Phemex Cuts Latency by Over 90%
Crypto exchange Phemex has upgraded its trading engine, dramatically reducing funding settlement latency from 10 seconds down to about 500 milliseconds. The more than 90% improvement is aimed at enabling smoother, uninterrupted trading, highlighting the ongoing performance arms race among digital asset venues.
The Phemex upgrade is part of a broader architectural shift, increasing its futures trading system's overall throughput by 60% to handle approximately 40,000 transactions per second (TPS). This enhancement is designed to maintain fast response times and system stability, particularly during peak trading and high market volatility. The new decentralized multi-node architecture also aims to eliminate single points of failure. Beyond latency reduction, the optimization has significantly improved system resource efficiency. Daily CPU usage for the trading engine has been reduced by over 50%, while the risk engine's memory consumption has decreased by about 30%. These gains in efficiency allow for greater processing capacity under the same hardware conditions, reducing the likelihood of order queuing during concurrent high-volume activity. The race for low latency in crypto often involves a choice between on-premises infrastructure for maximum control and speed, and cloud deployments for scalability and global reach. While co-location within the same data center as an exchange's matching engine can achieve single-digit microsecond latency, many crypto exchanges utilize a hybrid or optimized cloud approach across various public cloud regions. This requires traders to focus on selecting compute-optimized instances and specific availability zones to minimize network hops. To achieve sub-millisecond performance, advanced trading systems are increasingly turning to kernel bypass techniques and FPGAs (Field-Programmable Gate Arrays). Kernel bypass libraries like DPDK or OpenOnload allow applications to interact directly with network card buffers, avoiding the inherent latency of the operating system's network stack. FPGAs take this further by implementing trading logic directly in hardware, enabling deterministic, nanosecond-level latency by executing tasks in parallel without the context-switching delays of CPUs. This infrastructure overhaul aligns with a strategic pivot announced by Phemex CEO Federico Variola, focusing on becoming an "AI-native" platform. The goal is to embed artificial intelligence into the core operational and product development layers, moving beyond simply adding AI features. This suggests future developments will leverage the enhanced engine performance for more adaptive and intelligent trading tools.
