San Jose, California — July 9, 2026. Pebble (MyPebble Inc.) is featured in Google Cloud's public launch of AlphaEvolve, alongside other organizations using the system in production. In the section titled “Pebble: Optimizing serving performance on GPUs,” Google's authors hand the mic to Pebble's Head of AI, Keval Shah, to describe how Pebble used AlphaEvolve to sharpen the model that underpins Pebble Sonar™'s optimization decisions.
The problem AlphaEvolve solved for Pebble
Pebble Sonar decides, in real time, where each GPU should sit on its power-efficiency curve. Making that decision well requires an accurate latency model — a mathematical function that predicts how a given serving configuration will perform on a given piece of hardware. The industry-standard starting point — NVIDIA's AI Configurator latency model — ships with a single static empirical correction factor (0.8) applied uniformly across all workloads, and does not differentiate FP8 from BF16 efficiency behaviour. On real production traffic those simplifications caused Sonar's recommended configurations to drift away from the true optimum.
Rather than hand-tune a better formula, Keval's team pointed AlphaEvolve at the problem. AlphaEvolve autonomously discovered new GPU performance-modeling formulations directly from Pebble's training prior. The learned function replaces the static 0.8 correction with something that adapts across quantizations and hardware.
“AlphaEvolve solved this by autonomously discovering GPU performance modeling formulations directly from our training prior. This Gemini-powered evolutionary approach drastically cut our model errors — delivering a 56% relative error reduction.”
Why the mention matters
AlphaEvolve is Google DeepMind's evolutionary program-synthesis system, now generally available on Google Cloud. Google's public launch post is deliberately selective about who it names: the companies chosen are the ones Google considers meaningful early production users. Pebble sits in that group. For customers evaluating Pebble Sonar, the mention is a Google-verified signal that the optimization decisions Sonar makes on their infrastructure are backed by a model that has been sharpened using frontier AI research — not a static heuristic.
What Pebble intends to do with it
The learned efficiency function is being folded directly into Pebble Sonar's production decision loop. Beyond this specific fix, Keval's team plans to keep AlphaEvolve in the loop as a continuous instrument — using it to remap the efficiency landscape as new accelerators (AMD Instinct™ MI350X and MI400X, NVIDIA Blackwell and next-generation Hopper) reach production — so Sonar's recommendations continue to track the frontier without a manual retuning cycle.
The full Google Cloud launch post is available at cloud.google.com →.
About Pebble
Pebble (MyPebble Inc.) builds the intelligence layer between AI infrastructure and the power grid. Its flagship product, Pebble Sonar™, continuously profiles GPU clusters in real time, identifies the optimal power operating point for each workload, and unlocks throughput gains that static, hardware-default configurations leave on the table. Learn more at gopebble.com.
Media contact
Pebble (MyPebble Inc.)
Pradeep Gaddam
pradeep.gaddam@gopebble.com
240-505-9239