The Dilemma of AI Accelerator Chip Manufacturers
The Dilemma of AI Accelerator Chip Manufacturers When scaling out - the high-end market: To achieve optimal performance, it becomes an incredibly challenging task to design not only the accelerator chip and board but also incorporate HBM, high-bandwidth low-latency networking, a compiler that optimizes workloads for multiple chips within multiple boards, tiles, and layers on top of large-scale equipment. When not scaling out - the low-end market: Even if we simplify the workload optimization problem to L=M=1 and tolerate some inconvenience with the compiler, we still need to produce chips with high quality and performance at a lower cost than the RTX 4090, reaching around 4N in terms of production volume. Can NVIDIA lower the manufacturing cost of the RTX 4090 to compete with this market? It seems highly unlikely. In the short term, abandoning option 1 and pursuing option 2, focusing on a narrow market segment with low-cost high-performance solutions, seems to be the only viable ...