Jim Keller: 'AI Still Obeys the Old Laws of Compute'
Jim Keller discusses Tenstorrent's Blackhole chip, emphasizing that AI compute still follows traditional laws like Moore's Law and Dennard scaling. He outlines the company's focus on scalable architecture for AI inference and training, and hints at future IPO ambitions while criticizing current AI hardware inefficiencies.
Background
Jim Keller is a legendary microprocessor architect who worked on Apple's A4/A5 chips (used in early iPhones/iPads), AMD's Zen architecture (which revived the company's server CPUs), and Tesla's self-driving hardware. He is now CEO of Tenstorrent, a Toronto-based AI chip startup competing with Nvidia.
- Tenstorrent is building AI accelerators but takes a different approach from Nvidia: rather than relying on a proprietary interconnect (NVLink) and software lock-in (CUDA), it uses open-source standards like RISC-V and Ethernet-based networking to link chips together.
- The "Blackhole" chip mentioned is Tenstorrent's current-generation AI processor. Keller argues that scaling AI requires better system-level engineering (memory bandwidth, networking, cooling) — not just bigger individual chips. This contrasts with the prevailing industry bet that more transistors alone will solve AI's growth problems.
- Tenstorrent has IPO ambitions, meaning it plans to go public. Keller's reputation gives the company credibility, but it faces an uphill battle against Nvidia's entrenched ecosystem and massive R&D spending.