The Pulse
Intel Xeon 6990E+ Handles 27% More Sandbox Requests
Intel says its Xeon 6990E+ system sustained about 313 concurrent CubeSandbox creation requests before average latency exceeded 200 milliseconds, compared with about 246 on the tested AMD EPYC 9965 system. The comparison covers sandbox provi
AI.info Team · Written with AI ·
Intel says its Xeon 6990E+ system handled about 313 concurrent sandbox-creation requests before average latency exceeded 200 milliseconds, roughly 27% more than the approximately 246 requests supported by the AMD EPYC 9965 system in the company’s test. Intel also reported that its Xeon 6980P system stayed below the same latency threshold at about 277 concurrent requests, or approximately 12% more than the tested EPYC system.
Intel measured when sandbox creation crossed a latency target
Sandboxing gives an agent an isolated environment in which to run code or invoke tools. In Intel’s description of a typical workflow, the environment must be provisioned and initialized before task execution, and it may be torn down when the task is done. Creating and removing these environments repeatedly can make provisioning speed relevant to responsiveness and the number of simultaneous requests a system can handle.
“In a typical agentic AI deployment, sandbox creation sits directly on the critical path of task execution.”
Nithya Rao, System and Software Optimization Engineer, Intel Corporation, writes.
Intel tested Tencent Cloud’s open-source CubeSandbox using its cube-bench create/delete workload. The workload sent sandbox-creation requests through the CubeSandbox API while aggregate concurrency increased from 100 to 350. Intel used 200 milliseconds as the service-level target for delivering a usable sandbox and tracked average creation latency as concurrency rose. The reported request counts mark where average latency crossed that target.
The tested systems had different configurations
The Xeon 6990E+ test used a one-node Supermicro SYS-212HA-TN system with a 288-core processor and 768GB of DDR5-6400 memory. The Xeon 6980P test used Intel’s one-node Avenue City system, a 128-core processor and 768GB of DDR5-8800 memory. Intel tested the 192-core AMD EPYC 9965 in a one-node Supermicro AS-2126HS-TN system with 768GB of DDR5-6400 memory. Intel says the tests were conducted as of September 2026.
Each sandbox was provisioned with 2,000 milliCPU, equivalent to two CPUs, and 2,000 mebibytes of memory. Intel says the benchmark used a common CubeSandbox template across the platforms, with instance type cubebox and template version v2. The test ran on CentOS Stream 9. The configurations also differed in server model, memory speed and components such as networking equipment and storage, as listed in Intel’s results.
The result is limited to sandbox provisioning
The comparison is a result from Intel’s specified test setups, not a comparison of processors surrounded by identical hardware. Intel’s published figures describe average sandbox-creation latency as concurrent requests increased; they do not show how long an agent takes to complete an end-to-end task. The article also does not provide results for full agent workflows or deployment costs.
The benchmark therefore supports a narrower conclusion: on the systems Intel tested, the Xeon 6990E+ and Xeon 6980P sustained more concurrent CubeSandbox creation requests than the tested EPYC 9965 system before average latency exceeded 200 milliseconds. It does not establish that either platform makes every part of an agent deployment faster.