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。

An H100 GPU can run 10^15 FLOPs/s. But somehow, a model like Gemini Flash-Lite is stuck at 200 tokens/s. These numbers show just how much of a bottleneck memory bandwidth is. In this video, we look at how LLMs are evolving to work around this limitation. Thanks Inception for... show more

Julia Turc

24,153 subscribers

26,110 次观看 • 6 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰

GPU tradeoff series: A100 is not much more powerful than 4090 🫠 GPU Perf and Price: - 4090: 330 fp16 TFLOPs, $1,749 - A100 (80GB): 312 fp16 TFLOPs, $20,000 > A100 is 11.4X more pricy Training speed for GPT-2(124M) with llm.c: - 4090: 153K tokens/s - A100 (80GB): 195K tokens/s > A100 is only 1.3X faster (both trained using a single card, A100 llm.c training is shown in the video, 4090 video is in the quoted tweet) Conclusion: 4090 has a much better cost vs performance ratio Why: As in the H100 vs. 4090 comparison, the biggest difference between A100 and 4090 is their GPU memory size/bandwidth and cross-GPU communication bandwidth, which does not matter too much if your model can fit into a single 4090. Specs: 4090: - GPU memory size: 24GB - memory bandwidth: 1 TB/s - communication bandwidth: 64 GB/s A100: - GPU memory size: 80GB - memory bandwidth: 2 TB/s - communication bandwidth: 900 GB/s Nvidia killed off NVLink (a high-speed communication link that connects GPUs) on 4090. (Jensen Huang smiling face) If multiple 4090s could be interconnected via NVLink, their performance would be closer to datacenter-grade A100 GPUs, even for training larger models. Additionally, 4090 isn't allowed in datacenters, that's how Nvidia makes 💰💰💰

Yuchen Jin

234,951 次观看 • 2 年前

This is the fastest I've seen Llama 3.3 running anywhere! Llama 3.3 70B running at 652 t/s is lightning fast. And if you want Llama 3.1, here are the speeds I was able to get: • Llama 3.1 8B: 1006 t/s • Llama 3.1 70B: 709 t/s • Llama 3.1 405B: 206 t/s (You can access all of these models for free! See the link below.) This speed is incredible, but the interesting part is what's happening behind the scenes: These models aren't running on a GPU! In this video, I'm using the SambaNova cloud to access these models. They built a custom chip (SN40L) optimized for AI workflows. A single SN40L chip can hold hundreds of models (trillions of parameters) in memory! The speed alone is a huge deal, but the big advantage is for agentic workflows running multiple specialized models. A GPU can only host a single model and switch (unload and load) to a different model if necessary. An SN40L, on the other hand, can host every model at once, making it much faster. Here is the video where you can see how fast these chips are:

This is the fastest I've seen Llama 3.3 running anywhere! Llama 3.3 70B running at 652 t/s is lightning fast. And if you want Llama 3.1, here are the speeds I was able to get: • Llama 3.1 8B: 1006 t/s • Llama 3.1 70B: 709 t/s • Llama 3.1 405B: 206 t/s (You can access all of these models for free! See the link below.) This speed is incredible, but the interesting part is what's happening behind the scenes: These models aren't running on a GPU! In this video, I'm using the SambaNova cloud to access these models. They built a custom chip (SN40L) optimized for AI workflows. A single SN40L chip can hold hundreds of models (trillions of parameters) in memory! The speed alone is a huge deal, but the big advantage is for agentic workflows running multiple specialized models. A GPU can only host a single model and switch (unload and load) to a different model if necessary. An SN40L, on the other hand, can host every model at once, making it much faster. Here is the video where you can see how fast these chips are:

Santiago

97,561 次观看 • 1 年前

125B on a single RTX 4090 at 25 tok/s is actually insane Qwen3.8-Flash-Next + MTP is hitting: 80K context 25.35 tok/s decode 471 tok/s prefill On one consumer GPU. Yes, it’s an MoE, so you’re not running all 125B parameters for every token. But still, getting a model this large to run this fast on a 4090 is wild. MTP is doing a lot of the heavy lifting here too. Instead of generating one token at a time, it can speculate on multiple tokens and verify them together. And 25 tok/s is the important number.

125B on a single RTX 4090 at 25 tok/s is actually insane Qwen3.8-Flash-Next + MTP is hitting: 80K context 25.35 tok/s decode 471 tok/s prefill On one consumer GPU. Yes, it’s an MoE, so you’re not running all 125B parameters for every token. But still, getting a model this large to run this fast on a 4090 is wild. MTP is doing a lot of the heavy lifting here too. Instead of generating one token at a time, it can speculate on multiple tokens and verify them together. And 25 tok/s is the important number.

FHILY👑

39,811 次观看 • 18 天前

I’m at a loss for words at how condescending this is towards women. Do the Dems think marriages are stuck in the 1950’s?

I’m at a loss for words at how condescending this is towards women. Do the Dems think marriages are stuck in the 1950’s?

Jean

22,430,835 次观看 • 1 年前

Qwen3.8-Flash-Next just got a massive speed boost on my old hardware. This guy started with roughly 11 tok/s. After tuning the inference stack, I’m now getting around 24–28 tok/s. That’s roughly a 2× speedup without changing the GPU. The setup: → RTX 3090 24GB → 64GB DDR4 system RAM → Qwen3.8-Flash-Next → Cafe-llama.cpp The interesting part is that Cafe-llama.cpp isn’t just another llama.cpp fork. The optimizations are making a real difference for this particular architecture, especially when you’re dealing with a model that relies heavily on system memory and offloading. Going from 11 tok/s to nearly 30 tok/s completely changes how practical the model feels for local use. And now I’m curious what happens with DDR5. If memory bandwidth becomes another bottleneck, a faster RAM subsystem could squeeze even more performance out of the same 3090. This is why local AI optimization is getting so interesting. Sometimes you don’t need a new GPU. You just need a better inference stack. 11 tok/s → 24–28 tok/s.

Qwen3.8-Flash-Next just got a massive speed boost on my old hardware. This guy started with roughly 11 tok/s. After tuning the inference stack, I’m now getting around 24–28 tok/s. That’s roughly a 2× speedup without changing the GPU. The setup: → RTX 3090 24GB → 64GB DDR4 system RAM → Qwen3.8-Flash-Next → Cafe-llama.cpp The interesting part is that Cafe-llama.cpp isn’t just another llama.cpp fork. The optimizations are making a real difference for this particular architecture, especially when you’re dealing with a model that relies heavily on system memory and offloading. Going from 11 tok/s to nearly 30 tok/s completely changes how practical the model feels for local use. And now I’m curious what happens with DDR5. If memory bandwidth becomes another bottleneck, a faster RAM subsystem could squeeze even more performance out of the same 3090. This is why local AI optimization is getting so interesting. Sometimes you don’t need a new GPU. You just need a better inference stack. 11 tok/s → 24–28 tok/s.

FHILY👑

23,243 次观看 • 21 天前

If we get AGI before 2030 but can only produce ~200 GW a year of chips on advanced process nodes, can we ease the bottleneck by going back to older nodes? Some of the most singularity-pilled investors think so. Most of the FLOPS improvement between chip generations has come from numerics - not transistor density. But you could just design a new version of a 7nm chip like the A100 to have FP4. Holding numerics and die area constant, a B200 (2025 chip on 4nm) would only have ~3.5x more FLOPS than an A100 (2020 chip on 7nm). That seems like a manageable haircut - does it mean we can blow past the 200 GW ceiling? Dylan Patel explains why this is not an apples-to-apples comparison. If you run real-world performance analysis on DeepSeek V3 or Kimi K2.5, the performance gap between Hopper and Blackwell isn't 3x. It's ~20x. Why? Because inference is already memory-bandwidth-bound, and older chips have far fewer FLOPS per die, requiring even more communication between chips. Remember the bandwidth hierarchy: on-chip SRAM runs at tens of TB/s, HBM at single-digit TB/s, NVLink within a 72-GPU domain at high hundreds of GB/s, and between racks at low hundreds of GB/s. Each boundary is roughly a 10x drop. With less compute per chip, you need more chips to run the same model - which means more communication across those slower pipes that are already bottlenecking you.

If we get AGI before 2030 but can only produce ~200 GW a year of chips on advanced process nodes, can we ease the bottleneck by going back to older nodes? Some of the most singularity-pilled investors think so. Most of the FLOPS improvement between chip generations has come from numerics - not transistor density. But you could just design a new version of a 7nm chip like the A100 to have FP4. Holding numerics and die area constant, a B200 (2025 chip on 4nm) would only have ~3.5x more FLOPS than an A100 (2020 chip on 7nm). That seems like a manageable haircut - does it mean we can blow past the 200 GW ceiling? Dylan Patel explains why this is not an apples-to-apples comparison. If you run real-world performance analysis on DeepSeek V3 or Kimi K2.5, the performance gap between Hopper and Blackwell isn't 3x. It's ~20x. Why? Because inference is already memory-bandwidth-bound, and older chips have far fewer FLOPS per die, requiring even more communication between chips. Remember the bandwidth hierarchy: on-chip SRAM runs at tens of TB/s, HBM at single-digit TB/s, NVLink within a 72-GPU domain at high hundreds of GB/s, and between racks at low hundreds of GB/s. Each boundary is roughly a 10x drop. With less compute per chip, you need more chips to run the same model - which means more communication across those slower pipes that are already bottlenecking you.

Dwarkesh Patel

56,362 次观看 • 5 个月前

SK Hynix's Hoshik Kim admits high bandwidth memory "is not the answer" to the memory-wall problem, and SK is working on different solutions to solve the bottleneck. "No, actually, I kind of agree with Pat. So HBM is not the answer." [SK Hynix is admitting that HBM is a bad memory?] "Nothing is a bad memory. It's a good memory, but it's not the final answer. So actually, HBM will not solve the memory-wall problem. Memory-wall problem is an inherent AI problem, which you cannot avoid." "So we try our best to provide higher bandwidth with this memory, with HBM technology even though it has some limitation, as Pat already mentioned, with thermal, IOs, cost, a lot of stuff. "But that's the only thing we got right now to help ease this memory-wall problem, memory-bottleneck problem. But at the same time, we are working in different directions to solve this memory-wall problem with different solutions."

SK Hynix's Hoshik Kim admits high bandwidth memory "is not the answer" to the memory-wall problem, and SK is working on different solutions to solve the bottleneck. "No, actually, I kind of agree with Pat. So HBM is not the answer." [SK Hynix is admitting that HBM is a bad memory?] "Nothing is a bad memory. It's a good memory, but it's not the final answer. So actually, HBM will not solve the memory-wall problem. Memory-wall problem is an inherent AI problem, which you cannot avoid." "So we try our best to provide higher bandwidth with this memory, with HBM technology even though it has some limitation, as Pat already mentioned, with thermal, IOs, cost, a lot of stuff. "But that's the only thing we got right now to help ease this memory-wall problem, memory-bottleneck problem. But at the same time, we are working in different directions to solve this memory-wall problem with different solutions."

Fireside Alpha

140,571 次观看 • 1 个月前

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

We used to buy more VRAM. Now we’re finding ways around the VRAM limit. This little DDR5 SO-DIMM setup is paired with an old NVIDIA Tesla P40 in the video. The P40 is an old datacenter GPU with 24GB of VRAM. By today’s standards, it’s not fast. But combine that with a machine carrying 128GB of system memory, and suddenly you have a very different local AI box. The GPU handles what it’s good at. The system memory gives the model somewhere to go when VRAM runs out. You obviously pay for that with bandwidth and inference speed. But this is the part I find interesting: Local AI is making “how much memory do I have?” almost as important as “how powerful is my GPU?” We’ve spent years chasing faster GPUs. Maybe the next wave of local AI hardware will be about finding smarter ways to use all the memory we already have. More VRAM is still better. But it’s no longer the only way to run a bigger model.

KyzoroX

35,554 次观看 • 11 天前

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

DS4 running on DGX Spark (GB10 / CUDA), private branch for now. 12 tokens/sec, the memory bandwidth is limited in this system, at 270GB/sec. But prefill is ways more alighed to M3 Max at ~200 t/s. I'll release when more mature, but it is almost sure that it will get merged.

antirez

84,127 次观看 • 4 个月前

"Mac Minis for example are a very good fit" - Andrej Karpathy Andrej Karpathy shouted out my work on EXO Labs in his keynote at Y Combinator AI SUS! Here's the breakdown: Right now most AI workloads run in the cloud where requests from different users are continuously batched together. These workloads are FLOPS-bound and favors hardware with the best unit economics of $ per FLOP, i.e. enterprise GPUs. The personal computing revolution will shift these workloads to personal devices with lower batch sizes (mostly batch_size=1). batch_size=1 inference is memory-bound, because all of the model parameters need to be loaded into the GPU every time a token is generated. Apple Silicon with its Unified Memory architecture has a lot of memory and memory bandwidth per $ compared to other hardware: - M4 Pro Mac Mini, 24GB @ 273GB/s, $58.33/GB, $5.13/GB/s - H100, 80GB @ 3350GB/s, $625/GB, $14.93/GB/s The unit economics of Apple Silicon are becoming more compelling with every release of the Mac. The future of AI inference looks more like open weights models (OpenAI open weights model soonTM) run at low batch_size on personal devices.

"Mac Minis for example are a very good fit" - Andrej Karpathy Andrej Karpathy shouted out my work on EXO Labs in his keynote at Y Combinator AI SUS! Here's the breakdown: Right now most AI workloads run in the cloud where requests from different users are continuously batched together. These workloads are FLOPS-bound and favors hardware with the best unit economics of $ per FLOP, i.e. enterprise GPUs. The personal computing revolution will shift these workloads to personal devices with lower batch sizes (mostly batch_size=1). batch_size=1 inference is memory-bound, because all of the model parameters need to be loaded into the GPU every time a token is generated. Apple Silicon with its Unified Memory architecture has a lot of memory and memory bandwidth per $ compared to other hardware: - M4 Pro Mac Mini, 24GB @ 273GB/s, $58.33/GB, $5.13/GB/s - H100, 80GB @ 3350GB/s, $625/GB, $14.93/GB/s The unit economics of Apple Silicon are becoming more compelling with every release of the Mac. The future of AI inference looks more like open weights models (OpenAI open weights model soonTM) run at low batch_size on personal devices.

Alex Cheema

103,604 次观看 • 1 年前

DeepSeek R1 671B is ridiculously good, but running at 158 tokens per second is something out of this world. A model *this good* running *this fast* is 3 years ahead of this time. The secret: DeepSeek in this short video isn't running on GPUs. It's running on custom-made chips! This is running on SambaNova and their SN40L chips, which are designed and optimized for running AI workflows. SambaNova is sponsoring my work. We designed GPUs to run games and later realized we could use them for AI. This is very different from creating a chip specifically for AI from the ground up. Just think about this for a second: One single SNL40 chip can simultaneously hold 100+ models (trillions of parameters) in memory! An agentic workflow that uses multiple models simultaneously won't need to swap models from memory even once! By the way, the Llama family of models *screams* on the SambaNova Cloud. Here are some numbers from my latest tests: • Llama 3.3 70B: 367.32 t/s • Llama 3.2 1B: 2381.82 t/s • Llama 3.2 3B: 1335.71 t/s • Llama 3.1 8B: 929.58 t/s • Llama 3.1 70B: 369.16 t/s • Llama 3.1 405B: 103.37 t/s You can access all of these models for free!

DeepSeek R1 671B is ridiculously good, but running at 158 tokens per second is something out of this world. A model this good running this fast is 3 years ahead of this time. The secret: DeepSeek in this short video isn't running on GPUs. It's running on custom-made chips! This is running on SambaNova and their SN40L chips, which are designed and optimized for running AI workflows. SambaNova is sponsoring my work. We designed GPUs to run games and later realized we could use them for AI. This is very different from creating a chip specifically for AI from the ground up. Just think about this for a second: One single SNL40 chip can simultaneously hold 100+ models (trillions of parameters) in memory! An agentic workflow that uses multiple models simultaneously won't need to swap models from memory even once! By the way, the Llama family of models screams on the SambaNova Cloud. Here are some numbers from my latest tests: • Llama 3.3 70B: 367.32 t/s • Llama 3.2 1B: 2381.82 t/s • Llama 3.2 3B: 1335.71 t/s • Llama 3.1 8B: 929.58 t/s • Llama 3.1 70B: 369.16 t/s • Llama 3.1 405B: 103.37 t/s You can access all of these models for free!

Santiago

131,099 次观看 • 1 年前

$15,000 for a Mac sounds insane. Until you realize what Apple is actually selling. 256GB of unified memory. Not a huge GPU farm. Not an NVIDIA rack. Not 4 different machines. Just one box that can keep multiple large LLMs in memory and run them locally. That changes the equation for local AI. NVIDIA wins when you need raw GPU compute. Apple gets interesting when the bottleneck is memory capacity. Because once your model doesn't fit into 24GB, 32GB or 48GB of VRAM, buying a faster GPU doesn't solve the problem. You need more memory. And that's where machines like the Mac Studio start looking less crazy. The interesting question isn't: “Is a $15K Mac faster than an NVIDIA GPU?” It isn't. The real question is: “How much are you willing to pay to make huge local models actually fit?” That’s where unified memory gets interesting.

$15,000 for a Mac sounds insane. Until you realize what Apple is actually selling. 256GB of unified memory. Not a huge GPU farm. Not an NVIDIA rack. Not 4 different machines. Just one box that can keep multiple large LLMs in memory and run them locally. That changes the equation for local AI. NVIDIA wins when you need raw GPU compute. Apple gets interesting when the bottleneck is memory capacity. Because once your model doesn't fit into 24GB, 32GB or 48GB of VRAM, buying a faster GPU doesn't solve the problem. You need more memory. And that's where machines like the Mac Studio start looking less crazy. The interesting question isn't: “Is a $15K Mac faster than an NVIDIA GPU?” It isn't. The real question is: “How much are you willing to pay to make huge local models actually fit?” That’s where unified memory gets interesting.

KyzoroX

19,536 次观看 • 1 个月前

Etched is deploying two new technologies in chip design: low-voltage inference and cluster-scale memory. CEO Gavin Uberti says they'll make their chips much more power-efficient and way, way faster than today's leading GPUs. He breaks it down: "We looked at a lot of early research directions, and we realized the key things that models need are way more compute and way faster memory." "If you think about inference, there are two key parts: prefill and decode. For prefill, it's a compute-bound problem. You need to have more FLOPS, more operations per second on each of your chips." "On our GPU, the bottleneck's actually thermals. You can't really run a GPU at more than around 50% of what it could theoretically do, or it'll melt." "So we're using a new technology today called low-voltage inference to try to solve this problem. You bring the voltage of the chip down dramatically, which allows us to have way, way better efficiency in terms of how much power is drawn per unit of math, and thus fit way way more flops onto the chip..." "For decode, it's all about bandwidth. Not just bandwidth on a chip, but bandwidth across your cluster. That's why we have this technology we call cluster-scale memory. It reduces the amount of time it takes to communicate from one chip to another dramatically." "As a result we can go use all of our HBM, HBM bandwidth, SRAM, SRAM bandwidth, and our scale-up domain as a single coherent pool. And that means if you're a user, you can go get much faster tokens per second, while still keeping your costs low."

Etched is deploying two new technologies in chip design: low-voltage inference and cluster-scale memory. CEO Gavin Uberti says they'll make their chips much more power-efficient and way, way faster than today's leading GPUs. He breaks it down: "We looked at a lot of early research directions, and we realized the key things that models need are way more compute and way faster memory." "If you think about inference, there are two key parts: prefill and decode. For prefill, it's a compute-bound problem. You need to have more FLOPS, more operations per second on each of your chips." "On our GPU, the bottleneck's actually thermals. You can't really run a GPU at more than around 50% of what it could theoretically do, or it'll melt." "So we're using a new technology today called low-voltage inference to try to solve this problem. You bring the voltage of the chip down dramatically, which allows us to have way, way better efficiency in terms of how much power is drawn per unit of math, and thus fit way way more flops onto the chip..." "For decode, it's all about bandwidth. Not just bandwidth on a chip, but bandwidth across your cluster. That's why we have this technology we call cluster-scale memory. It reduces the amount of time it takes to communicate from one chip to another dramatically." "As a result we can go use all of our HBM, HBM bandwidth, SRAM, SRAM bandwidth, and our scale-up domain as a single coherent pool. And that means if you're a user, you can go get much faster tokens per second, while still keeping your costs low."

TBPN

20,404 次观看 • 2 个月前

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 次观看 • 2 年前

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

UC Berkeley just open-sourced FreeToken. (2–4x faster local LLM inference than Ollama) the results are wild: - Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s - DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s - GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s a 35B model at 16-bit precision needs about 70GB just for its weights. even at 4 bits it is close to 18GB, and FreeToken serves it on an 8GB GPU. let me explain how: all three models mentioned above are Mixture-of-Experts, and that is what FreeToken takes advantage of. each layer holds hundreds of separate experts plus a small router that picks a few of them per token. Qwen3.6-35B activates roughly 3B of its 35B parameters per token. DeepSeek-V4-Flash picks 6 of 256 experts per layer, so 13B of its 284B run at a time. so compute was never the bottleneck. the weights a single step touches fit comfortably on a consumer GPU. every expert the router might pick still has to exist somewhere. they sit in system RAM, and the GPU keeps a cache of the ones the model has been using recently. so everything comes down to what happens when the router picks an expert that is not on the GPU. there are two ways to serve that miss: 1. copy it over PCIe and run it on the GPU 2. run it on the CPU, where it already lives both read from the same system memory, so they compete for one pool of bandwidth instead of adding to each other. existing engines pick one option and freeze it when the model loads. but routing changes on every token, so a fixed choice misses most of what the model asks for. FreeToken measures both bandwidths on your machine and splits each step's misses between the two paths in proportion. the GPU and CPU results then merge exactly, with no approximation. two machines with the same GPU can end up wanting opposite strategies, which I did not expect. a 5090 in a gaming desktop should push nearly everything over PCIe, while an 8GB laptop is better off computing most misses on the CPU. none of that is readable off a spec sheet, so the engine profiles it once per machine. the second half of the design is about agents. coding agents constantly rewrite their own history, and every edit normally forces thousands of tokens back through prefill. FreeToken saves its checkpoints at the exact boundaries agent frameworks cut on, so it only reprocesses the new part. its slowest first token stays under 44 seconds, while llama.cpp peaks at 232 and KTransformers at 946. it serves the OpenAI and Anthropic APIs under Apache 2.0, so Claude Code and Codex can point at it directly. releasing weights publicly decides who can download a model, not who can afford to run one. frontier open models keep shipping, and running them still assumes a rented cluster. meanwhile there are over a hundred million consumer machines with discrete GPUs sitting mostly idle. closing that gap was never a hardware problem, and work like this is what turns open weights into something you can actually use. paper: repo: almost every idea in this post, from why memory bandwidth decides the outcome to why moving weights costs more than computing on them, comes straight out of how a GPU is built. I wrote a detailed primer on that. the article is quoted below.

Akshay 🚀

343,270 次观看 • 29 天前

Sharp question to S Jaishankar in Finland Question: Is India just stuck in the role of a frustrated observer in the Iran war? Is there much you can do? Jaishankar (smiling): I think at this point much of the world is stuck in the role of a frustrated observer

Sharp question to S Jaishankar in Finland Question: Is India just stuck in the role of a frustrated observer in the Iran war? Is there much you can do? Jaishankar (smiling): I think at this point much of the world is stuck in the role of a frustrated observer

Shashank Mattoo

54,140 次观看 • 3 个月前

Free is the part people are underpricing. Google just made Gemini 3.6 Flash the new default. Not a toy model hiding in a free tier. A near-frontier coding model normal users can try in Gemini and AI Studio without touching enterprise pricing. This is landing while people compare paid frontier APIs: > GPT-5.6 Sol at $5/$30 > Claude Sonnet 5 at $3/$15 > Grok 4.5 at $2/$6 > Gemini 3.5 Flash at $1.50/$9 Benchmarks matter. But a model people can actually use for free changes behavior faster than a model they only admire from pricing tables.

Free is the part people are underpricing. Google just made Gemini 3.6 Flash the new default. Not a toy model hiding in a free tier. A near-frontier coding model normal users can try in Gemini and AI Studio without touching enterprise pricing. This is landing while people compare paid frontier APIs: > GPT-5.6 Sol at $5/$30 > Claude Sonnet 5 at $3/$15 > Grok 4.5 at $2/$6 > Gemini 3.5 Flash at $1.50/$9 Benchmarks matter. But a model people can actually use for free changes behavior faster than a model they only admire from pricing tables.

ＺＥＦＩ

13,074 次观看 • 2 个月前

THE AI RACE IS OVER. DUST CLEARED. GROK 4 FAST ALREADY HIT 67 BILLION TOKENS Grok 4 Fast just bulldozed the leaderboard, clocking 67.1 billion tokens and leaving rivals gasping for bandwidth! Leaderboard Mode: • Qwen3’s 80B trails at 23B – miles behind • Gemini 2.5 Flash barely flickers in 3rd • Claude? Still waiting at the loading screen • TrustAI charts show Grok’s surge lighting up November It’s not just an AI model anymore – it’s a speed-run on the entire industry! Source: TrustAI, xAI, Medium

THE AI RACE IS OVER. DUST CLEARED. GROK 4 FAST ALREADY HIT 67 BILLION TOKENS Grok 4 Fast just bulldozed the leaderboard, clocking 67.1 billion tokens and leaving rivals gasping for bandwidth! Leaderboard Mode: • Qwen3’s 80B trails at 23B – miles behind • Gemini 2.5 Flash barely flickers in 3rd • Claude? Still waiting at the loading screen • TrustAI charts show Grok’s surge lighting up November It’s not just an AI model anymore – it’s a speed-run on the entire industry! Source: TrustAI, xAI, Medium

Mario Nawfal

196,729 次观看 • 10 个月前

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

13,282 次观看 • 3 个月前