Jim Keller

Jim Keller looks with interest at possible agreements with Intel or Qualcomm and, at the same time, relaunches Tenstorrent's ambitions in the artificial intelligence market. The company's CEO confirmed that he met with top executives from both companies, explaining that he hopes for an important business agreement to enhance the internally developed RISC-V intellectual property.

At the same time, Keller reiterated his belief that Tenstorrent can surpass competitors such as Cerebras and NVIDIA both in performance and in the overall cost of infrastructure.

During an interview with EE Times, Keller also addressed rumors circulating in recent months regarding Intel and Qualcomm's potential interest in Tenstorrent. While not confirming any corporate operation (it has been reported that Qualcomm was interested in investing up to $10 billion), the executive stated he had meetings with the CEOs of both companies and hopes for a "major deal," particularly related to the licensing of RISC-V IP developed by Tenstorrent.

According to Keller, one of the main hyperscalers is also considering using the company's AI intellectual property for a smaller chip.

The statements come at a time of significant expansion for Tenstorrent, which in recent weeks unveiled the BlackHole Galaxy system, a platform designed for AI workloads. The company claims that the architecture offers a total cost of ownership (TCO) significantly lower than competing solutions such as the NVIDIA GB300, positioning itself as an alternative for data centers and infrastructure dedicated to inference.

Keller also commented on the recent IPO of Cerebras, stating that he does not consider it a threat. On the contrary, he believes that the IPO could help to increase visibility in the sector and stated that Tenstorrent intends to "beat them in everything," referring to both performance and economic efficiency of their respective platforms. This reference also relates to the results achieved by Cerebras with the Wafer Scale Engine CS-3 system in executing the Kimi K2.6 model with one trillion parameters, in which Keller asserts that large-scale implementations of the BlackHole servers achieve superior performance at a significantly lower cost.

One of the technical aspects addressed during the interview concerns the management of the KV Cache, a fundamental element for the performance of modern language models. Unlike some competing architectures that adopt dedicated solutions, Tenstorrent stores the cache directly in the DRAM of the same chips responsible for the decoding phase. When possible, the data is kept in the local SRAM of the Tensor chips, while in cases where capacity is insufficient, it is transferred to DRAM. Although this solution incurs a performance penalty compared to SRAM memory, Keller argues that it still represents a more efficient approach than architectures lacking local DRAM memory.

On the commercial front, Tenstorrent is also preparing for its IPO. According to Keller, investors show a strong interest in the company, which is meanwhile expanding its international presence and consolidating its supply chain. The manufacturer also reports a significant increase in orders for Galaxy servers: among these is a contract from an offshore client for a pod consisting of 96 Galaxy servers, equivalent to 3,072 BlackHole chips.

The growth in demand is also said to be favored by the difficulties some competitors face in meeting all market requests. Keller claims that an AI infrastructure worth about $100 million based on NVIDIA hardware could be realized with a cost close to $20 million by choosing Tenstorrent, thanks to a particularly competitive cost proposal. The company also states that it is working on the production of about 1,000 Galaxy servers, of which about half are already sold.