In this episode, Ian and George discuss NVIDIA’s $20 billion deal with Groq, analyzing its unusual structure, motivations, and implications for industry consolidation, while also reviewing NVIDIA’s upcoming Olympus CPU core and its technical choices. They wrap up with thoughts on the risks of vendor lock-in, upcoming CES announcements from major chipmakers, and some lighthearted holiday banter.
The episode opens with the hosts, Ian and George, sharing their holiday experiences and discussing various festive foods and drinks, including traditional dishes from their families and some unique UK treats. They also chat about tech-themed gifts, such as Lego kits of server hardware, and reminisce about collecting branded tech swag from industry events. This lighthearted introduction sets the stage for a deep dive into recent developments in the semiconductor and AI hardware industry.
The main topic is NVIDIA’s surprising $20 billion deal with Groq, an AI hardware startup known for its deterministic, low-latency inference chips. The hosts clarify that NVIDIA is not acquiring Groq outright but is instead licensing its IP, acquiring physical assets, and hiring most of the key staff, while Groq continues to operate its cloud business independently. They analyze the unusual structure of the deal, speculate on NVIDIA’s motivations—ranging from eliminating a potential competitor to acquiring unique technology or talent—and question whether the price tag is justified given Groq’s relatively old architecture and limited market traction outside the Middle East.
The conversation then shifts to broader industry consolidation, with NVIDIA’s recent acquisitions of other companies like ScheduleMD (the main developer behind the Slurm workload manager) raising concerns about vendor lock-in and the future of open-source software in high-performance computing. The hosts discuss the importance of workload managers like Slurm in multi-tenant environments and the potential risks of a dominant hardware vendor controlling critical software infrastructure. They note that while alternatives exist, replacing well-established tools is challenging due to decades of accumulated improvements and ecosystem support.
Next, the hosts review technical documentation for NVIDIA’s upcoming Olympus CPU core, which will power the Vera Rubin data center CPUs. They highlight the unusual approach NVIDIA is taking with simultaneous multi-threading (SMT), opting for more static resource partitioning within the core rather than the dynamic allocation seen in AMD and Intel designs. This could result in more predictable performance at the expense of peak throughput, which may be better suited for data center and AI workloads where determinism and regularity are valued over bursty single-threaded performance.
Finally, the episode previews upcoming industry events, particularly CES, where AMD, Intel, and NVIDIA are all expected to make major announcements. The hosts speculate on what each company might reveal, including potential Zen 6 teasers from AMD and new product launches from Intel and NVIDIA. They also offer practical advice for navigating CES in Las Vegas and reflect on the rapid pace of change and consolidation in the tech industry, closing with some light banter and reminders about supporting their channels and upcoming content.