Abstract:
OpenAI claims that its new "Jalapeno" chip performed better in tests than Nvidia's current product lineup, highlighting the progress the company has made in developing its own AI processors. Richard, head of OpenAI chip business Ho said in an interview that the chip was compared to the Nvidia GB300, a test that considered the GB300 to be the most advanced product available. In the test, Jalapeno had the advantage in two indicators: the amount of AI workload it can handle per unit of power consumption, and the speed of response.
When OpenAI publicly tested Jalapeno this time, it used its own smaller open source model GPT‑OSS 120B, as well as third-party models from DeepSeek and Moonshot AI. On Moonshot’s model Kimi K2.5 1T, Jalapeno’s advantages are even more obvious.
OpenAI stated that in internal tests, Jalapeno also performed well when running some large-scale advanced models that have not yet been released. This suggests that the chip's design "will further increase in value" as workloads become larger and more complex.
OpenAI collaborated with Broadcom to develop Jalapeno. The company plans to use the chip to support its AI models starting later this year. This is part of its plan to build its own AI infrastructure and promote self-research of chips on a larger scale.
Broadcom produces customized chips for many customers. The two companies announced their cooperation last year and vigorously promoted the development speed of this processor in June this year, saying that it was developed at a record speed.
Currently, more and more companies are developing self-developed AI chips, and NVIDIA still dominates this field.
Normally, some chips have more processing power when running AI tasks, while others are better at responding quickly, but Jalapeno has the best of both worlds, Ho said.
OpenAI will decide which AI models run on Jalapeno, allowing customers to choose lower-cost or higher-performance solutions based on their needs.
Ho said: "In laboratory tests, Jalapeno showed strong performance in the high-throughput field, which means it can serve a large number of customers at a lower cost; at the same time, it also performed well in the low-latency field, which means that for customers who value response speed, its response time will be very, very fast." For data centers, electricity is one of the most important costs.
It should be noted that Jalapeno was not tested with Nvidia's latest generation Vera Rubin chips, which have just begun shipping. In addition, Jalapeno is not designed for AI model training, and training is exactly one of the areas where NVIDIA technology is best at.
Jalapeno is mainly oriented to the AI inference stage, that is, after the model completes training, it generates responses and performs tasks based on user prompts. The advantage in speed means that OpenAI can use it to achieve effects that in the past could only be achieved by processors using different memory types and chip architectures.
OpenAI still relies on Cerebras technology to run some models, and this type of chip is more suitable for smaller models. In comparison, Jalapeno is capable of handling much larger models. Ho said OpenAI's technology will not replace suppliers such as Cerebras in the short term.
Ho said: "Our demand for computing power is very huge, which is why we have signed contracts with so many different suppliers. This situation will continue for some time." He said that OpenAI hopes to publicly display its research and development results to promote innovation in the field of AI chips.
Other startups are exploring similar directions. Etched said last week it had secured a valuation of about $21 billion in a funding round and had begun shipping a low-voltage chip. MatX, founded by two former Google chip employees, is currently developing a semiconductor product with the same goal of combining high throughput and low latency.
OpenAI also said that the company used its own AI model to speed up the chip development process. The second-generation chip has also entered a more in-depth research and development stage. Ho said the company expects to complete tape-out of the chip "in the coming months," which is when the chip design enters the finalization stage.
At the same time, OpenAI has begun designing its third-generation chips, hoping to further reduce the cost of the AI infrastructure it is building on a large scale around the world. Ho said: "We have now reached a cost level and power consumption level that can reduce infrastructure costs. This is only the first step."
Although OpenAI is actively promoting self-developed chips and comparing their performance with NVIDIA products, Ho still specifically emphasized that OpenAI still regards NVIDIA as a key chip supplier.
"NVIDIA is a very good partner, and we still need to use a lot of NVIDIA chips."
Comments