Intel and Tencent Cloud Release New-Generation Cloud Instances Based on Xeon 6+, with Over 30% Overall Performance Improvement

Author | Chen Yuelin

Editor | Bao Yonggang

That a single AI agent can complete a task smoothly does not mean that hundreds or thousands of agents can run efficiently at the same time.

For cloud servers that carry a large number of agent tasks, beyond increasing computing speed, how to quickly prepare execution environments when tasks arrive in bursts and how to reduce the memory resources occupied by temporarily idle tasks have also become urgent problems to solve.

In response, at the annual Tencent Global Digital Ecosystem Summit, Intel and Tencent Cloud jointly released a new generation of cloud instances based on Xeon 6+ processors. After optimization around key factors such as task startup efficiency and memory resource utilization, their overall performance is more than 30% higher than the previous generation.

From the earlier S9, S9e, and S9pro powered by Xeon 6 processors to today's new-generation cloud servers (CVM), the collaboration between Intel and Tencent Cloud has not stopped at upgrading processor performance. Facing trends such as continuously rising model invocation volumes and agents beginning to take on more work in actual business operations, how the CPU coordinates with the operating system, memory, and other components is becoming a new challenge for cloud servers—and this is also the focus of the two companies' joint optimization this time.

288-core Xeon 6+ processors: supporting Agent deployment still requires overcoming the memory barrier

AI applications are climbing to an entirely new scale. The "Large Model Token Service Performance Monitoring Report (2026)" released by the China Academy of Information and Communications Technology shows that in June 2026, domestic average daily Token consumption had reached about 500 trillion, nearly four times that of December last year.

This "carnival" around AI usage is not happening only in China; it is also sweeping the world in the application form of Agents. In a McKinsey survey covering 97 countries worldwide, among respondents from companies with annual revenue exceeding $1 billion, 40% said their companies had already deployed Agents at scale in at least one business function, up from 27% in 2025.

Unlike traditional one-question-one-answer interactions, an agent can break a task into multiple steps, repeatedly invoke models, use tools, execute code, and adjust actions based on feedback. This may not only lead to more Token consumption, but also require the CPU to take on more work beyond model inference.

According to Guo Wei, Vice President of Intel's Marketing Group and General Manager of Intel China, in the first half of this year the industry was still discussing the CPU-to-GPU configuration ratio in terms such as 1:8 and 1:4; entering the second half, expectations that CPU demand will continue to grow have become increasingly evident, and some even suggested that it may evolve toward 1:1 in the future.

In its server CPU product layout, Intel made clear as early as 2023 its dual-track direction of Performance-cores (P-cores) and Efficient-cores (E-cores) to meet the computing performance, core density, and energy efficiency requirements of different workloads. Among them, P-cores focus on computing performance and are suitable for workloads with high single-core performance requirements; E-cores emphasize core density and energy efficiency and are better suited to large-scale parallel, high-throughput computing tasks.

In June this year, Intel launched the Xeon 6+ processor (codenamed Clearwater Forest) manufactured using the Intel 18A process. The processor integrates up to 288 E-cores, supports 12 memory channels, and DDR5 delivers a maximum transfer rate of 8000MT/s. Intel also disclosed that a single Xeon 6+ processor can support the stable operation of more than 1,000 agents.

However, an increase in the number of CPU cores and higher memory transfer speeds do not mean a server can carry more agents proportionally. When analyzing agent workloads, Intel technical experts found that CPU usage is markedly bursty—when an agent starts an execution environment and executes tasks, it generates short-term peaks in computing demand, but over the course of a single Agent's operation, overall CPU utilization is generally not high.

In contrast, memory is different. Intel experts also pointed out that even when an agent has no computing task for the time being, its running state may still continuously occupy memory. They did the math on site: assuming each Agent is allocated 2GB of memory, running 4 Agents on one vCPU resource would require 8GB of memory; while in an example of a 256-core CPU with 12 memory channels and 128GB of memory per channel, the average physical memory capacity corresponding to each vCPU is only 6GB.

This means that when cloud providers seek to increase agent deployment density through resource reuse, memory capacity may become a limiting factor earlier than CPU computing power. Therefore, Intel and Tencent Cloud have further extended this collaboration based on Xeon 6+ to areas including the operating system kernel, resource scheduling, and memory management, attempting to translate hardware capabilities into performance improvements in actual business operations.

Memory must be saved, and sandboxes must also start quickly

Data from TrendForce shows that in the second quarter of 2026, the average contract price of server DRAM rose 53%–58% quarter on quarter. Although the pace of increase has slowed somewhat, the institution still expects prices to continue rising 13%–18% in the third quarter.

Facing continuously increasing procurement costs, for cloud providers hoping to run more Agent tasks on a single server, finding ways to make existing memory hold more data is urgent.

Compressing and storing memory data that does not need to be accessed frequently for the time being, and then decompressing it when needed, is one of the key paths to improving server memory utilization. But this brings a new problem: while saving space, compression and decompression also consume computing and memory access resources. Especially when a large number of tasks request decompression at the same time, the extra latency may slow down business response. The Tencent Cloud TencentOS Server product R&D team once tried using the LZ4 software solution for memory compression, but found that the maximum latency of multi-channel concurrent decompression was as high as 5 milliseconds, which could not meet actual business requirements.

To solve this problem, Tencent Cloud turned to the IAA (In-Memory Analytics Accelerator) built into Intel Xeon CPUs, using dedicated hardware to perform compression and decompression tasks, reducing the occupation of CPU computing resources and improving processing latency. On this basis, the Tencent Cloud team also adapted the IAA driver, simplified unnecessary call paths, and tested a variety of batch parallel configurations, ultimately choosing the IAA-batch-8 solution.

Test results show that the advantage of IAA lies not only in shortening decompression time, but also in stability under high concurrency. When the concurrency scale increased from 32 channels to 256 channels, the decompression latency of the LZ4 software solution rose accordingly, while the latency of IAA-batch-8 remained basically stable. In a comprehensive stress test closer to actual business, for the decompression of 2MB memory huge pages, compared with the software solution, the average latency and maximum latency of IAA improved by about 16 times and 9 times respectively, and both were at the microsecond level.

In addition to memory resources, sandbox creation efficiency is also an important factor affecting the high-density deployment and high-concurrency execution of agents.

The so-called sandbox is an isolated environment provided for agents to execute code and call tools, avoiding interference between different tasks. The Tencent Cloud CubeSandbox team pointed out that in some large-model reinforcement learning scenarios, one round of training may require starting thousands or even hundreds of thousands of sandboxes for model evaluation and inference. How to complete the creation of a large number of execution environments in a short time while controlling resource usage has become a problem that cloud providers need to solve.

However, different isolation technologies each have limitations: traditional containers start quickly, but usually share the server's operating system kernel; full virtual machines have independent kernels, but require more resources. Therefore, Tencent Cloud CubeSandbox adopts technologies such as micro virtual machines (MicroVM), snapshots, and resource reuse, enabling multiple sandboxes to share the same basic resources and independently store only the changed parts, thereby reducing duplicate memory usage.

On this basis, Intel engineers also optimized some operating system kernel execution paths involved in the sandbox creation process, and introduced the relevant improvements into the kernel of OpenCloudOS, an open-source operating system jointly initiated by Tencent and other enterprises, further reducing the latency when a large number of sandboxes start concurrently.

Test data disclosed on site shows that, under the condition of meeting the 200-millisecond service level objective (SLA), a server equipped with Xeon 6+ processors can support 350 concurrent sandbox creation requests per machine when running CubeSandbox, an improvement of about 46% over other mainstream test platforms.

More noteworthy is the performance after workloads are superimposed. Tencent Cloud's actual tests found that even when the server simultaneously carries other tasks and the overall load rises, Xeon 6+ still maintains relatively stable concurrent sandbox creation capability.

However, Intel does not believe that existing products can already fully meet the operating requirements of agents. Intel technical experts said on site that there is still room for further optimization between agent applications and the software and hardware ecosystem, and the team basically reports the experience accumulated from actual workload optimization with partners to the back-end CPU design team every week, providing reference for the improvement of subsequent products.

The cooperation between the two sides will also continue to the next generation of products. According to Xia Shiyue, senior product manager at Tencent Cloud, Tencent Cloud plans to continue cooperating with Intel to develop cloud server instances based on the next-generation Xeon processor Diamond Rapids. According to Intel's previously announced product plan, Diamond Rapids will adopt the Intel 18A-P process, integrate up to 256 CPU cores, support 16-channel memory and 128 PCIe 6.0 lanes, and introduce technologies such as Advanced Performance Extensions (APX) and enhanced Advanced Matrix Extensions (AMX).

Leiphone Leiphone