Instinct™ MI350X
MiTACComputing
G8825Z5 servers
San Jose, California — June 26, 2026. Pebble (MyPebble Inc.), the company building the intelligence layer between AI infrastructure and the power grid, today released results from a joint demonstration with MiTAC Computing on AMD Instinct™ MI350X hardware. With Pebble Sonar™ managing the cluster, AI inference throughput climbed 28% on a strictly power-constrained rig — without drawing a single additional watt.
The demonstration
Power, not silicon, is increasingly the ceiling on how much compute an operator can deploy. The demonstration targeted that constraint directly. Running on MiTAC Computing infrastructure with AMD Instinct MI350X GPUs, a cluster built on MiTAC G8825Z5 servers was held to a strict 6 kW power ceiling while serving Meta's Llama 3.1 405B — one of the largest publicly available models. Pebble Sonar profiled the live workload, then recovered roughly 15% of power headroom the default configuration had been wasting as heat rather than useful compute. That reclaimed budget was used to bring two additional GPUs online inside the same 6 kW envelope.
The result
| Metric | Baseline | With Pebble Sonar | Change |
|---|---|---|---|
| Total throughput | 4,582 tok/s | 5,866 tok/s | +28% |
| Energy efficiency | baseline | +14.7% | improved |
| Total rig power | 6.0 kW | 6.0 kW | unchanged (capped) |
| GPUs online | 6 | 8 | +2 free GPUs |
| P99 TTFT | baseline | statistically unchanged | no regression |
Latency and thermals held steady. The 99th-percentile time-to-first-token was statistically unchanged versus the baseline, and GPU temperatures stayed well within MI350X operating limits.
“Operators are being told the only way to serve more AI is to buy more power, and the wait for new grid capacity can stretch for years. This demonstration shows another path. The headroom is already sitting inside the power you have. Sonar finds it, and turns stranded GPUs into deployable capacity at zero additional energy cost.”
“MiTAC is committed to delivering maximum performance-per-watt across our AI infrastructure. Validating these significant, software-driven efficiency gains by deploying Pebble Sonar on the MiTAC G8825Z5 server racks featuring AMD Instinct MI350X GPUs proves that smarter power management is the key to unlocking sustainable, next-generation compute.”
How Pebble Sonar works
Sonar runs alongside the inference stack. It continuously profiles real-time power draw, GPU utilization, clock frequencies, and token throughput across every GPU in the rig, then builds a live model of the workload's power-efficiency curve. From that model it identifies the operating point at which each watt produces the most useful compute — jointly tuning hardware power caps and serving-engine parameters, rather than treating them as independent knobs.
It also scales down to 70B
In a companion test on the same AMD Instinct MI350X platform, Sonar applied the same approach to an 8-GPU Llama 3.1 70B deployment, improving tokens-per-watt by 82.4% by identifying a more efficient per-GPU operating point. The two results together — 70B and frontier-scale 405B — point to the same conclusion: the next efficiency frontier in AI inference is not a bigger server, but smarter power management. For operators running frontier models on AMD Instinct hardware, that means more AI output from the infrastructure and the power they already own — without new procurement cycles, infrastructure upgrades, or model changes.
Full methodology and per-trial measurements are available in the accompanying technical case study at gopebble.com. The original wire release is published at mitaccomputing.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. A second product, Pebble Flex™, enables GPU clusters to participate in grid flexibility programs — automatically reducing and restoring power in response to utility signals without interrupting AI workloads, turning stranded grid capacity into a deployable asset. Learn more at gopebble.com.
About MiTAC Computing
MiTAC Computing Technology Corp., a subsidiary of MiTAC Holdings, specializes in AI, HPC, cloud, and edge computing. With end-to-end capabilities from R&D and manufacturing through global support, MiTAC Computing provides customized platforms for hyperscale data centers, HPC, and AI applications. Learn more at mitaccomputing.com.
Media contacts
Pebble (MyPebble Inc.)
Pradeep Gaddam
pradeep.gaddam@gopebble.com
240-505-9239
MiTAC Computing
Raymond Huang
raymond.y.huang@mitaccomputing.com
510-651-8868 ext. 6915