ThunderAgent Engine: 2× Throughput, Near-Linear Multi-Node Scaling for Agent Workflows
togethercompute · x · 2026-07-30
Together Compute's ThunderAgent engine targets the inference bottlenecks in complex agent workflows. Traditional request-level engines fail to recognize that a series of LLM calls belongs to a single workflow, causing severe KV cache thrashing and cascading recomputation.
ThunderAgent tackles this by introducing a program-level scheduling abstraction:
- Smart Scheduling: Treats each agent workflow as a schedulable program, tracking its phase, KV footprint, and node placement. It pauses low-priority workflows under memory pressure to significantly boost cache hit rates.
- Single-Node Performance: On a single 8×H100 node with batch 192, compared to SGLang dropping to 390 tok/s (65s latency), ThunderAgent maintains 803 tok/s (10.6s latency)—delivering 2× throughput and roughly 6× lower latency.
- Multi-Node Scaling: Scaling from 16 to 64 GPUs increases processing speed from 671 to 2,248 steps/min, achieving near-linear scaling. The performance advantage grows with cluster size (1.79× at 2 nodes, 2.39× at 8 nodes).
- Seamless Integration: Designed as a drop-in solution compatible with existing configs like KV offloading and speculative decoding. Already adopted by SkyRL and NVIDIA Dynamo.
Related event: ThunderAgent Doubles Agent Inference Throughput(6 posts)→
More from coding & agent
- AI Bug Fixes Often Incomplete, Creating Messy Open Source Security — curious_vii · 2026-07-30
- Codex Tip: Turn Prompts into Repeatable, Scriptable Workflows — reach_vb · 2026-07-30
- ShadKit Open Source: Porting shadcn/ui and Vercel AI Elements to SwiftUI — jasonkneen · 2026-07-30
- Test: Enabling Two API Settings Boosts GPT-5.6 ARC-AGI-3 Score by 3x — sandersted · 2026-07-30
- dspy-monty-interpreter v0.3.0 Released: Multithreading and Isolated Execution — dbreunig · 2026-07-30
- Current AI Agent Memory Systems Are Just Hacky RAG Wrappers — Trick_Stretch_4746 · 2026-07-30