How ApexQuant Achieved Sub-5 ms Latency for Crypto Trading

About the client

ApexQuant is a quantitative crypto trading company operating automated strategies across several digital asset markets.

The company’s platform continuously receives order book updates, calculates trading signals, manages risk, and submits orders through exchange APIs.

Unlike longer-term trading systems, several of ApexQuant’s strategies are highly sensitive to network latency. Even a few additional milliseconds between receiving a market update and submitting an order can affect execution quality.

Challenge

ApexQuant initially operated its trading infrastructure from general-purpose cloud environments.

Compute performance was sufficient, but network paths to several of the company’s primary exchange gateways were inconsistent.

Depending on routing and network conditions, latency to critical trading endpoints could fluctuate between 8 and 20+ ms, with occasional spikes significantly above that range.

For latency-sensitive strategies, average ping was not the only concern. The team needed both low latency and predictable latency.

The key requirements were:

• below 5 ms RTT to primary exchange endpoints;
• low jitter and stable network routing;
• strong single-thread CPU performance;
• dedicated physical resources;
• NVMe storage for high-volume market data;
• sufficient network capacity for multiple real-time data feeds.

Solution

Together with Unihost, ApexQuant selected a dedicated infrastructure location based on actual network tests to the exchange gateways used by its trading platform.

Instead of choosing the server purely by country or advertised bandwidth, the team tested routing from available Unihost locations and deployed the execution environment where the best network path was available.

The company selected an AMD Ryzen 9 9950X dedicated server for its latency-sensitive trading stack.

The 16-core Ryzen platform offers stable frequencies of up to 4.3 GHz, making it particularly suitable for workloads where fast execution of individual processes is more important than simply maximizing the number of CPU cores. Unihost lists high-frequency trading among the target workloads for this platform.

The server runs:

• real-time market-data collectors;
• order book processing;
• trading strategy engines;
• risk-management services;
• Redis;
• exchange API connectors;
• order execution services.

Research, backtesting, historical datasets, and other non-latency-sensitive workloads remain on separate infrastructure.

Unihost solution used

Ryzen 9 9950X Dedicated Server

AMD Ryzen 9 9950X, 16 cores / 32 threads, up to 4.3 GHz, DDR5 RAM, NVMe storage, up to 10 Gbps connectivity.

Used for: market-data ingestion, strategy calculation, risk checks, exchange API connectivity, and low-latency order execution.

The dedicated hardware eliminates resource contention with neighboring virtual machines, while the high-frequency CPU helps minimize processing time between receiving a market event and producing an order.

For ApexQuant, however, the most important part of the deployment was the network path: the final location was selected only after testing latency to the customer’s actual exchange infrastructure.

Results

After moving its execution stack to the selected Unihost location, ApexQuant measured 1.8 – 4.6 ms round-trip latency to its primary exchange endpoints under normal network conditions.

Previously, the same connections typically ranged between 8 and 20+ ms.

The improvement was especially noticeable in latency-sensitive strategies that process large numbers of order book updates and continuously adjust open orders.

The new architecture also reduced jitter. Instead of optimizing for the best occasional response time, ApexQuant gained a much tighter and more predictable latency range.

“For our execution infrastructure, 20 ms versus 5 ms is not a small difference. We selected the server based on the actual route to the exchange, and after deployment our primary connections consistently stayed below 5 ms.”

Architecture

Market data feeds

↓

Dedicated Ryzen execution node

↓

Order book processing

↓

Strategy + risk engine

↓

Exchange gateway – 1.8 – 4.6 ms RTT

Non-critical workloads such as research, backtesting, analytics, and long-term data storage remain outside the low-latency execution environment.

This separation keeps the trading node focused on one task: processing market events and transmitting orders as quickly and consistently as possible.

What’s next

ApexQuant plans to deploy a second execution node for redundancy and use active-passive failover between the two systems.

The company is also testing additional Unihost locations against other exchange gateways, with the goal of maintaining sub-5 ms connectivity for a larger percentage of its trading infrastructure.

Results at a glance: 1.8 – 4.6 ms RTT, sub-5 ms exchange connectivity, lower jitter, dedicated high-frequency CPU, isolated execution environment.