Inside Agentic AI Workloads
Agentic workloads repeatedly re-enter model inference over growing state. We characterize how trajectories, KV-cache retention, and tool use shape serving demand.
Gimlet Labs Blog
A blog about our lab's research on high performance AI systems.
Previous Posts
Showing 6 previous posts tagged Hardware.
Scaling low-latency inference with Gimlet Cloud and Cerebras
Gimlet brings Cerebras’ wafer-scale compute into Gimlet Cloud, enabling inference at up to 3,000 tokens per second while expanding the performance and efficiency available for agentic and real-time AI applications.
Low-Latency Inference with Speculative Decoding on d-Matrix Corsair and GPU
We evaluated running gpt-oss-120b with a 1.6B parameter speculative decoder on d-Matrix Corsair. Compared to the same speculative decoder on GPU and equivalent energy consumption, we've found that the Corsair-based solution delivers 2-5X end-to-end request speedup on configurations optimized for interactivity, and up to 10X end-to-end speedup for energy-optimized configurations.
The emerging role of SRAM-centric chips in AI inference
In this post, we'll discuss the major differences between GPUs and SRAM-centric accelerators (e.g. Cerebras, Groq, and d-Matrix), explaining why near-compute memory versus far-compute memory is the key tradeoff being made by these architectures, and what this means for inference workloads.
Benchmarking AI-generated CUDA kernels on an H100
We extended our kernel generation research to CUDA, benchmarking on an H100 where generated kernels achieve around 1.8X speedups over baseline PyTorch (including torch.compile).
Splitting LLM inference across different hardware platforms
Separating prefill and decode stages of LLM inference improves token throughput because their resource needs differ. Although most deployments use NVIDIA hardware for both stages, multivendor disaggregation can actually improve efficiency while maintaining SLAs. Based on our models using NVIDIA B200s and Intel Gaudi 3, common workloads can see 1.7X TCO improvement compared to single-vendor disaggregation.
Speeding up PyTorch inference on Apple devices with AI-generated Metal kernels
Our lab investigated whether frontier models can write optimized GPU kernels for Apple devices to speed up inference. We found that they can: our AI-generated Metal kernels were 1.24x faster across KernelBench v0.1 problems, and 1.87x faster across KernelBench v0 problems.