Myrtle.ai Cuts Inference Latency in Half with FPGA Benchmark Record
Myrtle.ai's VOLLO software achieved 2-microsecond latency (99th percentile) across financial machine learning models in a STAC benchmark audit released today. The result halves the previous record for inference speed in this category, according to the benchmark authority that sets standards for financial technology performance.
STAC-ML (Markets) Inference measures how quickly systems can run machine learning models on real-time market data. The benchmark was designed by quantitative traders and technologists from major financial firms to test performance, efficiency, and reliability across standardized models.
VOLLO ran on an AMD Versal Premium series chip installed in a Supermicro server with a Silicom accelerator card. The combination delivered lower latency than all previously audited systems across all three benchmark models.
Why This Matters for Trading
Microsecond-level latency translates directly to speed in trading decisions. Firms using VOLLO can run more complex models faster when analyzing market data, evaluating risk, or generating quotes. That speed advantage compounds when multiplied across thousands of daily decisions.
Myrtle.ai said hundreds of thousands of hours of production trading already run on VOLLO across major trading firms. Traders develop models in standard machine learning tools, then compile them into VOLLO for deployment on FPGA hardware.
The Hardware Stack
VOLLO doesn't require FPGA expertise to deploy. Developers can test their models' performance on the platform without learning specialized FPGA tools, according to the company.
Peter Baldwin, CEO of myrtle.ai, said the company has spent three years expanding which models VOLLO supports and which hardware platforms it runs on since first setting records in this benchmark in 2023.
AMD's Versal Premium series chip contains more than 3.3 million programmable logic units and PCIe Gen5 connectivity, features suited to applications requiring deterministic, low-latency inference.
The full benchmark results are available in STAC Report SUT ID MRTL260323. Developers can request an evaluation at vollo.myrtle.ai or contact fintech@myrtle.ai.
For more on AI for Finance or how financial leaders approach AI infrastructure, see the AI Learning Path for CFOs.
Your membership also unlocks:
