Scene 1 · 0:00
Every AI request triggers two decisions Every AI request hits
two questions before it's answered.
Scene 2 · 0:05
Question 1
Which GPU?
summarize this document…
Question 2
How hard should it run?
700
watts
Scene 3 · 0:13
Inference Router
Picks GPU by load
Power Caps
Set 6 months ago
Two systems. Never talk to each other.
1000
watts
Wasted electricity +38 %
Scene 4 · 0:21
Inference Router
tuned by Pebble
pebble
Power Caps
tuned per workload
Pebble tunes both. Together.
Power-cap each GPU at its efficiency knee — reclaim the stranded headroom.
Scene 5 · 0:30
Pebble's atlas — every model × every GPU.
0
profiles · updated continuously
Scene 6 · 0:42
Real-time · per-request optimization
Cluster tokens / watt
2.84
Requests / sec
142
SLA breaches
0
Pebble decisions / sec
412
Scene 7 · 0:50
The math of the efficiency knee Power-cap each GPU at its knee — fit more in the same envelope.
Default 1000 W cap 6 × 1000 W = 6 kW
1000W 1000W 1000W 1000W 1000W 1000W
↓ Sonar caps at the efficiency knee
Sonar 750 W cap 8 × 750 W = 6 kW
750W 750W 750W 750W 750W 750W 750W 750W
+2 GPUs same envelope · same SLAs · +33% inference throughput
Scene 8 · 0:58

Same power envelope. Same SLAs.
More GPUs.

+2 GPUs
per 6 deployed · same kW
+82 %
Tokens per watt
0
SLA breaches
Scene 9 · 1:03
pebble
The first full-stack AI inference optimizer.
Request a Demo → Watch the Flex explainer →
← Back to Pebble