OpenAI Runs Jalapeño ASICs With AMD Turin Host CPUs

OpenAI is running its Jalapeño inference ASIC alongside AMD processors, according to reporting from Tom’s Hardware cited by AI Weekly. The systems use AMD EPYC “Turin” CPUs as their host silicon, putting the custom OpenAI accelerator in a server configuration that does not rely on Nvidia’s processors.

Each host system reportedly contains two Turin-class CPUs and 1.5 terabytes of DRAM. The arrangement provides the surrounding computing platform for Jalapeño, OpenAI’s custom chip for inference workloads. The information offers a glimpse of how the company is assembling the infrastructure needed to operate its own silicon, as well as the alternatives it is considering for the components around it.

A two-part server design

An inference ASIC is built for the computation required when an AI system produces an answer or prediction. The host processor is a separate part of the system, responsible for supporting the accelerator and providing the broader server platform. In OpenAI’s Jalapeño deployment, AMD’s EPYC Turin CPUs fill that host role.

The reported configuration has two Turin processors in each host and 1.5TB of memory. Those specifications are the clearest disclosed details about the machines carrying the ASIC. They also show that the choice of accelerator does not determine every other component in the server: OpenAI’s own inference silicon is being paired with a host platform supplied by AMD.

That distinction matters because discussions of AI hardware often focus on the accelerator alone. In practice, a system also depends on the processors, memory and server design that allow the accelerator to operate. OpenAI’s reported setup identifies both sides of that arrangement: Jalapeño is the custom inference component, while Turin provides the host CPUs.

Why AMD is being used

Richard Ho, OpenAI’s vice-president and head of hardware, described the decision to use Turin as a pragmatic effort to reduce risk, according to the reporting. That explanation presents the processor selection as a practical engineering choice rather than a declaration that one supplier has replaced another across OpenAI’s infrastructure.

The available account does not say that AMD is supplying the Jalapeño ASIC itself. Instead, AMD’s EPYC processors serve as the host silicon in the systems where the OpenAI chip is deployed. Keeping those roles separate is important: the report concerns the CPU platform surrounding Jalapeño, not the manufacture of OpenAI’s custom accelerator.

The reported use of AMD hardware also indicates that OpenAI’s custom chip is being tested in a mixed-supplier environment. Jalapeño comes from OpenAI’s own hardware effort, while the host processors come from AMD. The result is a server design built from components with different origins and functions.

Nvidia remains part of the comparison

Nvidia’s Vera standalone platform was the alternative discussed in Ho’s comments. He judged it to be “a little bit behind” in maturity, as reported by Tom’s Hardware and AI Weekly. That assessment helps explain why OpenAI selected Turin for the hosts in the Jalapeño systems described by the report.

The comment is narrower than a general verdict on Nvidia’s hardware. It relates specifically to Vera’s maturity as the alternative under consideration for this deployment. The report does not say that Nvidia hardware is absent from all OpenAI systems, nor does it establish that AMD will be the company’s permanent host-CPU supplier.

What it does show is a particular stage in OpenAI’s hardware strategy. The company has a custom inference ASIC running in systems with two AMD EPYC Turin processors and 1.5TB of DRAM per host. Its hardware leadership has also compared the maturity of Nvidia’s Vera standalone option with the practical needs of the current design.

For now, the disclosed details are limited, but they establish the basic architecture: OpenAI’s Jalapeño chip is being deployed with AMD Turin CPUs around it, while Nvidia Vera is being viewed as a less mature alternative for that host role. The choice reflects the immediate requirements of getting a custom inference system into operation, rather than a broader announcement about every platform OpenAI may use.