个人助理要跑在多少核的 CPU 上

Mike @MikeLongTerm原推发布于 2026.09.30

在 X 上打开原推

$AMD $META @OpenAI $AAPL $AMZN $GOOGL @AnthropicAI | Why Consumers/Enterprises Agentic must have the highest Core/Thread 🧵 Not Financial Advice! DYOR! I do own AMD shares! Think of every personal agent (Muse, Dots, and the ones Amazon, Google, Microsoft, and Apple to compete on consumers agentic ) as a remote employee who gets their own laptop that never turns off. The GPU is the employee’s brain, it does the thinking. The CPU is the laptop itself: the OS, the browser, the files, Slack, email, shopping tabs, code they run, and the ability to keep working after you close ChatGPT. That laptop has to stay isolated so one agent cannot peek at another user’s mail. That isolation is a small virtual machine, typically about 2 vCPUs (EPYC Turin) and 8 GB of RAM. I will tag various threads so u understand more, one good example is @AMD CTO @mpaper2000 sat with Mike Clark, AMD SVP, Corporate Fellow and Chief Architect on why AMD EYPC is the ultimate option to scale large numbers of agents and training agents with RL. Thread density is the strongest AMD selling point for scaling agents and RL rollouts. 1. Why highest core and thread count is a must for scaling the number of agents If you want 100 million people or billions to each have one of those always-on laptops, you need hundreds of millions of tiny computers running at the same time. You do not buy 100 million separate chips. You slice one giant chip into many small apartments. ~A current best dense AMD EPYC can put 128–192 cores (256–384 threads) in one socket (2 yr old) ~The next one, Venice ramp from Q4 26, goes to 256 cores and 512 threads on a single chip. Simple packing math: one 256 core chip can host on the order of 100+ agent VMs if each guest gets 2 vCPUs and you allow normal cloud oversubscription. A 32 core chip hosts a handful. To serve the same city of agents you would need eight times as many servers, eight times the racks, power, and cooling. That is why Dr. Lisa Su has been saying the old “1 CPU for every 4–8 GPUs” ratio is flipping to 2 CPU 1 GPU or even much higher ratio in more agentic workloads. Persistent agents eat CPUs the way chatbots ate GPUs. Highest thread density is how you fit more agents per watt, per dollar, and per rack. x86 still wins a lot of this work because the agent has to drive the real web and existing software, not a custom island rewritten for ARM. And yes $INTC will certainly benefit as we have a severe shortage that could last 3-5 years. 2. Why the same high-core chips matter for training agents with reinforcement learning Serving agents is “give each person a laptop.” Training agents with RL is “run a massive internship program.”RL learns by trying. The model takes an action, sees what happened, gets a reward, and updates. One try is slow. Ten thousand tries in parallel is how you get a capable agent in a reasonable time. ~Classic example: OpenAI once needed almost 30,000 CPU cores just to train a robot on a Rubik’s Cube because each parallel environment lived on a CPU. And yes, that was why @OpenAI started a long term strategic partnership with @AMD in the early day for GPU and CPU. ~For computer use agents (browse, click, write code, use apps), the “environment” is still a real OS + browser + tools. That work does not all move to the GPU. You still need many CPU threads running many sandboxes at once so the learner is not sitting idle waiting for the next batch of experience. More threads = more intern desks running at the same time = the model sees more outcomes per hour = it gets good faster. GPUs still do the heavy model updates. High-core CPUs keep the experience factory fed. One picture that ties it together Imagine you are opening a 24-hour virtual call center. ~Each agent is one employee who needs their own desk, computer, and browser that stays logged in even on lunch break. ~A high core EPYC is a skyscraper floor plan: hundreds of desks in one building. ~A low core chip is a small office. You would need a whole industrial park to employ the same workforce. ~RL training is new hire boot camp: you do not train one intern at a time. You run thousands of practice shifts in parallel so the company playbook improves overnight. 3. $META and @OpenAI will have proritized CPU allocations OpenAI signed a 6GW AMD deal in October 2025. Meta signed a matching 6GW deal in February 2026. Those headlines were about Instinct GPUs (Helios Rack), but the racks are not GPU only. Helios pairs the accelerators with 6th gen EPYC Venice CPUs. Meta is also a named lead customer for Venice, after already running millions of EPYC chips in its fleet. That means the first Venice wafers that can pack the most agent sandboxes per socket are spoken for by the two companies now launching consumer agents at scale. OpenAI and Meta can each potentially own on the order of 10% of AMD after deployments. That is why people say they have “prioritized CPU allocation as shareholders.” They get early systems and co-design access because they are strategic partners. Connect that back to the apartment building analogy. Muse and Dots each need a private laptop that stays on. High core Venice is the skyscraper that can hold the most of those laptops per floor. OpenAI and Meta are the two landlords who already reserved the first towers, and they hold cheap options on the construction company. When more agents show up, they do not have to scramble for leftover cores. They already have first claim on the densest x86 inventory, plus a reason to keep buying it. In 2022 the market was already pivoting hard to GPUs. Dr. Su still treated the data center CPU as AMD’s largest growth engine and kept pouring cores into EPYC instead of treating the server processor as a leftover host chip. By June 2023 she was saying most AI workloads still ran on CPUs and calling Genoa “the best CPU for AI.” That was not the agentic slogan yet. It was the early version of the same idea: AI is not only matrix math on a GPU. A lot of the job is general purpose compute, orchestration, and packing many small tenants onto one socket. She kept the high core roadmap through that whole GPU mania: Genoa, Bergamo, Turin (still beat most $ARM based CPUs today, then Venice at 256 cores and 512 threads. The product plan was already a bet that someone would need extreme thread density. Agentic AI later gave that bet a name. By Advancing AI 2024 she was talking about agentic workloads needing “significantly more general purpose compute as well as AI compute.” By 2026 explosive Agentic AI demand send CPU GPU ratio to 2 to 1, and many new estimates are pointing to 4-8 CPU 1 GPU or even extreme case 40 CPU 1 GPU; and that enough agents could mean more CPUs than GPUs. Dr. Su also said CPU demand had “far exceeded” her own already bullish forecast. Not Financial Advice! DYOR! I do own AMD shares!

这条是引用着另一条写的：「Muse 每个用户一台 Linux 虚拟机，CPU 需求暴涨」