Cerebras SpeedHub

Evidence that the advertised speed is oversold. decode (wafer) is the number Cerebras markets (~1000 tok/s); roundtrip is what your application actually sees end-to-end. The reality ratio below is roundtrip ÷ advertised.
Fleet median decode (wafer)
tok/s
the advertised / marketed number
Fleet median roundtrip
tok/s
what your app actually gets
Reality ratio (roundtrip ÷ advertised)
green ≥80% · amber 40–80% · red <40%

Fleet

One row per worker. Location is editable inline (saved on change). auto-refresh every 15s
Status Location Region Server Last run decode
wafer
roundtrip
app
TTFT
ms
net rtt
ms
Promise vs reality Enabled Mode Schedule
UTC
loading…

Runs

Click a run to expand its per-output-size speed curve.
Time (UTC) Mode Trig decode roundtrip TTFT net tcp net tls net rtt rpm tpm 429
pick a server to see its runs