Cerebras
The startup building chips the size of dinner plates — because training AI models demands more silicon than any GPU can deliver.
NVIDIA's most powerful AI chip has 80 billion transistors, originally designed for gaming graphics workloads. Seribris asked, "What if we designed a chip from scratch "specifically for AI computation? "They built the largest chip ever manufactured, "four trillion transistors, "50 times more than NVIDIA's most advanced chip. "The entire wafer becomes the processor. "The chip is wafer-sized with more transistors "than any chip ever built. "Every design decision optimizes for AI speed. "Transistor connections are dramatically shorter. "Memory bandwidth is massive. "The computation memory bottleneck "that plagued GPU training is eliminated. "Training that takes weeks on GPU clusters "takes days on Seribris. "The speed up is not two times, "but 10 times on certain workloads.
Watch the full reel free on MoonReelz — moonreelz.com