Case studies & benchmarks
Real workloads, real numbers — GPU efficiency, power-bill, and training-cost outcomes from Pebble deployments.
AMD × Pebble · New
AMD's Ted Marena puts Pebble on stage — +86.5% throughput, −17.5% power on Instinct™ MI350X.
Pebble Sonar on 8× AMD Instinct MI350X GPUs nearly doubled inference throughput while drawing less wall power than the default configuration. Featured by AMD at Plug and Play San Jose, May 2026.
LLM Inference Benchmark
Maximizing GPU Efficiency: NVIDIA MPS vs Dedicated GPU Allocation for LLM Inference
50% GPU cost reduction with only 7.5% throughput loss when sharing via MPS — a benchmark on Qwen3-4B-FP8 across NVIDIA A100-80GB GPUs.
Healthcare AI
IOTAIRx cuts AI training cost 40% and power 48% with Pebble Sonar
40% infrastructure cost reduction and 48% power cut on long-running DeepSeek fine-tuning jobs running on NVIDIA A100s in GKE.
