Tigris | 会讲课教授是好老师 · 2025-12-26 · 帖子

People keep turning $NVDA Groq’s SRAM story into a meme: “More SRAM = no more HBM.” - Wrong, Why? Quick question: have you ever cooked and wondered why you’re “slow”? I t’s rarely the stove. It’s the constant trips to the fridge. SRAM is the prep counter next to the stove. Put the hot ingredients and tools within arm’s reach, and the workflow becomes smoother, more predictable. That’s the real Groq signal: not “less thinking”, but less waiting inside the system. But a bigger prep counter doesn’t eliminate the fridge. HBM is the fridge + pantry: the capacity and bandwidth you need to serve many customers, long contexts, and high concurrency. So no, SRAM doesn’t “kill” HBM. SRAM buys determinism. HBM buys scale. In the inference era, if anything, scaling real deployments can make HBM even more important. $MU I turned this into a simple visual. Link in first reply.