Meta Scraps Advanced AI Chip
Meta has reportedly abandoned its most advanced in-house AI chip project, a major retreat that deepens its reliance on Nvidia and AMD. The move signals that even tech giants are struggling to compete with the pace of dedicated chipmakers, reinforcing the strategic value of Apple's vertical integration in silicon.
The now-scrapped chip, internally codenamed "Olympus," was Meta's second-generation attempt at a powerful in-house training chip. Its predecessor, a simpler design codenamed "Iris," was also discarded. The failure of Olympus was attributed to its complex design, software stability issues, and the high risk of delays as Meta races to compete with OpenAI and Google. This retreat from custom training silicon follows earlier setbacks with Meta's first-generation inference chip, the MTIA v1. While deployed for some recommendation models, the MTIA, a 7nm chip running at 800 MHz, showed only marginal efficiency gains on low-to-medium complexity tasks and lagged behind GPUs on more complex workloads. To fill the gap, Meta is massively increasing its spending with external vendors. The company recently announced a multi-year, multi-billion dollar deal to rent Google's TPUs and has separately committed to purchasing millions of next-generation GPUs from Nvidia. A five-year, $100 billion partnership with AMD for its MI450 GPUs is also in place. The design for Olympus was reportedly based on a more complex architecture similar to Nvidia's, known as SIMT (single instruction, multiple threads), which is easier for software engineers to program but harder for hardware teams to design. This contrasts with the simpler SIMD approach of the failed "Iris" chip, which proved difficult for software engineers to program for training AI models. This struggle highlights the immense challenge and cost of developing cutting-edge AI accelerators, where design costs for a 5nm chip can exceed $500 million. It also underscores the strategic advantage of Apple's long-standing vertical integration, where deep co-design of hardware like the A-series and M-series chips and software provides unmatched performance and efficiency. The decision to scrap Olympus points to internal skepticism about matching Nvidia's capabilities and the immense engineering resources required to debug complex chips and manage power consumption effectively. This move suggests Meta is prioritizing immediate access to high-performance computing to keep pace in the AI race, even at the cost of long-term silicon independence.
