TileRT Takes the Top Spot on AgentX with AMD Instinct MI355X GPUs (www.tilert.ai)

TileRT has achieved a remarkable milestone by securing the top position on the InferenceX AgentX leaderboard, delivering 469 tokens per second (tok/s) with the GLM-5.3 model on an 8× AMD Instinct MI355X GPU setup. This performance surpasses the second-place NVIDIA GB300 NVL72 system by over 100 tok/s and highlights TileRT’s ability to maintain high generation speeds across extended contexts, a crucial factor for real-world coding-agent tasks. Unlike traditional benchmarks, which utilize fixed-length prompts, AgentX evaluates inference performance through realistic coding-agent trajectories, emphasizing the importance of sustaining efficiency in long-running processes. The significance of this achievement lies not just in speed but also in the innovative architectural optimizations that TileRT implements. By employing a persistent Engine Kernel, TileRT minimizes latency through continuous execution without traditional kernel boundaries, allowing for improved data prefetching and overlapping computation with communication. Such optimizations leverage AMD’s CDNA 4 architecture, which supports comprehensive software control over execution paths and memory management, resulting in enhanced performance even as context length expands dramatically. TileRT’s advancements signal an important evolution in managing AI workloads, reinforcing the vital role of inference speed in scaling complex agent interactions.