AI Performance Engineering Meetup (Atlanta)

Monday, October 19, 2026 · 9:00–10:00 AM PDT · Online
Zoom link: https://us02web.zoom.us/j/82308186562
Join us for three perspectives on AI performance: getting more from GPUs, turning agent failures into learning, and exploring faster network transport.
Talk #0: Introductions and Meetup Updates
by Chris Fregly and Antje Barth
Talk #1: Why Does Serving More Users Change GPU Utilization?
by Rajat “Raj” Dandekar, Co-founder of Vizuara
In this 5–10 minute technical segment, Raj will use a visual explanation and code to explore how concurrent users change GPU utilization. He’ll connect vLLM measurements on an H100 to the roofline model and briefly introduce Vizuara’s Kernel Engineering workshop.
Bio: Rajat Dandekar is a co-founder of Vizuara. He holds a PhD in Mechanical Engineering from Purdue University and BTech and MTech degrees from IIT Madras.
Talk #2: From Agent Traces to Agent Training: Closing the Continuous Learning Loop
by Rustem Feyzkhanov, Engineering Manager at Snorkel AI
Production traces show where agents fail, but how do we turn those failures into learning? This talk explores a practical loop from traces to reproducible simulations, verifiable rewards, and model improvement through fine-tuning or reinforcement learning. We’ll also look at when it’s better to change prompts, tools, or the agent harness instead of updating model weights.
Bio: Rustem Feyzkhanov is an Engineering Manager at Snorkel AI.
Talk #3: Homa: A Fast Lane for Small Messages
by John Ousterhout, Professor Emeritus of Computer Science at Stanford University
John will explain how Homa can dramatically reduce small-message latency and explore its potential relevance to AI workloads. Drawing on his expertise in transport protocols, he’ll discuss what Homa could offer when small-message latency matters. Its importance for AI workloads remains an open, workload-dependent question.
Bio: John Ousterhout is a Professor Emeritus of Computer Science at Stanford University, the creator of Tcl, and a systems researcher whose current work focuses on the Homa transport protocol.
Zoom link: https://us02web.zoom.us/j/82308186562
Related Links
Github Repo: http://github.com/cfregly/ai-performance-engineering/
O’Reilly Book: https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Algorithms/dp/B0F47689K8/
YouTube: https://www.youtube.com/@AIPerformanceEngineering
Generative AI Free Course on DeepLearning.ai: https://bit.ly/gllm

Monday, November 16, 2026 · 9:00–10:00 AM PST · Online
Zoom link: https://us02web.zoom.us/j/82308186562
Join us for two deep dives into hidden AI performance costs: surprise PyTorch recompilations during inference and compute, memory, and communication bottlenecks in Jev training.
Talk #0: Introductions and Meetup Updates
by Chris Fregly and Antje Barth
Talk #1: Overcoming the PyTorch Recompilation Dilemma in Real-Time Inference Workloads
by Chaim Rand
`torch.compile` can accelerate steady-state inference, yet a guard failure on a new input shape or layout can trigger an expensive recompilation in the request path. Drawing on his latest post, Chaim will compare six approaches using a dynamic graph-neural-network example: cache preloading, bounded recompilation with eager fallback, Dynamo and input-layout tuning, warmup seeding, background compilation, and ahead-of-time compilation with `torch.export` and AOTInductor. He’ll explain their tradeoffs and how the right choice depends on the model, input distribution, runtime environment, and latency requirements.
Bio: Chaim Rand is a machine-learning algorithm developer who writes about PyTorch and practical AI/ML performance optimization.
Based on: Overcoming the PyTorch Recompilation Dilemma in Real-Time Inference Workloads
Talk #2: From Profiles to Throughput: How TypeSafe Optimized Jev with Zymtrace
by Israel Ogbole, Co-founder and CEO of Zymtrace
Israel will explore the latest Zymtrace/TypeSafe case study on training and serving Jev. He’ll discuss how continuous CPU/GPU profiling and MCP-accessible profile data supported an iterative optimization loop across compute, memory, communication, and inference synchronization. The published case study reports approximately 46% higher training throughput on the same hardware after three successive changes, plus approximately 10% better inference performance after removing a hidden device-to-host synchronization. The talk will focus on how the team found each bottleneck, measured each change, and identified the next constraint.
Bio: Israel Ogbole is co-founder and CEO of Zymtrace, which develops continuous CPU/GPU profiling tools for performance optimization.
Based on: How TypeSafe achieved 46% higher GPU throughput training Jev — with the help of Zymtrace
Zoom link: https://us02web.zoom.us/j/82308186562
Related Links
Github Repo: http://github.com/cfregly/ai-performance-engineering/
O’Reilly Book: https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Algorithms/dp/B0F47689K8/
YouTube: https://www.youtube.com/@AIPerformanceEngineering
Generative AI Free Course on DeepLearning.ai: https://bit.ly/gllm

Zoom link: https://us02web.zoom.us/j/82308186562

Talk #0: Introductions and Meetup Updates
by Chris Fregly and Antje Barth

Talk #1: GPU, PyTorch, and CUDA Performance Optimizations

Talk #2: GPU, PyTorch, and CUDA Performance Optimizations

Zoom link: https://us02web.zoom.us/j/82308186562

Related Links
Github Repo: http://github.com/cfregly/ai-performance-engineering/
O'Reilly Book: https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Algorithms/dp/B0F47689K8/
YouTube: https://www.youtube.com/@AIPerformanceEngineering
Generative AI Free Course on DeepLearning.ai: https://bit.ly/gllm