Грядет Возвращение Центрального Процессора

Ранее в этом году руководители Amazon Web Services поставили перед своими инженерами новые задачи: они должны любой ценой экономить время работы процессора. Сообщается, что AWS столкнулась с резким увеличением времени ожидания загрузки ЦП-серверов, поскольку рабочие нагрузки на ИИ перегружают облачную инфраструктуру компании.

Проблема, по-видимому, застала AWS врасплох, и на то были веские причины. Бум искусственного интеллекта привел к резкому росту спроса на графические процессоры, а затем и на память. Процессоры в основном не упоминались, поскольку из-за их относительного отсутствия распараллеливания они плохо подходили для вывода моделей искусственного интеллекта - процесса запуска и предоставления пользователям больших языковых моделей (LLM).

Но появление агентных систем ИИ, которые позволяют моделям ИИ работать автономно и обращаться к субагентам, меняет картину происходящего.

Мэтт Кимболл, вице-президент и главный аналитик центра обработки данных Moor Insights & Strategy в Остине, штат Техас, говорит, что в 2026 году спрос на процессоры резко возрастет, во многом благодаря агентному ИИ. "Одно дело - иметь такую агентурную нагрузку, и, скажем, это приводит к появлению 100 агентов. Если я собираюсь внедрить это на своем предприятии, эти 100 агентов превратятся в десятки тысяч, сотни тысяч или миллионы", - говорит Кимбалл. "У вас есть агенты, создающие субагентов, выполняющие вызовы API [интерфейса прикладного программирования] и взаимодействующие с другими агентами через контекстный протокол модели [Anthropic]".

Комментарии Кимбалла частично относятся к "использованию инструментов", что означает способность магистра права получать доступ к Интернету, открывать файлы на рабочем столе и в целом использовать различные программы для выполнения своих задач.

Магистры, обученные использованию инструментов, учатся обращаться к другому программному обеспечению. Хотя логические выводы LLM по-прежнему в основном выполняются на графическом процессоре или аналогичном ускорителе искусственного интеллекта, вызовы инструментов, которые выполняет LLM, обычно передаются в центральный процессор.

"Многие компоненты агентной задачи искусственного интеллекта по своей сути являются заданиями на базе центрального процессора", - объясняетSouvik Kundu, senior staff research scientist at Intel. "The CPU does the job of parsing output, figuring out which tool to invoke, making the API call or running the code, collecting the result, and feeding it back." Madhu Rangarajan, vice president of compute and enterprise AI products at AMD, makes a similar claim, saying, "In our testing, seven of the eight stages in realistic agentic AI pipelines run entirely on the CPU."

An LLM tasked with programming software, for example, will likely make tool calls to write code to files, move or replace files, download required packages, and build the software once the LLM believes it’s complete.

Kundu co-authored a paper on agentic AI optimization alongside researchers from Georgia Tech in Atlanta. They found the CPU is often idle while LLM inference is executed on a GPU and that, conversely, the GPU is often idle when tool calls are executed on the CPU. To optimize this, Kundu and his colleagues propose scheduling optimizations that can cut end-to-end latency (the time between the start and finish of the agentic workload) by up to 1.8-times under sustained load.

It’s a start, but the gains chase a moving target. Agentic systems generate work at machine speed and multiply it as they go. OpenAI’s inadvertent hack of Hugging Face saw its model fire off as many as 300 actions an hour, and a single agent can spawn sub-agents that make tool calls of their own.

And there’s one more important complication that may increase the workload on a CPU as models become more complex: safety guardrails.

Safety and policy checks on an agent’s actions are often specific rules that inspect syntax and log files, Kundu says. Guardrails may also use small models (under a billion parameters) to analyze task complexity or intent. Though they could be executed on a GPU, they often aren’t, because their small size and the need to minimize latency keeps the work on the CPU.

Increasing the number of CPUs available significantly decreases the latency for Llama-8B responses over longer sequence lengths.Source: Euijun Chung, Yuxiao Jia, et al.

Euijun Chung, a PhD student at Georgia Tech, recently co-authored another paper, with findings that complement Kundu’s work. Chung and his co-authors found that when a server has too few CPU cores, it falls behind on dispatching work to the GPUs. That causes the GPUs to stall as they wait for instructions.

In addition to that, the paper touches on another key element of LLM workloads: tokenization.

Tokenization is a key first step in LLM inference. It converts text into integer token IDs that can be processed by the model. Unlike the matrix math required for most LLM inference, tokenization is branchy, data-dependent sequential string manipulation. Though it can be parallelized by chunking text, it’s not massively parallel in the same way as the bulk of LLM inference.

Tokenization of small prompts is a relatively trivial task and won’t tax even an entry-level CPU. However, an agentic model that makes tool calls must parse and tokenize the results of the call.

"If you have an ongoing sequence of, say, 100,000 tokens, and you have a tool result of 1,000 tokens, the tokenizer will have to tokenize the whole sequence again. And you have to do tokenization at every agentic tool call," Chung says. This both increases the frequency of tokenization and increases the number of tokens involved. It’s conceivable that future tokenizers will find ways to mitigate this, Chung says, but it remains a problem for modern LLM inference.

The paper finds that time-to-first-token latency (the time required for the model to produce the first word of its reply) can increase dramatically as the sequence length grows. CPUs with more cores can reduce the problem. In test runs at longer sequence lengths, increasing CPU core counts can reduce time-to-first-token latency by roughly 1.5 to 7 times.

Chung and his colleagues were only able to test smaller models, such as Alibaba’s Qwen 3-30B and Meta’s Llama 3.1-70B, due to limitations of the hardware available for testing. He speculates that larger models will experience less dramatic bottlenecks due to their higher overall GPU demand, but also expects agentic AI will push token lengths far beyond the longest he and his co-authors tested.

"If you think about something like Anthropic’s Claude, you can easily hit 500,000, even a million tokens," Chung says. "In the world of agentic AI, the average sequence length will grow and grow, so I’m expecting this problem to get worse in future workloads."

Amazon’s crackdown on use of CPU resources is one of several indicators that Kundu and Chung have identified issues with real-world relevance.

Intel has sold out of server CPUs through at least the end of the year. AMD has doubled its server CPU forecast. Arm and Qualcomm have both announced new CPUs designed to accelerate agentic AI. Even Nvidia has prioritized Vera, its Arm-based CPU for agentic AI, which is part of Nvidia’s Vera Rubin platform.

Kimball says these developments make it clear that the AI industry is placing more emphasis on CPU performance. He sees the surge in demand as an "absolute tell" that CPUs are now considered a key part of an agentic AI system.

Unfortunately, this may translate to broader CPU shortages and increased prices, much as has already occurred with GPUs and memory.

"You’re already seeing a CPU crunch to some degree. When you look at the constraints in the market, it even trickles down into the consumer space," Kimball says. He adds that Intel has cut production of client CPUs in favor of server CPUs, even as Intel’s new 18A production process has grown the company’s sales in the client segment. Kimball sees that as a sign that CPU makers will follow the money.

>