myrtle.ai indicates that its inference accelerator, VOLLO ®, set a new record in the performance tests of gradient boosting trees conducted from STAC to ML Markets ( Inference ), reducing the 99th percentile latency by more than 30% and increasing throughput to at least 5 times the previous best result. The certification results of STAC ® were announced today at the STAC Summit held in London.
The test was conducted on a server equipped with AMD Alveo ™ V80LL computing accelerators and Blackcore token issuance N 3132- SM. The p99 latency on all three models was less than 2 microseconds. The smallest model achieved a throughput of 50 million inferences per second, with a latency of only 1.77 microseconds.
In electronic trading, the delay between receiving market data and making decisions directly affects profits. myrtle.ai indicates that with low and predictable delays, companies can use larger and more accurate models without missing market opportunities, thus there is no longer a need to compromise between model quality and speed.
myrtle.ai CEO Peter Baldwin stated: "Trading companies want to use increasingly powerful models without sacrificing speed, and these results show that this is feasible. Developers can now test their models on VOLLO without the need for FPGA expertise and see the differences for themselves."
Following the release of the results of STAC Tacana in April, VOLLO now holds the record for both deterministic latency in decision trees and neural networks. The company states that this solution has been verified in a production environment and has run for hundreds of thousands of hours under real trading conditions, creating alpha for many large trading firms around the world. The flexibility of its model also makes it the preferred platform in the fields of telecommunications, network security, and national defense.
STAC - ML Markets ( Inference ) is a technical benchmark for real-time market data inference execution. This tool was designed by quantitative analysts and technical experts from large financial companies and is used to evaluate the performance, resource efficiency, and quality of any technology stack that can run the provided models. The complete results are contained in the STAC report ( SUT MRTL2026905 ), available at the URL www.STACresearch.com / MRTL2026905.
Machine learning developers can now evaluate the performance of their models on VOLLO without the need for FPGA tools or specialized knowledge. Please visit myrtle.ai / vollo-trees or contact [ email protected ].
About myrtle.ai
myrtle.ai is a company that focuses on AI and machine learning software, providing ultra-low latency inference accelerators based on major FPGA vendor platforms. Its accelerators are applied in fields such as financial transactions, wireless communications, LLM, speech processing, and recommendation systems.
The symbols “VOLLO”, “VOLLO Accelerator” and “VOLLO” are all registered trademarks of “myrtle.ai”. “STAC” and all names ending with “STAC” are trademarks or registered trademarks of “Strategic Technology Analysis Center”, “LLC”. The symbols “AMD”, “AMD Arrow”, “Alveo” and their combinations are all trademarks of “Advanced Micro Devices”, “Inc”.
