Earnings
Home›Earnings›Analyst Ratings›Cerebras targets faster AI inference with wafer-scale…
Cerebras targets faster AI inference with wafer-scale chips
The company says embedding 44GB of on-chip SRAM helps eliminate data shuttling that slows token generation, aiming for up to 14x faster output than conventional setups.
AI spending is shifting from training clusters toward real-time inference, where how quickly systems generate tokens can determine whether products work for users, according to MarketBeat Ratings.
Cerebras Systems is positioned around that change with a wafer-scale architecture designed to address the “memory wall” seen in traditional chips, which rely on repeated moves of data between compute and external memory during the decode phase.
The company says its approach embeds 44GB of static random-access memory directly on the chip and uses a single-silicon-wafer design, with the goal of delivering tokens up to 14 times faster than conventional setups.
Cerebras also points to infrastructure partnerships and an inference approach that separates prefill and decode, including an AMD alliance linking Helios rack platforms to Cerebras hardware.