Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

"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...

alexocheema's profile picture

Alex Cheema

52,706 subscribers

103,604 Aufrufe • vor 1 Jahr •via X (Twitter)

Wissenschaft & Technologie
LIVE

Anya Rossi• Live Now

Private livecam show

10 Kommentare

Profilbild von Alex Cheema - e/acc
Alex Cheema - e/accvor 1 Jahr

First post on @karpathy slides:

Profilbild von Alex Cheema - e/acc
Alex Cheema - e/accvor 1 Jahr

Second post on @karpathy slides:

Profilbild von Amal
Amalvor 1 Jahr

@karpathy @exolabs @ycombinator It's funny ai people are calling the thing AI SUS 😆

Profilbild von —
—vor 1 Jahr

@karpathy @exolabs @ycombinator Congratulations :)

Profilbild von UHD K1ng
UHD K1ngvor 1 Jahr

@karpathy @exolabs @ycombinator Rtx A6000 Quadro from 2023 setup is always better than Mac Mini and extra for. rAM that u can run does not give u an advantage and agent should be short 4 billion parameters as LLM.

Profilbild von $MIA
$MIAvor 1 Jahr

@karpathy @exolabs @ycombinator centralized AI? ancient history lol

Profilbild von dcbuilder.eth ⚪️
dcbuilder.eth ⚪️vor 1 Jahr

@karpathy @exolabs @ycombinator You must be over the moon

Profilbild von Blake Fullerton
Blake Fullertonvor 1 Jahr

@karpathy @exolabs @ycombinator Alex is cracked ✊ nice work man

Profilbild von Andrew — e/acc
Andrew — e/accvor 1 Jahr

@karpathy @exolabs @ycombinator congrats!🔥🚀

Profilbild von BrenJ🔺
BrenJ🔺vor 1 Jahr

@karpathy @exolabs @ycombinator How do we best align market forces to accelerate the shift to personal devices? Aka how do we make it +EV for a “normie” to plunk down $1.5k for a decent Mac mini?

Ähnliche Videos

Someone bought a Mac Mini for local AI. 64GB of unified memory seemed like the obvious choice. More memory. Bigger models. No VRAM limits. Then they put it against an RTX 3060. And things got weird. The Mac Mini M4 Pro pushed Qwen 2.5 at 38 tok/s. The old RTX 3060 with just 12GB VRAM hit 52 tok/s. That’s the interesting part about local AI. Apple and NVIDIA are solving different problems. Apple gives you a huge pool of memory. NVIDIA gives you dedicated compute and CUDA. One helps you fit bigger models. The other can run smaller models faster. Memory gets the model in. Compute gets the tokens out.
1:11

Someone bought a Mac Mini for local AI. 64GB of unified memory seemed like the obvious choice. More memory. Bigger models. No VRAM limits. Then they put it against an RTX 3060. And things got weird. The Mac Mini M4 Pro pushed Qwen 2.5 at 38 tok/s. The old RTX 3060 with just 12GB VRAM hit 52 tok/s. That’s the interesting part about local AI. Apple and NVIDIA are solving different problems. Apple gives you a huge pool of memory. NVIDIA gives you dedicated compute and CUDA. One helps you fit bigger models. The other can run smaller models faster. Memory gets the model in. Compute gets the tokens out.

KyzoroX

26,269 Aufrufe • vor 28 Tagen

NVIDIA sent us 2 DGX Sparks. For a while we wondered what we would do with them. The memory bandwidth is 273GB/s making it 3x slower than an M3 Ultra (819GB/s) for batch_size=1 inference. But it has 4x more FLOPS (100 TFLOPS compared to 26 TFLOPS). So we thought, what if we could combine the DGX Spark & M3 Ultra, and make use of both the massive compute on the DGX Spark and the massive memory-bandwidth on the M3 Ultra. We came up with a way to split inference across both devices and achieve a speedup of up to 4x for long prompts compared to the M3 Ultra on its own. Full details in the blog post linked below.
1:13

NVIDIA sent us 2 DGX Sparks. For a while we wondered what we would do with them. The memory bandwidth is 273GB/s making it 3x slower than an M3 Ultra (819GB/s) for batch_size=1 inference. But it has 4x more FLOPS (100 TFLOPS compared to 26 TFLOPS). So we thought, what if we could combine the DGX Spark & M3 Ultra, and make use of both the massive compute on the DGX Spark and the massive memory-bandwidth on the M3 Ultra. We came up with a way to split inference across both devices and achieve a speedup of up to 4x for long prompts compared to the M3 Ultra on its own. Full details in the blog post linked below.

Alex Cheema

281,225 Aufrufe • vor 11 Monaten

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.
3:14

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 Aufrufe • vor 3 Monaten

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
0:29

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 Aufrufe • vor 1 Jahr

Micron is going to $4,000 and once you understand what inference actually is, the number stops sounding crazy (Save this). Dylan Patel just said that by 2030, OpenAI and Anthropic alone will need over 100 gigawatts of compute combined and by 2040, we may not even be measuring AI infrastructure in gigawatts anymore. We may be talking about terawatts. Every single one of those gigawatts needs memory to function. Without it, the compute is worthless. Most people heard that and thought about Nvidia but they should be thinking about Micron. Every AI model generating a response has two phases. The first is prefill, processing your prompt which is compute-heavy and the second is decode generating each word one token at a time and that phase is almost entirely memory-bound, not compute-bound. During decode, the GPU's processing units sit idle more than 95% of the time, waiting for data to arrive from memory. Google confirmed it in a research paper that decode-phase bottlenecks are dominated by memory bandwidth and capacity not raw compute. The GPU is not the bottleneck but the memory feeding the GPU is. This matters because inference is now where all the money lives. Training a model happens once, Inference happens billions of times a day every ChatGPT response, every Claude output, every agentic workflow running in the background and every one of those token streams is a billing event tied directly to memory performance. Adding more GPUs does not fix this because GPUs are already underutilized in inference because they are sitting idle waiting on memory. Adding more memory bandwidth and capacity is what directly reduces token cost, reduces latency, and allows the same cluster to serve dramatically more users simultaneously. Longer context windows compound the problem further, a model running a 1 million token context window requires dramatically more memory per session than a 10,000 token window, and every new model generation pushes context longer. The market treats memory as a downstream beneficiary of Nvidia orders. The correct framework is the opposite, Micron is the upstream constraint on how much value every Nvidia GPU can actually generate at inference scale. Micron guided Q4 to $50 billion in revenue, has HBM4 ramping at twice the pace of the prior generation, and CEO Sanjay Mehrotra has said supply will not catch demand before the end of 2027. At 8x forward earnings on $112 projected FY2027 EPS, Micron is the most undervalued infrastructure company in the entire AI stack. Inference is memory. Memory is Micron and the inference ramp has barely started. Milk Road Pro members are already up massively on this position and we're just getting started. If you want the full breakdown of what we're buying and why, come join us for just a dollar using the link below!
0:39

Micron is going to $4,000 and once you understand what inference actually is, the number stops sounding crazy (Save this). Dylan Patel just said that by 2030, OpenAI and Anthropic alone will need over 100 gigawatts of compute combined and by 2040, we may not even be measuring AI infrastructure in gigawatts anymore. We may be talking about terawatts. Every single one of those gigawatts needs memory to function. Without it, the compute is worthless. Most people heard that and thought about Nvidia but they should be thinking about Micron. Every AI model generating a response has two phases. The first is prefill, processing your prompt which is compute-heavy and the second is decode generating each word one token at a time and that phase is almost entirely memory-bound, not compute-bound. During decode, the GPU's processing units sit idle more than 95% of the time, waiting for data to arrive from memory. Google confirmed it in a research paper that decode-phase bottlenecks are dominated by memory bandwidth and capacity not raw compute. The GPU is not the bottleneck but the memory feeding the GPU is. This matters because inference is now where all the money lives. Training a model happens once, Inference happens billions of times a day every ChatGPT response, every Claude output, every agentic workflow running in the background and every one of those token streams is a billing event tied directly to memory performance. Adding more GPUs does not fix this because GPUs are already underutilized in inference because they are sitting idle waiting on memory. Adding more memory bandwidth and capacity is what directly reduces token cost, reduces latency, and allows the same cluster to serve dramatically more users simultaneously. Longer context windows compound the problem further, a model running a 1 million token context window requires dramatically more memory per session than a 10,000 token window, and every new model generation pushes context longer. The market treats memory as a downstream beneficiary of Nvidia orders. The correct framework is the opposite, Micron is the upstream constraint on how much value every Nvidia GPU can actually generate at inference scale. Micron guided Q4 to $50 billion in revenue, has HBM4 ramping at twice the pace of the prior generation, and CEO Sanjay Mehrotra has said supply will not catch demand before the end of 2027. At 8x forward earnings on $112 projected FY2027 EPS, Micron is the most undervalued infrastructure company in the entire AI stack. Inference is memory. Memory is Micron and the inference ramp has barely started. Milk Road Pro members are already up massively on this position and we're just getting started. If you want the full breakdown of what we're buying and why, come join us for just a dollar using the link below!

Milk Road AI

130,756 Aufrufe • vor 2 Monaten

$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.
0:44

$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 Aufrufe • vor 1 Monat

Apple is no longer hiding it. These machines are built for local AI. Apple unveiled the new Mac Mini with M6 their first 2nm chip. Starting at $899. What you can run on the $899 Mac Mini (32GB) 👇 → Qwen 3.6 27B : fits comfortably, near-GPT-4 quality → Qwen 3 Coder 32B : full coding assistant on your desk → Llama 4 8B at full precision → DeepSeek R1 32B for reasoning tasks → 13.5x faster LLM processing than M1. 4x faster than M4. What you can run on the Mac Studio with M5 Ultra (512GB) 👇 → Llama 4 70B at high quality quantization → DeepSeek V4 with hundreds of billions of parameters → Frontier-scale models entirely in local memory → Connect four Mac Studios over Thunderbolt 5 cluster them to run trillion-parameter models locally Apple said their Mac business grew 30% last quarter largely because people are buying these to run AI models on their own desks instead of paying for cloud tokens. The future of inference is local. Apple built the hardware for it.
0:38

Apple is no longer hiding it. These machines are built for local AI. Apple unveiled the new Mac Mini with M6 their first 2nm chip. Starting at $899. What you can run on the $899 Mac Mini (32GB) 👇 → Qwen 3.6 27B : fits comfortably, near-GPT-4 quality → Qwen 3 Coder 32B : full coding assistant on your desk → Llama 4 8B at full precision → DeepSeek R1 32B for reasoning tasks → 13.5x faster LLM processing than M1. 4x faster than M4. What you can run on the Mac Studio with M5 Ultra (512GB) 👇 → Llama 4 70B at high quality quantization → DeepSeek V4 with hundreds of billions of parameters → Frontier-scale models entirely in local memory → Connect four Mac Studios over Thunderbolt 5 cluster them to run trillion-parameter models locally Apple said their Mac business grew 30% last quarter largely because people are buying these to run AI models on their own desks instead of paying for cloud tokens. The future of inference is local. Apple built the hardware for it.

Vaibhav Sisinty

512,470 Aufrufe • vor 19 Tagen

What if we could connect all the dark compute across the globe to build the world's biggest AI data center? Most of the compute in the world is dark: phones, laptops, Tesla's, PS5's, TV's. These devices have powerful GPUs but are mostly sitting idle. Today, EXO Labs is announcing the first step in activating all of this dark compute for AI workloads: evML, a distributed computing protocol that enforces honest behavior through hardware security and spot-checks with just 5% performance overhead. Version 0.0.1 of evML is compatible with Apple, Samsung and LG devices. Our preliminary report on evML is now available (link below).
0:24

What if we could connect all the dark compute across the globe to build the world's biggest AI data center? Most of the compute in the world is dark: phones, laptops, Tesla's, PS5's, TV's. These devices have powerful GPUs but are mostly sitting idle. Today, EXO Labs is announcing the first step in activating all of this dark compute for AI workloads: evML, a distributed computing protocol that enforces honest behavior through hardware security and spot-checks with just 5% performance overhead. Version 0.0.1 of evML is compatible with Apple, Samsung and LG devices. Our preliminary report on evML is now available (link below).

EXO Labs

137,969 Aufrufe • vor 1 Jahr

$AMZN CEO Andy Jassy describes AI adoption as very barbelled, with most existing enterprise workloads still not using inference though he expects that to change significantly over time. "I think it's actually kind of useful to look at at least our view of what we see in the demand and adoption curve right now, which is we see this adoption curve in AI right now is very barbelled." "There is on one end of the barbell, the AI labs are consuming gobs and gobs of compute. And there are a few runaway successful generative AI applications like Claude Code and ChatGPT." "And on the other end of the barbell are enterprises who are getting real value from AI in cost avoidance and productivity. And these are things like automating customer service or business process automation or fraud." "In the middle of the barbell is all of the current enterprise production workloads, some of which are using inference in a pervasive way, but most of which aren't. And that is going to change very significantly over time." "And that will be, in my opinion, that will be the largest absolute segment, the existing production workloads in the enterprise and new businesses and workloads that startups build too."
1:10

$AMZN CEO Andy Jassy describes AI adoption as very barbelled, with most existing enterprise workloads still not using inference though he expects that to change significantly over time. "I think it's actually kind of useful to look at at least our view of what we see in the demand and adoption curve right now, which is we see this adoption curve in AI right now is very barbelled." "There is on one end of the barbell, the AI labs are consuming gobs and gobs of compute. And there are a few runaway successful generative AI applications like Claude Code and ChatGPT." "And on the other end of the barbell are enterprises who are getting real value from AI in cost avoidance and productivity. And these are things like automating customer service or business process automation or fraud." "In the middle of the barbell is all of the current enterprise production workloads, some of which are using inference in a pervasive way, but most of which aren't. And that is going to change very significantly over time." "And that will be, in my opinion, that will be the largest absolute segment, the existing production workloads in the enterprise and new businesses and workloads that startups build too."

Fireside Alpha

29,242 Aufrufe • vor 1 Monat

Cloud GPU training is a scam. A single M4 MacBook does 2.9 TFLOPS. Seven friends with MacBooks match an NVIDIA A100. Alexander Hayes just open-sourced a tool that makes this work over Wi-Fi. It's called AirTrain. Here's how it works: Traditional distributed training (DDP) syncs gradients after every single step. For a 124M parameter model, that's ~500MB exchanged per step. You need 50 GB/s of sustained bandwidth. Impossible over Wi-Fi. AirTrain uses the DiLoCo algorithm. Each Mac trains independently for 500 steps, then syncs only the difference. One sync per 500 steps instead of one per step. 500x less network communication. Wi-Fi actually works. The entire sync takes ~2 seconds. Here's what makes it wild: → Zero-config discovery. Devices find each other automatically via mDNS/Bonjour. Same protocol as AirDrop. → Fault tolerant. Nodes can join and leave mid-training without killing the run. → Checkpoint relay. Train for a few hours, export a checkpoint, hand it off to someone else to continue. Like a relay race for ML training. → Built on Apple's MLX framework. Native to M1/M2/M3/M4/M5 unified memory. No host-to-device copy overhead. → Local dashboard. Real-time loss curves, peer monitoring, throughput metrics in your browser. Here's the wildest part: An M4 Max with 128GB unified memory can train a 70B parameter model without offloading. An NVIDIA RTX 4090 has 24GB VRAM. Apple Silicon gets ~245-460 GFLOPS per watt. Training on MacBooks costs almost nothing in electricity compared to cloud GPUs. And there are hundreds of millions of Apple Silicon Macs in the world. The math: Traditional DDP: 1 sync per step = 50 GB/s required AirTrain (DiLoCo): 1 sync per 500 steps = 0.1 GB/s required Wi-Fi handles 0.1 GB/s. That's it. That's the breakthrough. They even built a community platform at with live session browsing, checkpoint sharing, and a contributor leaderboard. Training a 124M parameter GPT-2? Instead of renting cloud GPUs at $3/hr, pool three MacBooks in a coffee shop and train for free. MIT licensed. Built in Python. 1 contributor. Early stage but the idea is insane. 100% Open Source. (Link in the comments)
0:44

Cloud GPU training is a scam. A single M4 MacBook does 2.9 TFLOPS. Seven friends with MacBooks match an NVIDIA A100. Alexander Hayes just open-sourced a tool that makes this work over Wi-Fi. It's called AirTrain. Here's how it works: Traditional distributed training (DDP) syncs gradients after every single step. For a 124M parameter model, that's ~500MB exchanged per step. You need 50 GB/s of sustained bandwidth. Impossible over Wi-Fi. AirTrain uses the DiLoCo algorithm. Each Mac trains independently for 500 steps, then syncs only the difference. One sync per 500 steps instead of one per step. 500x less network communication. Wi-Fi actually works. The entire sync takes ~2 seconds. Here's what makes it wild: → Zero-config discovery. Devices find each other automatically via mDNS/Bonjour. Same protocol as AirDrop. → Fault tolerant. Nodes can join and leave mid-training without killing the run. → Checkpoint relay. Train for a few hours, export a checkpoint, hand it off to someone else to continue. Like a relay race for ML training. → Built on Apple's MLX framework. Native to M1/M2/M3/M4/M5 unified memory. No host-to-device copy overhead. → Local dashboard. Real-time loss curves, peer monitoring, throughput metrics in your browser. Here's the wildest part: An M4 Max with 128GB unified memory can train a 70B parameter model without offloading. An NVIDIA RTX 4090 has 24GB VRAM. Apple Silicon gets ~245-460 GFLOPS per watt. Training on MacBooks costs almost nothing in electricity compared to cloud GPUs. And there are hundreds of millions of Apple Silicon Macs in the world. The math: Traditional DDP: 1 sync per step = 50 GB/s required AirTrain (DiLoCo): 1 sync per 500 steps = 0.1 GB/s required Wi-Fi handles 0.1 GB/s. That's it. That's the breakthrough. They even built a community platform at with live session browsing, checkpoint sharing, and a contributor leaderboard. Training a 124M parameter GPT-2? Instead of renting cloud GPUs at $3/hr, pool three MacBooks in a coffee shop and train for free. MIT licensed. Built in Python. 1 contributor. Early stage but the idea is insane. 100% Open Source. (Link in the comments)

Guri Singh

160,201 Aufrufe • vor 4 Monaten

BREAKING $AMD is ready to ship Ryzen AI Halo 🆕 in June 2026!!!! 🚀🚀🚀 AKA Personal/Local AI system. Ryzen AI Max+ 395 (Strix Halo APU) with 128 GB of unified LPDDR5X memory, 40 RDNA 3.5 GPU compute units, and a full ROCm software stack preloaded for both Windows and Linux. AMD Halo is $2,000-$3,000 vs $NVDA Spark $4,699 The main purpose of the AMD Ryzen AI Halo is to serve as a compact, developer focused workstation for running intensive AI workloads locally, specifically building, testing, prototyping, fine-tuning, and deploying generative AI models (like LLMs) and Agentic AI applications without relying on the cloud. Run models with up to 200 billion parameters on your desk using 128 GB of unified memory (shared across CPU, GPU, and NPU). This eliminates cloud costs, latency, data privacy concerns, and internet dependency. ~This is for AI developers, researchers, engineers, and creators. ~Teams or individuals wanting faster iteration, data privacy, and cost savings compared to cloud APIs. ~Users experimenting with open-source models, fine-tuning, inference, and production-like local deployment Source: AMD CES 2026
2:47

BREAKING $AMD is ready to ship Ryzen AI Halo 🆕 in June 2026!!!! 🚀🚀🚀 AKA Personal/Local AI system. Ryzen AI Max+ 395 (Strix Halo APU) with 128 GB of unified LPDDR5X memory, 40 RDNA 3.5 GPU compute units, and a full ROCm software stack preloaded for both Windows and Linux. AMD Halo is $2,000-$3,000 vs $NVDA Spark $4,699 The main purpose of the AMD Ryzen AI Halo is to serve as a compact, developer focused workstation for running intensive AI workloads locally, specifically building, testing, prototyping, fine-tuning, and deploying generative AI models (like LLMs) and Agentic AI applications without relying on the cloud. Run models with up to 200 billion parameters on your desk using 128 GB of unified memory (shared across CPU, GPU, and NPU). This eliminates cloud costs, latency, data privacy concerns, and internet dependency. ~This is for AI developers, researchers, engineers, and creators. ~Teams or individuals wanting faster iteration, data privacy, and cost savings compared to cloud APIs. ~Users experimenting with open-source models, fine-tuning, inference, and production-like local deployment Source: AMD CES 2026

Mike

16,097 Aufrufe • vor 3 Monaten

Andrew Feldman, co-founder and CEO of Cerebras gives the best explanation of why Cerebras' wafer-scale architecture is 2,500X faster than a GPU during LLM inference. During inference, there are 2 stages: - pre-fill, where the model first processes the user's prompt, and - decode, where it generates the answer 1 token at a time in sequence. During that sequencial Decode phase, before each token is calculated the model weights have to be moved from memory into compute. On a GPU those weights are moved from HBM, while Cerebras keeps them in much faster SRAM spread across its very large wafer-scale processor, so the memory-to-compute movement that must happen for every token is about 2,500× faster ---- From The MAD Podcast with Matt Turck and Cerebras YouTube channel, (full video link in comment)
2:48

Andrew Feldman, co-founder and CEO of Cerebras gives the best explanation of why Cerebras' wafer-scale architecture is 2,500X faster than a GPU during LLM inference. During inference, there are 2 stages: - pre-fill, where the model first processes the user's prompt, and - decode, where it generates the answer 1 token at a time in sequence. During that sequencial Decode phase, before each token is calculated the model weights have to be moved from memory into compute. On a GPU those weights are moved from HBM, while Cerebras keeps them in much faster SRAM spread across its very large wafer-scale processor, so the memory-to-compute movement that must happen for every token is about 2,500× faster ---- From The MAD Podcast with Matt Turck and Cerebras YouTube channel, (full video link in comment)

Rohan Paul

39,999 Aufrufe • vor 16 Tagen

Wake the world's sleeping compute. Look at the Mac nearest to you. What's it doing? Probably nothing. There are 100M+ Macs with Apple Silicon out there. Apple quietly made them *really* good at inference. A $3k Mac runs a 60B model at 30 watts. Most sit idle most of the day. Meanwhile every AI API call passes through three layers of margin before reaching the hardware. We call this the Inference Tax. We got curious: what happens if you connect idle Macs directly to inference demand? This is Darkbloom. Private inference network for idle Macs. darkbloom [dot] dev -- paper + code open. Reply for invite + free credits ↓
2:58

Wake the world's sleeping compute. Look at the Mac nearest to you. What's it doing? Probably nothing. There are 100M+ Macs with Apple Silicon out there. Apple quietly made them *really* good at inference. A $3k Mac runs a 60B model at 30 watts. Most sit idle most of the day. Meanwhile every AI API call passes through three layers of margin before reaching the hardware. We call this the Inference Tax. We got curious: what happens if you connect idle Macs directly to inference demand? This is Darkbloom. Private inference network for idle Macs. darkbloom [dot] dev -- paper + code open. Reply for invite + free credits ↓

Gajesh

511,813 Aufrufe • vor 5 Monaten

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 💰💰💰
0:32

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 Aufrufe • vor 2 Jahren

Everyone wrote Apple off as the AI loser, but one hardware spec might flip that story upside down (Save this). @jason called Apple a screaming buy on the back of a single chip detail. The rumored M7 Ultra, expected around 2028, is designed to support up to 1.5TB of unified memory, enough to run frontier class trillion parameter AI models locally, with no cloud required. The Street's bear case on Apple is straightforward. Apple has no frontier model of its own, Siri has stumbled for years and the company effectively rents OpenAI's models for its hardest queries. That narrative treats Apple as the one Magnificent Seven name that missed the AI wave entirely but the bull case flips that framing on its head. If frontier AI models keep shrinking and getting cheaper to run, Apple doesn't need the smartest model in the world, it just needs to own the device that model runs on. And unified memory is the mechanism that makes this possible. Unlike traditional systems where the CPU and GPU each need separate memory, Apple's architecture lets the CPU, GPU and Neural Engine draw from one shared pool. A fully specced M7 Ultra could theoretically run something on the scale of a 1.2 trillion parameter model locally and that capability plugs directly into the one advantage Apple has spent over a decade building: privacy. Apple has already shipped Private Cloud Compute, a system designed so even Apple can't access user data processed off device. Apple doubled down on this at WWDC 2026, framing on device privacy as non-negotiable while rivals default to the cloud. If the best AI models get small enough to run on Apple silicon, the moat stops being the model and becomes the hardware it has to sit on. Milk Road Pro remains bullish on Apple and it remains as one of our core positions, if you want the full thesis + our full AI trades, come join us using the link below for just a $1.
2:41

Everyone wrote Apple off as the AI loser, but one hardware spec might flip that story upside down (Save this). @jason called Apple a screaming buy on the back of a single chip detail. The rumored M7 Ultra, expected around 2028, is designed to support up to 1.5TB of unified memory, enough to run frontier class trillion parameter AI models locally, with no cloud required. The Street's bear case on Apple is straightforward. Apple has no frontier model of its own, Siri has stumbled for years and the company effectively rents OpenAI's models for its hardest queries. That narrative treats Apple as the one Magnificent Seven name that missed the AI wave entirely but the bull case flips that framing on its head. If frontier AI models keep shrinking and getting cheaper to run, Apple doesn't need the smartest model in the world, it just needs to own the device that model runs on. And unified memory is the mechanism that makes this possible. Unlike traditional systems where the CPU and GPU each need separate memory, Apple's architecture lets the CPU, GPU and Neural Engine draw from one shared pool. A fully specced M7 Ultra could theoretically run something on the scale of a 1.2 trillion parameter model locally and that capability plugs directly into the one advantage Apple has spent over a decade building: privacy. Apple has already shipped Private Cloud Compute, a system designed so even Apple can't access user data processed off device. Apple doubled down on this at WWDC 2026, framing on device privacy as non-negotiable while rivals default to the cloud. If the best AI models get small enough to run on Apple silicon, the moat stops being the model and becomes the hardware it has to sit on. Milk Road Pro remains bullish on Apple and it remains as one of our core positions, if you want the full thesis + our full AI trades, come join us using the link below for just a $1.

Milk Road AI

37,459 Aufrufe • vor 1 Monat

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.
0:27

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 Aufrufe • vor 4 Tagen

Google Ironwood TPU Memory Hierarchy in 9 levels by hand ✍️ 1. Bit – The most basic unit of information, the on–off decision from which every number, tensor, and model state is ultimately constructed. 2. FP8 (1×8 → 8 bits) – Eight bits are grouped to form a floating-point value, typically used for inference, where reduced precision is a deliberate trade-off to maximize throughput and efficiency. 3. BF16 (×2 → 16 bits) – Two FP8-scale chunks are combined to gain more dynamic range and stability, while still staying friendly to high-throughput hardware. 4. Tensor tile (×1024 → 1K) – Data moves through the chip in blocks of 1024 values at a time, defining the granularity at which tensors are fetched and manipulated. 5. Matrix Multiplication Unit (MXU) (×64 → 64K) – A systolic array where matrix multiplication is not abstract but physical, with tensor tiles flowing through fixed hardware to achieve the highest possible throughput. 6. Vector Memory (VMEM) (×2048 → 128M) – On-chip working memory that holds activations, partial results, and intermediates, sized specifically to keep the systolic array busy without stalling. 7. Common Memory (CMEM) (×8 → 1 GB) – A small but critical shared memory sitting between VMEM and HBM, used for staging, accumulation, synchronization, and cross-lane coordination. 8. HBM (×96 → 96 GB) – Off-chip high-bandwidth memory where model weights and large states live, implemented as HBM3e with 16 stacks at 6 GB each, for a total of 96 GB. 9. Dual-Die (x2 → 192GB) – Two tightly coupled compute dies operate as a single logical accelerator, each with its own local HBM, effectively doubling memory capacity and bandwidth while allowing tensors and activations to stream seamlessly across dies as if they lived on one chip. I created this drawing for this week's seminar. I’ll take you through these 9 levels in a beginner-friendly way by hand ✍️. RSVP 👉
0:27

Google Ironwood TPU Memory Hierarchy in 9 levels by hand ✍️ 1. Bit – The most basic unit of information, the on–off decision from which every number, tensor, and model state is ultimately constructed. 2. FP8 (1×8 → 8 bits) – Eight bits are grouped to form a floating-point value, typically used for inference, where reduced precision is a deliberate trade-off to maximize throughput and efficiency. 3. BF16 (×2 → 16 bits) – Two FP8-scale chunks are combined to gain more dynamic range and stability, while still staying friendly to high-throughput hardware. 4. Tensor tile (×1024 → 1K) – Data moves through the chip in blocks of 1024 values at a time, defining the granularity at which tensors are fetched and manipulated. 5. Matrix Multiplication Unit (MXU) (×64 → 64K) – A systolic array where matrix multiplication is not abstract but physical, with tensor tiles flowing through fixed hardware to achieve the highest possible throughput. 6. Vector Memory (VMEM) (×2048 → 128M) – On-chip working memory that holds activations, partial results, and intermediates, sized specifically to keep the systolic array busy without stalling. 7. Common Memory (CMEM) (×8 → 1 GB) – A small but critical shared memory sitting between VMEM and HBM, used for staging, accumulation, synchronization, and cross-lane coordination. 8. HBM (×96 → 96 GB) – Off-chip high-bandwidth memory where model weights and large states live, implemented as HBM3e with 16 stacks at 6 GB each, for a total of 96 GB. 9. Dual-Die (x2 → 192GB) – Two tightly coupled compute dies operate as a single logical accelerator, each with its own local HBM, effectively doubling memory capacity and bandwidth while allowing tensors and activations to stream seamlessly across dies as if they lived on one chip. I created this drawing for this week's seminar. I’ll take you through these 9 levels in a beginner-friendly way by hand ✍️. RSVP 👉

Tom Yeh

30,489 Aufrufe • vor 7 Monaten

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.
1:08

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 Aufrufe • vor 3 Monaten

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.
0:42

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 🚀

341,701 Aufrufe • vor 22 Tagen