Myrtle.ai's VOLLO Breaks New Records for Gradient-Boosted Inference at STAC-ML

Myrtle.ai's VOLLO Sets New STAC-ML Records

In the competitive landscape of high-frequency trading, performance and speed are essential. Myrtle.ai recently announced record-breaking achievements with its VOLLO® inference accelerator, which has set new benchmarks on the STAC-ML Markets for gradient-boosted tree inference. Unveiled during the STAC Summit in London, these impressive results demonstrate Myrtle.ai's commitment to enhancing AI-driven trading systems.

At the core of VOLLO's success is its unprecedented ability to reduce the 99th percentile latency by over 30%, while simultaneously increasing throughput by a staggering five times compared to previous records. This means that trading firms can now process and analyze vast amounts of market data in real-time with remarkable efficiency.

Incredible Performance Metrics

The VOLLO accelerator operates on an AMD Alveo™ V80LL Compute Accelerator housed in a Blackcore ICON 3132-SM+ server configuration. This cutting-edge setup allows VOLLO to maintain p99 latencies below 2 microseconds across all tested models. For instance, the smallest model achieved the extraordinary feat of sustaining 50 million inferences per second while achieving a latency of just 1.77 microseconds. Such speed is crucial since in the fast-paced world of electronic trading, milliseconds can significantly impact profit margins.

Myrtle.ai's CEO, Peter Baldwin, emphasized the importance of these records for trading firms: "Latency is critical. With our VOLLO accelerator, traders can deploy larger and more accurate models without sacrificing speed. The ability to run sophisticated models can be a game-changer in making timely trading decisions."

Bridging the Gap Between Complexity and Performance

One of VOLLO’s standout features is its accessibility. Developers without extensive FPGA knowledge can test their own models against VOLLO and observe performance improvements firsthand. This democratization of technology enables trading firms to experiment with advanced strategies while still adhering to the speed demands of today’s markets. By leveraging VOLLO, firms can maintain model quality without compromising latency—a common trade-off in the industry.

The results of the STAC-ML Markets (Inference) benchmarks serve as a standardized measure for technology performance in real-time trading environments. Formed by leading financial technology professionals, STAC's rigorous assessment ensures that any claims about latency and model execution can be substantiated with transparency. You can find the full performance results in the STAC Report.

Proven Success Across Industries

Beyond financial trading, VOLLO has already shown its potential in various fields, including telecommunications, network security, and defense sectors. This versatility speaks to the accelerator's robust architecture and its ability to cater to diverse AI application needs. As the demand for real-time data analysis continues to grow, myrtle.ai is poised to capitalize on this overarching trend with its innovative technology.

In conclusion, Myrtle.ai's VOLLO has not only set a new standard for gradient-boosted tree inference on the STAC-ML benchmarks but has also reshaped the narrative around the intersection of speed and model complexity in trading technology. Continued enhancements and rigorous testing will ensure that myrtle.ai remains a leader in providing ultra-low-latency inference solutions to its clients.

For more information on how your models could benefit from VOLLO's capabilities, visit myrtle.ai/vollo-trees or reach out directly at [email protected].

About Myrtle.ai

Myrtle.ai specializes in AI and machine learning software, creating inference accelerators designed for FPGA-based platforms from various suppliers. With applications spanning financial services, telecom, LLMs, and more, Myrtle.ai is dedicated to advancing real-time data processing capabilities.