2 ms·
Cerebras CS-3 wafer-scale 1e6 cores 125 PFLOPS 25kW (30kA)
- burnt-resistor 11mo ago- 4e12 transistors in 46,225 mm^2 (8.65e7 transistors/mm^2), basically an entire wafer squared off. - 5 nm process in current third gen. - RAM is spread out "IRAM"-style. - Most of the server packaging space is taken up by DC power supplies. - No decoupling capacitors. - Executes dummy operations to prevent "inductive transient"-like backlash from changing current draw too quickly. [I cry in reversible computing.] - Custom core, not ARM. - "~20x faster" than NVIDIA GPUs. - In customer production use now including US National Labs. - 2 WSE-3 systems per rack, so that's 50 kW/rack and therefore needs watercooling or rack-dedicated airhandling.