Drive down the cost per token of Gemini inference by optimizing throughput, accelerator utilization, batching, scheduling, and KV cache efficiency. Achieve these gains through expert optimization and by building agentic workflows that enable AI agents to discover, implement, and validate performance improvements at scale. Be among the first to bring Gemini to new TPU generations by characterizing hardware at launch, tuning kernels, sharding, and serving configurations for specific memory hierarchies and interconnects, and delivering production-ready performance.
Share these insights to influence future TPU hardware and compiler roadmaps. Implement model compression techniques—such as quantization and sparsity—to reduce the cost-per-token for production Gemini models. Evaluate new architectures, optimizers, and training methods to achieve target quality with reduced compute requirements.
Job description provided by Google.
View the original posting
