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.

GLM-5.2 8bit running on two M3 Ultra 512GB with MLX distributed? Here it is! 🚀 Decode speed: 17.9 tokens/sec 🔥 Memory used: ~ 760GB 👀 Again keep in mind it's a preliminary PR by super Pedro Cuenca still a WIP!

Ivan Fioravanti ᯅ

39,669 subscribers

34,140 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

MLX GLM 5.2 Distributed on two M3 Ultra 512GB 🔥 One M3 Ultra: 18.8 tokens/sec Two M3 Ultra: 23.4 tokens/sec Context: - PR by Pedro Cuenca is still open and probably there is room for improvement: - basic generation test to measure decoding performance here, I will do a full context benchmarking once PR is more mature - nvfp4 quantization used - Video alternates standard speed and x20, with one Mac first and distributed later. Enjoy! 🙌🏻

MLX GLM 5.2 Distributed on two M3 Ultra 512GB 🔥 One M3 Ultra: 18.8 tokens/sec Two M3 Ultra: 23.4 tokens/sec Context: - PR by Pedro Cuenca is still open and probably there is room for improvement: - basic generation test to measure decoding performance here, I will do a full context benchmarking once PR is more mature - nvfp4 quantization used - Video alternates standard speed and x20, with one Mac first and distributed later. Enjoy! 🙌🏻

Ivan Fioravanti ᯅ

87,805 views • 1 month ago

GLM-4.7-8bit (350GB) running at 19 toks/s on two M3 Ultra 512GB using Tensor Parallelism with EXO - MLX, versus 14 toks/s with single node. 🚀 Now context benchmarking & then OpenCode tests 🔥 Note: this is from sources, I had to change things to run it.

GLM-4.7-8bit (350GB) running at 19 toks/s on two M3 Ultra 512GB using Tensor Parallelism with EXO - MLX, versus 14 toks/s with single node. 🚀 Now context benchmarking & then OpenCode tests 🔥 Note: this is from sources, I had to change things to run it.

Ivan Fioravanti ᯅ

327,687 views • 6 months ago

GLM-5 runs with mlx-lm on a single 512GB M3 Ultra in Q4. It's quite good in my initial testing and pretty fast as well. It generated a highly functional space invaders game using 7.1k tokens at 15.4 tok/s and 419GB memory. Thanks to Gökdeniz Gülmez and Tarjei Mandt for the port.

GLM-5 runs with mlx-lm on a single 512GB M3 Ultra in Q4. It's quite good in my initial testing and pretty fast as well. It generated a highly functional space invaders game using 7.1k tokens at 15.4 tok/s and 419GB memory. Thanks to Gökdeniz Gülmez and Tarjei Mandt for the port.

Awni Hannun

60,599 views • 5 months ago

I did it! It works! Using GLM-4.7-4bit with mlx_lm.server and opencode to fix real code locally! 🔥 Here single M3 Ultra 512GB, nex step phase will be 2 using Tensor Parallelism and then apply same changes to exo. Prefill is slow on a single machine, but generation is good.

I did it! It works! Using GLM-4.7-4bit with mlx_lm.server and opencode to fix real code locally! 🔥 Here single M3 Ultra 512GB, nex step phase will be 2 using Tensor Parallelism and then apply same changes to exo. Prefill is slow on a single machine, but generation is good.

Ivan Fioravanti ᯅ

44,000 views • 6 months ago

GLM 4.6 runs quite fast on an M3 Ultra with mlx-lm even at higher precision. Pretty remarkable that it benchmarks competitive to the just-released Sonnet 4.5. Hope those benchmarks hold-up in day-to-day use. Here's a run using 5.5 bpw quantized model, generating 5.3k tokens at 17+ tok/sec using 244 GB. What prompts should I test?

GLM 4.6 runs quite fast on an M3 Ultra with mlx-lm even at higher precision. Pretty remarkable that it benchmarks competitive to the just-released Sonnet 4.5. Hope those benchmarks hold-up in day-to-day use. Here's a run using 5.5 bpw quantized model, generating 5.3k tokens at 17+ tok/sec using 244 GB. What prompts should I test?

Awni Hannun

68,539 views • 9 months ago

Linear scaling achieved with multiple DeepSeek v3.1 instances. 4x macs = 4x throughput. 2x M3 Ultra Mac Studios = 1x DeepSeek @ 14 tok/sec 4x M3 Ultra Mac Studios = 2x DeepSeek @ 28 tok/sec DeepSeek V3.1 is a 671B parameter model - so at its native 8-bit quantization, it requires ~700GB of memory to run the model. EXO puts half of the layers on each device, combining their memory. EXO uses MLX distributed with TB5 interconnect, optimized for Apple Silicon. If we need higher throughput, adding two more devices lets us serve more users at once. EXO Labs handles all of this seamlessly - adding more devices to the cluster for linear scaling as we need it. The new EXO 1.0 will be open-source soonTM

Linear scaling achieved with multiple DeepSeek v3.1 instances. 4x macs = 4x throughput. 2x M3 Ultra Mac Studios = 1x DeepSeek @ 14 tok/sec 4x M3 Ultra Mac Studios = 2x DeepSeek @ 28 tok/sec DeepSeek V3.1 is a 671B parameter model - so at its native 8-bit quantization, it requires ~700GB of memory to run the model. EXO puts half of the layers on each device, combining their memory. EXO uses MLX distributed with TB5 interconnect, optimized for Apple Silicon. If we need higher throughput, adding two more devices lets us serve more users at once. EXO Labs handles all of this seamlessly - adding more devices to the cluster for linear scaling as we need it. The new EXO 1.0 will be open-source soonTM

Matt Beton

158,485 views • 11 months ago

1,000 tok/s, SWE-1.7 is near-frontier intelligence at real-time speed. Building this way is a new category of DX. A task that used to justify walking away finishes before you've mentally moved on. There's a weird middle mode now: it's technically async, but fast enough that you just... watch. In this example I'm running it head to head with GLM-5.2 Max. • SWE-1.7 is now free in Devin CLI or Desktop • SWE-1.7 Fast (Lightning mode) is $2.5 / MTok In · $12.5 / MTok Out

1,000 tok/s, SWE-1.7 is near-frontier intelligence at real-time speed. Building this way is a new category of DX. A task that used to justify walking away finishes before you've mentally moved on. There's a weird middle mode now: it's technically async, but fast enough that you just... watch. In this example I'm running it head to head with GLM-5.2 Max. • SWE-1.7 is now free in Devin CLI or Desktop • SWE-1.7 Fast (Lightning mode) is $2.5 / MTok In · $12.5 / MTok Out

nader dabit

175,485 views • 15 days ago

Memory Skill for OpenClaw with 26k+ users in 1 week🚀 OpenClaw's memory system is broken by default. It requires curating massive MEMORY.md files or relying on duplicate-heavy generation. Hours are wasted tuning, and massive amounts of tokens are burned. It's time to stop. So we built the memory skill to solve that prob Here is our superpower ⚔️ 🎯 Top #1 market accuracy (92.19%) after 8+ months of intense architecture iteration 🧠 The ultimate solution to keep the timeline, facts, and meaning perfectly in place ☁️ Local & Cloud + Version control ⚡ Super easy setup

Memory Skill for OpenClaw with 26k+ users in 1 week🚀 OpenClaw's memory system is broken by default. It requires curating massive MEMORY.md files or relying on duplicate-heavy generation. Hours are wasted tuning, and massive amounts of tokens are burned. It's time to stop. So we built the memory skill to solve that prob Here is our superpower ⚔️ 🎯 Top #1 market accuracy (92.19%) after 8+ months of intense architecture iteration 🧠 The ultimate solution to keep the timeline, facts, and meaning perfectly in place ☁️ Local & Cloud + Version control ⚡ Super easy setup

andy nguyen

588,887 views • 4 months ago

Next mlx-vlm release will ship with continuous batching support on the server 🚀 What's coming: → Continuous batching — new requests join the active batch immediately, no waiting. Mixed image + text batches supported → OpenAI-compatible API — field-for-field match with mlx-lm, reasoning/content split for thinking models, tag-aware streaming → Multi-turn tool calling — full tool use support across streaming and non-streaming, works with Gemma4 and other templates → Vision feature caching — cache image embeddings across turns. Gemma4: 228x speedup, Qwen3.5: 23x on cache hit All running locally on Apple Silicon. Check our this demo running 4 concurrent requests (mixed image + text) to gemma-4-26B-A4B-IT by Google Gemma in bf16 using Pi + MLX-VLM server on my M3 Ultra. One of the requests ingests a 8K resolution image!

Next mlx-vlm release will ship with continuous batching support on the server 🚀 What's coming: → Continuous batching — new requests join the active batch immediately, no waiting. Mixed image + text batches supported → OpenAI-compatible API — field-for-field match with mlx-lm, reasoning/content split for thinking models, tag-aware streaming → Multi-turn tool calling — full tool use support across streaming and non-streaming, works with Gemma4 and other templates → Vision feature caching — cache image embeddings across turns. Gemma4: 228x speedup, Qwen3.5: 23x on cache hit All running locally on Apple Silicon. Check our this demo running 4 concurrent requests (mixed image + text) to gemma-4-26B-A4B-IT by Google Gemma in bf16 using Pi + MLX-VLM server on my M3 Ultra. One of the requests ingests a 8K resolution image!

Prince Canuma

82,169 views • 3 months ago

timelapse #162 (17 hrs) - Idk what it is, grok makes it super easy to get into flow state. No model does this for me - Went deep into spec decoding architecture learning with grok - Continued Minecraft reimplementation in C. The biggest bottleneck is samples to verify against (capturing my own gameplay for a few mins rn) as well as fly wheel speed for an agent to iterate. was bottlenecked by replay and compile, now bottlenecked by tokens. this is the hardest task ive given an llm to date. - Contract work - Getting tinygrad’s glm 5.2 at c=1 to be stable - more meetings

timelapse #162 (17 hrs) - Idk what it is, grok makes it super easy to get into flow state. No model does this for me - Went deep into spec decoding architecture learning with grok - Continued Minecraft reimplementation in C. The biggest bottleneck is samples to verify against (capturing my own gameplay for a few mins rn) as well as fly wheel speed for an agent to iterate. was bottlenecked by replay and compile, now bottlenecked by tokens. this is the hardest task ive given an llm to date. - Contract work - Getting tinygrad’s glm 5.2 at c=1 to be stable - more meetings

Elliot Arledge

47,900 views • 12 days ago

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

Avi Chawla

268,010 views • 26 days ago

"In my mind, it's still second place" - Justin Gaethje says he isn’t satisfied with being interim champion 😲👀 “I’m not satisfied because it is, in my mind, still second place. However, I do know it’s a great achievement. In history, I’m gonna be a two-time interim champion, which on paper makes me a champion. In this one specific area, I have to be pessimistic and not be content because, obviously, same as last time.” (Via Justin Gaethje 🇺🇸 )

"In my mind, it's still second place" - Justin Gaethje says he isn’t satisfied with being interim champion 😲👀 “I’m not satisfied because it is, in my mind, still second place. However, I do know it’s a great achievement. In history, I’m gonna be a two-time interim champion, which on paper makes me a champion. In this one specific area, I have to be pessimistic and not be content because, obviously, same as last time.” (Via Justin Gaethje 🇺🇸 )

Red Corner MMA

47,703 views • 5 months ago

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

PowerInfer - a high-speed inference engine for deploying LLMs locally. Just came across this super interesting project on speeding up inference. It's not MoE but it's a simple approach that exploits the high locality in LLM inference to design a GPU-CPU hybrid inference engine. Hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons (the majority) are computed on the CPU. This approach significantly reduces GPU memory demands and CPU-GPU data transfer. It achieves an average token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs on a single NVIDIA RTX 4090 GPU. It's on only 18% lower than that achieved by a top-tier server-grade A100 GPU. It also significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy. There is a lot more innovation around inference that's coming fast. Really encouraged by the study on sparse computation to enhance the computational efficiency of LLMs. It's now possible to use PowerInfer with Llama 2 and Faclon 40B. Mistral-7B support is coming soon!

elvis

261,622 views • 2 years ago

Starting off 2026 with a massive new drop. 🔊Turn up the sound. 👀Watch the video. Introducing MECHFX - ultra premium dice and holographic tokens compatible with the Gundam Card Game. This is my most ambitious project yet and DuelSpec’s first #Kickstarter. There are only 20 Super Early Pilot tiers and it's already going quickly. Have a store? We have bulk tiers available for vendors as well. 🎁 GIVEAWAY: Follow, Like & Repost your favorite Gundam for a chance to win a MECHFX set FREE! 👉 Back it on KICKSTARTER here: 🔗

Starting off 2026 with a massive new drop. 🔊Turn up the sound. 👀Watch the video. Introducing MECHFX - ultra premium dice and holographic tokens compatible with the Gundam Card Game. This is my most ambitious project yet and DuelSpec’s first #Kickstarter. There are only 20 Super Early Pilot tiers and it's already going quickly. Have a store? We have bulk tiers available for vendors as well. 🎁 GIVEAWAY: Follow, Like & Repost your favorite Gundam for a chance to win a MECHFX set FREE! 👉 Back it on KICKSTARTER here: 🔗

DuelSpec

23,583 views • 6 months ago

At 1AM, we never stop cooking 🧑‍🍳 Full demo of our experimental NFT + token launch feature, live in beta 6.0.0. Sign up with a new wallet → sync it → launch tokens and NFTs on Midnight. In the time most wallets take just to sync, you've already minted. 🔥 All contracts used here are powered by the awesome work by NMKR Look, feel, experience and manage your NFT and token assets on Midnight right inside your 1AM Wallet 🚀 There is a cool Easter Egg in the demo if you watch closely 👀 What do you guys think about this new update? Let's bring Midnight to the masses 🖤

At 1AM, we never stop cooking 🧑‍🍳 Full demo of our experimental NFT + token launch feature, live in beta 6.0.0. Sign up with a new wallet → sync it → launch tokens and NFTs on Midnight. In the time most wallets take just to sync, you've already minted. 🔥 All contracts used here are powered by the awesome work by NMKR Look, feel, experience and manage your NFT and token assets on Midnight right inside your 1AM Wallet 🚀 There is a cool Easter Egg in the demo if you watch closely 👀 What do you guys think about this new update? Let's bring Midnight to the masses 🖤

OneAm

86,564 views • 3 months ago

New video of Qualcomm CEO Cristiano Amon: AI will require “gazillions” of tokens. Because, Agentic AI will consume dramatically more tokens because it performs autonomous tasks, uses multiple systems, and interacts with tools. AI demand will grow hugely when software starts letting agents act, not just answer. A chatbot spends tokens on language; an agent spends tokens on deciding, checking, calling tools, reading outputs, revising plans, and coordinating with other software. Today a single human-AI exchange may be large, a reasoning task may be much larger, but we are already entering the agentic era, where an autonomous workflow can become exponentially larger still because the model is no longer producing one response. It is running a process. When a SaaS product redesigns itself around human-agent interaction, every task can become a chain of hidden micro-decisions, and each micro-decision consumes context, memory, tool calls, verification, and output tokens. Another point is people hear “more tokens” and think only of bigger data centers, when the deeper change is economic: software usage may stop being measured mainly by clicks, seats, or sessions. It will be measured by how much machine reasoning/tokens gets spent on behalf of each user. ---- From "Reuters" YouTube channel, (link in comment)

New video of Qualcomm CEO Cristiano Amon: AI will require “gazillions” of tokens. Because, Agentic AI will consume dramatically more tokens because it performs autonomous tasks, uses multiple systems, and interacts with tools. AI demand will grow hugely when software starts letting agents act, not just answer. A chatbot spends tokens on language; an agent spends tokens on deciding, checking, calling tools, reading outputs, revising plans, and coordinating with other software. Today a single human-AI exchange may be large, a reasoning task may be much larger, but we are already entering the agentic era, where an autonomous workflow can become exponentially larger still because the model is no longer producing one response. It is running a process. When a SaaS product redesigns itself around human-agent interaction, every task can become a chain of hidden micro-decisions, and each micro-decision consumes context, memory, tool calls, verification, and output tokens. Another point is people hear “more tokens” and think only of bigger data centers, when the deeper change is economic: software usage may stop being measured mainly by clicks, seats, or sessions. It will be measured by how much machine reasoning/tokens gets spent on behalf of each user. ---- From "Reuters" YouTube channel, (link in comment)

Rohan Paul

31,617 views • 1 month ago

"I love Ukraine so much and it still breaks my heart with what's going on. This war is cruel. It's pointless. It's gone on for far too long and it needs to end soon. I'm still inspired by Ukraine's resilience as I was two years ago. But this has to stop. It's time to consolidate all our efforts and put an end to this inhumane war. Hold on, my friends. We're still with you. We love you and we're here for you." Bring Me The Horizon’s Bring Me The Horizon Oli Sykes Oliver Sykes has recorded a message of support to Ukrainians. Thank you, Oli💙💛 We look forward to seeing you again in Ukraine!

"I love Ukraine so much and it still breaks my heart with what's going on. This war is cruel. It's pointless. It's gone on for far too long and it needs to end soon. I'm still inspired by Ukraine's resilience as I was two years ago. But this has to stop. It's time to consolidate all our efforts and put an end to this inhumane war. Hold on, my friends. We're still with you. We love you and we're here for you." Bring Me The Horizon’s Bring Me The Horizon Oli Sykes Oliver Sykes has recorded a message of support to Ukrainians. Thank you, Oli💙💛 We look forward to seeing you again in Ukraine!

UNITED24

56,518 views • 2 years ago

I'm running Llama 4 Maverick at 620 t/s! I'm living in the future! Honestly, a large language model running this fast is something straight out of a sci-fi movie. Speeds like this will enable a whole new world of applications that aren't possible today. For reference, GPT-4o, which is probably the most popular OpenAI model, runs between 60 and 110 t/s. The secret here: I'm not running AI at Meta's Llama 4 Maverick on a GPU. I'm using the SambaNova Cloud (my sponsor) and their custom SN40L chips. They are optimized from the ground up for running AI workflows. Right now, SambaNova Cloud runs DeepSeek, Qwen, Whisper, and the entire family of Llama models on these chips. You can check the speed of each of these models using SambaNova Cloud's Playground (see the attached video). It's completely free, and that's how I'm measuring their speeds. For example, I also tried DeepSeek R1 (the latest version from May) and, oh boy! DeepSeek R1 is a huge 671B parameter model. It's probably the best open reasoning model in the world, and it runs at 140 tokens per second! !!! Inference time on an SN40L is night and day from what you'll get from a GPU. Here is why this is big: If you are running an agentic workflow that uses multiple models simultaneously on a GPU, it will need to swap models in and out of memory (because not every model fits). A single SNL40 chip can simultaneously hold over 100 models (trillions of parameters) in memory. If you are using open models, try the SambaCloud API to see what lightning speed looks like. Here is how: 1. Create a free account at: 2. Check the QuickStart guide: If you try the playground, check the speed you're getting with Llama 4 and DeepSeek, and post the results below. I've seen much higher numbers than I posted here, so I'm curious to see whether geography affects the speed.

I'm running Llama 4 Maverick at 620 t/s! I'm living in the future! Honestly, a large language model running this fast is something straight out of a sci-fi movie. Speeds like this will enable a whole new world of applications that aren't possible today. For reference, GPT-4o, which is probably the most popular OpenAI model, runs between 60 and 110 t/s. The secret here: I'm not running AI at Meta's Llama 4 Maverick on a GPU. I'm using the SambaNova Cloud (my sponsor) and their custom SN40L chips. They are optimized from the ground up for running AI workflows. Right now, SambaNova Cloud runs DeepSeek, Qwen, Whisper, and the entire family of Llama models on these chips. You can check the speed of each of these models using SambaNova Cloud's Playground (see the attached video). It's completely free, and that's how I'm measuring their speeds. For example, I also tried DeepSeek R1 (the latest version from May) and, oh boy! DeepSeek R1 is a huge 671B parameter model. It's probably the best open reasoning model in the world, and it runs at 140 tokens per second! !!! Inference time on an SN40L is night and day from what you'll get from a GPU. Here is why this is big: If you are running an agentic workflow that uses multiple models simultaneously on a GPU, it will need to swap models in and out of memory (because not every model fits). A single SNL40 chip can simultaneously hold over 100 models (trillions of parameters) in memory. If you are using open models, try the SambaCloud API to see what lightning speed looks like. Here is how: 1. Create a free account at: 2. Check the QuickStart guide: If you try the playground, check the speed you're getting with Llama 4 and DeepSeek, and post the results below. I've seen much higher numbers than I posted here, so I'm curious to see whether geography affects the speed.

Santiago

34,148 views • 1 year ago

I had to test it myself to believe this unreal inference speed. 3,000 tokens/s for 1 user on standard datacenter GPUs. They leveraged a hidden efficiency gap in how GPUs generate tokens. Kog just achieved 3,000 tokens/s on 8× AMD MI300X GPUs and 2,100 on 8× NVIDIA H200 (FP16, no speculative decoding). Their tech preview is on a 2B model, and they show how their techniques will scale to large frontier MoE models at similar speeds. That's a huge number because normal low-batch GPU decoding for 2B to 8B models is usually closer to 100 to 300 tokens/s per request, so Kog is claiming something like a 10X to 30X jump in the speed one user actually feels. Their trick: they are getting the speed by treating LLM decoding as a memory streaming problem, not mainly a math problem. For 1 user at batch size 1, the GPU is not doing big, efficient matrix-matrix work like in training or large-batch serving; it is repeatedly pulling the model’s active weights from high-bandwidth memory for each new token, so speed depends on how smoothly those weights keep flowing. Normal inference stacks keep breaking that flow. They run many separate GPU programs for different parts of the model, move intermediate results through memory, wait at synchronization points, talk back to the CPU for scheduling or sampling, and then repeat this token after token. Kog’s answer is to co-design 3 things that are usually tuned separately: the runtime, the low-level GPU code, and the model architecture. The biggest engineering move is the monokernel, where the whole decode pass runs as 1 persistent GPU-resident program, including sampling, so the system does not keep stopping for kernel launches, CPU scheduling, and intermediate memory round trips. They also rebuilt synchronization, because their own measurements say grid sync was eating around 35% of token-generation time; instead of making every compute unit wait at a broad barrier, each unit waits only for the exact data it needs. On AMD MI300X, they also map memory access around the chiplet layout, because memory latency changes depending on which die makes the request. Then their Laneformer model uses Delayed Tensor Parallelism, which lets cross-GPU communication happen in the background instead of blocking every layer.

I had to test it myself to believe this unreal inference speed. 3,000 tokens/s for 1 user on standard datacenter GPUs. They leveraged a hidden efficiency gap in how GPUs generate tokens. Kog just achieved 3,000 tokens/s on 8× AMD MI300X GPUs and 2,100 on 8× NVIDIA H200 (FP16, no speculative decoding). Their tech preview is on a 2B model, and they show how their techniques will scale to large frontier MoE models at similar speeds. That's a huge number because normal low-batch GPU decoding for 2B to 8B models is usually closer to 100 to 300 tokens/s per request, so Kog is claiming something like a 10X to 30X jump in the speed one user actually feels. Their trick: they are getting the speed by treating LLM decoding as a memory streaming problem, not mainly a math problem. For 1 user at batch size 1, the GPU is not doing big, efficient matrix-matrix work like in training or large-batch serving; it is repeatedly pulling the model’s active weights from high-bandwidth memory for each new token, so speed depends on how smoothly those weights keep flowing. Normal inference stacks keep breaking that flow. They run many separate GPU programs for different parts of the model, move intermediate results through memory, wait at synchronization points, talk back to the CPU for scheduling or sampling, and then repeat this token after token. Kog’s answer is to co-design 3 things that are usually tuned separately: the runtime, the low-level GPU code, and the model architecture. The biggest engineering move is the monokernel, where the whole decode pass runs as 1 persistent GPU-resident program, including sampling, so the system does not keep stopping for kernel launches, CPU scheduling, and intermediate memory round trips. They also rebuilt synchronization, because their own measurements say grid sync was eating around 35% of token-generation time; instead of making every compute unit wait at a broad barrier, each unit waits only for the exact data it needs. On AMD MI300X, they also map memory access around the chiplet layout, because memory latency changes depending on which die makes the request. Then their Laneformer model uses Delayed Tensor Parallelism, which lets cross-GPU communication happen in the background instead of blocking every layer.

Rohan Paul

13,148 views • 1 month ago