ⚡ 3.3M tokens on a single A100 GPU

Introducing DuoAttention: Our new framework slashes both memory and latency for long-context LLMs without sacrificing performance! By applying full KV cache only to critical heads, we achieve: ⚡ 2.55x memory reduction ⚡ 2.18x decoding speedup ⚡ 3.3M tokens on a single A100 GPU

Guangxuan Xiao

31,023 views • 1 year ago

It's not only about how long your context is, but how well you use it. Great to see Gemini 2.5 models dominating MRCR and other benchmarks on long context! See 2.5 Pro tackle a complex coding task by reasoning over an entire repo (>500k tokens). Performance and effective use of the (loooong) context windows are what really matter!

It's not only about how long your context is, but how well you use it. Great to see Gemini 2.5 models dominating MRCR and other benchmarks on long context! See 2.5 Pro tackle a complex coding task by reasoning over an entire repo (>500k tokens). Performance and effective use of the (loooong) context windows are what really matter!

Oriol Vinyals

27,008 views • 1 year ago

This was a great read. So I started building an open-source data agent with 6 layers of context: 1. Table usage 2. Human annotations 3. Query patterns 4. Institutional knowledge 5. Memory 6. Runtime context Early version if you want to follow along:

This was a great read. So I started building an open-source data agent with 6 layers of context: 1. Table usage 2. Human annotations 3. Query patterns 4. Institutional knowledge 5. Memory 6. Runtime context Early version if you want to follow along:

Ashpreet Bedi

51,251 views • 5 months ago

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.

nvidia's 3B mamba destroyed alibaba's 3B deltanet on the same RTX 3090. only 24 days between releases. same active parameters, same VRAM tier, completely different architectures. nemotron cascade 2: 187 tok/s. flat from 4K to 625K context. zero speed loss. flags: -ngl 99 -np 1. that's it. no context flags, no KV cache tricks. auto-allocates 625K. qwen 3.5 35B-A3B: 112 tok/s. flat from 4K to 262K context. zero speed loss. flags: -ngl 99 -np 1 -c 262144 --cache-type-k q8_0 --cache-type-v q8_0. needed KV cache quantization to fit 262K. both models held a flat line across every context level. both architectures are context-independent. but nvidia's mamba2 is 67% faster at generating tokens on the exact same hardware and needs fewer flags to get there. same node, same GPU, same everything. the only variable is the model. gold medal math olympiad winner running at 187 tokens per second on single RTX 3090 a card from 6 years ago. nvidia cooked.

Sudo su

186,573 views • 3 months ago

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

New short course: LLMs as Operating Systems: Agent Memory, created with Letta, and taught by its founders Charles Packer and Sarah Wooders. An LLM's input context window has limited space. Using a longer input context also costs more and results in slower processing. So, managing what's stored in this context window is important. In the innovative paper MemGPT: Towards LLMs as Operating Systems, its authors (which include the instructors) proposed using an LLM agent to manage this context window. Their system uses a large persistent memory that stores everything that could be included in the input context, and an agent decides what is actually included. Take the example of building a chatbot that needs to remember what's been said earlier in a conversation (perhaps over many days of interaction with a user). As the conversation's length grows, the memory management agent will move information from the input context to a persistent searchable database; summarize information to keep relevant facts in the input context; and restore relevant conversation elements from further back in time. This allows a chatbot to keep what's currently most relevant in its input context memory to generate the next response. When I read the original MemGPT paper, I thought it was an innovative technique for handling memory for LLMs. The open-source Letta framework, which we'll use in this course, makes MemGPT easy to implement. It adds memory to your LLM agents and gives them transparent long-term memory. In detail, you’ll learn: - How to build an agent that can edit its own limited input context memory, using tools and multi-step reasoning - What is a memory hierarchy (an idea from computer operating systems, which use a cache to speed up memory access), and how these ideas apply to managing the LLM input context (where the input context window is a "cache" storing the most relevant information; and an agent decides what to move in and out of this to/from a larger persistent storage system) - How to implement multi-agent collaboration by letting different agents share blocks of memory This course will give you a sophisticated understanding of memory management for LLMs, which is important for chatbots having long conversations, and for complex agentic workflows. Please sign up here!

Andrew Ng

200,752 views • 1 year ago

If u this clip sounds kinda bad out of context, wait til u hear how much worse it is with full context! Link:

If u this clip sounds kinda bad out of context, wait til u hear how much worse it is with full context! Link:

Kurt Metzger

41,808 views • 6 months ago

Out of Context Human Race This is how to raise hour kids. Not with screens.

Out of Context Human Race This is how to raise hour kids. Not with screens.

Lizy Nyanjwa

124,129 views • 2 years ago

this is good example of how context implied changes over text vs video (or is deliberately taken out of context to sensationalize)

this is good example of how context implied changes over text vs video (or is deliberately taken out of context to sensationalize)

Cani

14,519 views • 4 months ago

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

New course: Efficient Inference with SGLang: Text and Image Generation, built in partnership with LMSys LMSYS Org and RadixArk RadixArk, and taught by Richard Chen Richard Chen, a Member of Technical Staff at RadixArk. Running LLMs in production is expensive, and much of that cost comes from redundant computation. This short course teaches you to eliminate that waste using SGLang, an open-source inference framework that caches computation already done and reuses it across future requests. When ten users share the same system prompt, SGLang processes it once, not ten times. The speedups compound quickly, especially when there's a lot of shared context across requests. Skills you'll gain: - Implement a KV cache from scratch to eliminate redundant computation within a single request - Scale caching across users and requests with RadixAttention, so shared context is only processed once - Accelerate image generation with diffusion models using SGLang's caching and multi-GPU parallelism Join and learn to make LLM inference faster and more cost-efficient at scale!

Andrew Ng

98,660 views • 2 months ago

A deep conversation with Nikolay Savinov, the Gemini long context pre-training co-lead… We go from the basics to what is needed to scale to infinite context to long context best practices for devs:

A deep conversation with Nikolay Savinov, the Gemini long context pre-training co-lead… We go from the basics to what is needed to scale to infinite context to long context best practices for devs:

Logan Kilpatrick

252,367 views • 1 year ago

Context memory essentially unlocks Agentic AI Much needed for Opus 4.6's "multi-agent swarms" In this SemiDoped pod, Vikram Sekar talks to Val Bercovici from Weka about context storage. - How token warehouses save inference costs - A new networking tier? Context Storage Network! - High Bandwidth Flash for context? - Weka's Augmented Memory Grid for context storage - Where this is all headed The convo is info packed. Don't miss out on it! b/acc, context platform engineer Chapters (00:00) Introduction to Weka and AI Storage Solutions (05:18) The Evolution of Context Memory in AI (09:30) Understanding Memory Hierarchies and Their Impact (16:24) Latency Challenges in Modern Storage Solutions (21:32) The Role of Networking in AI Storage Efficiency (29:42) Dynamic Resource Utilization in AI Networks (30:04) Introducing the Context Memory Network (31:13) High Bandwidth Flash: A Game Changer (32:54) Weka’s Neural Mesh and Storage Solutions (35:01) Axon: Transforming GPU Storage into Memory (39:00) Augmented Memory Grid Explained (42:00) Pooling DRAM and CXL Innovations (46:02) Token Warehouses and Inference Economics (52:10) The Future of Storage Innovations

Context memory essentially unlocks Agentic AI Much needed for Opus 4.6's "multi-agent swarms" In this SemiDoped pod, Vikram Sekar talks to Val Bercovici from Weka about context storage. - How token warehouses save inference costs - A new networking tier? Context Storage Network! - High Bandwidth Flash for context? - Weka's Augmented Memory Grid for context storage - Where this is all headed The convo is info packed. Don't miss out on it! b/acc, context platform engineer Chapters (00:00) Introduction to Weka and AI Storage Solutions (05:18) The Evolution of Context Memory in AI (09:30) Understanding Memory Hierarchies and Their Impact (16:24) Latency Challenges in Modern Storage Solutions (21:32) The Role of Networking in AI Storage Efficiency (29:42) Dynamic Resource Utilization in AI Networks (30:04) Introducing the Context Memory Network (31:13) High Bandwidth Flash: A Game Changer (32:54) Weka’s Neural Mesh and Storage Solutions (35:01) Axon: Transforming GPU Storage into Memory (39:00) Augmented Memory Grid Explained (42:00) Pooling DRAM and CXL Innovations (46:02) Token Warehouses and Inference Economics (52:10) The Future of Storage Innovations

Semi Doped

12,796 views • 5 months ago

Build Multi Agent Systems with Reasoning and Context Thanks to sonnet-4, we have level 4 autonomous multi-agent systems working. Learn how to add: -> Reasoning Tools (think -> analyze) -> Shared Agentic Context -> Agentic Memory Code below 👇

Build Multi Agent Systems with Reasoning and Context Thanks to sonnet-4, we have level 4 autonomous multi-agent systems working. Learn how to add: -> Reasoning Tools (think -> analyze) -> Shared Agentic Context -> Agentic Memory Code below 👇

Ashpreet Bedi

38,576 views • 1 year ago

tested out antirez' ds4.c this morning. so impressive and delivers. on a M3 max, 128GB, stock ds4 settings: - 14–15 t/s at 62K pre-filled actual coding conversation - memory usage was flat during gen ~85GB res - disk cache is ~8GB for a full 100K context window - thermals were normal, light fan activity - inference server is rock solid so far biggest constraint: anytime there's a compact, we pay the wait-time price of a fresh prefill (~1min per 10k context) before we are back in action. sequential inference + multiple agents in parallel performance is unclear, will report back. I'm so amped.

tested out antirez' ds4.c this morning. so impressive and delivers. on a M3 max, 128GB, stock ds4 settings: - 14–15 t/s at 62K pre-filled actual coding conversation - memory usage was flat during gen ~85GB res - disk cache is ~8GB for a full 100K context window - thermals were normal, light fan activity - inference server is rock solid so far biggest constraint: anytime there's a compact, we pay the wait-time price of a fresh prefill (~1min per 10k context) before we are back in action. sequential inference + multiple agents in parallel performance is unclear, will report back. I'm so amped.

pradeep

162,339 views • 1 month ago