BaseThesis x Groq Teardown: From Silicon to Systems
Hosted on LumaFetched 2 months ago
Saturday, December 20, 2025
to Saturday, December 20, 2025
Artificial IntelligenceMachine LearningInnovation
Student only
This hackathons is only open to students. Double check the event page for more information as this may mean only those from a particular university/country are eligible.
Event Type
in person
Organizers
Inference is often treated as interchangeable APIs and optimisations often happen at model level (better prompts, different models, clever techniques) while the actual constraint lives in execution architecture (memory hierarchy, numerical precision, compiler decisions, scheduling determinism).You write a prompt, the model generates tokens, your application receives text. Between these steps lies an entire computational substrate that determines whether the system works or fails in production. In this teardown we go into the architectural decisions at the silicon level cascade all the way to whether users adopt or abandon your product. Why does one voice agent feel conversational while another feels broken, even using the same model? Why do some real-time systems hold up under load while others degrade unpredictably?The answer lives in how the system executes inference (LPU vs GPU, SRAM vs DRAM, deterministic vs dynamic scheduling). This teardown exists to close that gap.What This Is? A deep technical exploration of low-latency AI systems with Groq. We're investigating why Groq behaves differently, how architectural choices affect model behavior, and where those advantages matter in production. Groq's LPU architecture eliminates the memory bottleneck through SRAM-based design and deterministic scheduling. We're investigating why different execution architectures produce different behaviors in production. Agenda: 11 AM - 12 PM : Settle In and & Meet Folks12 PM - 1 PM : Teardown with DJ Biswas, Demo + Code Review1 PM - 1.15 PM : Q&A1.15 PM - 2 PM : Mixer + Light Lunch Provided Who This Is For Builders who: Shipped production systems (real users, real constraints)Hit real limits- rate caps, latency budgets, unpredictable behavior under loadThink in tradeoffs- speed, cost, qualityQuestion received wisdom: "temperature 0.7" isn't satisfying without understanding whyCare about system execution, not just model prompting Why Attend? Develop hardware-aware AI thinking and go from "How do I prompt better?" to "How does the system execute my model?" You'll understand: How inference hardware shapes model behaviorWhy memory architecture determines which interactions are possibleWhere numerical precision decisions happen and why they matterHow to think in systems, numerics, and real-time constraints You'll bridge the gap between: Model-level thinkingSystem-level executionProduct-level impact What to bring? Laptop, curiosity, willingness to question assumptions.About the speaker DJ Biswas is Groq's AI Solutions Architect. He's spent a decade understanding how data structures map to memory, how queries translate to cache behavior, how indexing decisions affect disk I/O. Works at layers where physics becomes computation becomes application behavior. Passionate about questioning received wisdom, navigating tradeoffs, and understanding production reality, and first-principles thinking. He thinks at the layer where physics becomes computation becomes user behavior.We believe that its important for builders to think in terms of systems, numerics, and real-time constraints, not just model APIs.Are you one of them? Plus: Attendees get priority access to apply for BaseThesis's January hackathon including cloud credits, TTS model access, and compute credits so you can immediately apply what you learn to build production systems. Note: Tea, Coffee, Snacks and Light Lunch Provided