myrtle.ai's VOLLO Sets New STAC-ML Records for Gradient Boosted Tree Inference
myrtle.ai has announced that its inference accelerator, VOLLO®, set new performance records in the STAC-ML Markets (Inference) benchmark tests for gradient boosted trees. The results showed a p99 latency of under 2 microseconds and throughput up to 50 million inferences per second on an AMD Alveo V80LL card within a Blackcore ICON server. These improvements represent over a 30% reduction in latency at the 99th percentile compared to previous bests, with five times higher throughput. The performance is particularly significant for electronic trading where low and deterministic latency directly impacts returns by enabling faster decision-making without compromising model accuracy or size. VOLLO now holds records for both gradient boosted trees and neural networks in terms of deterministic latency following earlier STAC Tacana results from April 2026. It has already been deployed successfully with major trading firms, generating alpha over hundreds of thousands of real-world trading hours.
Tickers: $AMD
Topics: AI inférence machine learning latence STAC Summit trading électronique