Blog

One SmartNIC Foundation for AI Data Movement and 400G Network Security

As AI factories scale, networking becomes a critical part of the AI data path, demanding higher throughput, lower latency, and minimal CPU involvement. This blog explores how Xelera’s 400G SmartNIC platform enables efficient network-to-accelerator data movement and high-speed IPsec offload, providing a common infrastructure for both AI workloads and secure datacenter connectivity.

Category:

One SmartNIC Foundation for AI Data Movement and 400G Network Security

Three Execution Modes, One API: How Xelera Silva Matches ML Inference to Your Latency Constraint

Xelera Silva covers three execution modes under a single API: CPU-Only for cache-fit models, Accelerator for larger GBT and LSTM topologies, and Inline Silicon for hard latency determinism. Bring your trained model; Silva handles the execution.

Xelera Sets the First STAC-ML El Popo World Record

Xelera achieves the first STAC-ML El Popo world record: Silva scores GBT inference at 1.60μs p99 (GBT_A) with AMD Alveo V80 on HPE ProLiant DL385. Independently audited. Configuration XLRA260312 now available in the STAC Vault.

Where Does the Time Go? A Systems View of GBT Inference Latency

Model-level optimisation has a ceiling set by the compute platform. Well-optimised CPU inference is genuinely fast for small models, but it carries two structural weaknesses: maximum-latency spikes that grow by orders of magnitude under concurrent load, and a determinism ceiling set by interrupts, scheduling, and shared resources. Offload accelerators (PCIe-attached) trade raw speed for consistency. Inline accelerators (network-attached FPGAs) remove data movement from the hot path entirely, delivering sub-microsecond latency with near-zero jitter. Many production deployments use all three.

Gradient Boosted Trees: Two Types of Numbers, Two Types of Failure

Quantising a GBT model is not the same as quantising a neural network. A neural network weight rounded slightly produces a small, continuous error. A GBT threshold rounded slightly can send an input down the wrong branch, producing a catastrophic error, greater than 100 % on real benchmark data. The correct approach is to quantise the input data before training and compute leaf bit-widths analytically from the trained ensemble.

Gradient Boosted Tree Inference is not what you think it is

GBT inference speed is not limited by arithmetic; it is limited by memory access. Two structural choices made at training time, tree depth and ensemble-to-core mapping, determine whether latency is bounded or openended. These are not tuning knobs; they are architectural decisions that should be set before deployment, not discovered in production.

Introducing Xelera Silva CPU-Only: Sub-Microsecond GBT Inference on Any Machine

In high-frequency trading, microseconds matter—but consistency matters more, since unpredictable latency can turn even fast systems into unreliable ones with real financial risk. Xelera Silva addresses this by delivering deterministic, microsecond-level Gradient Boosting Tree inference without requiring specialized hardware expertise, now expanded with a new deployment option backed by benchmark data.

IPsec Performance at 400GbE: Eliminating CPU Bottlenecks with SmartNIC Offload Using Xelera Net

At 100GbE network speeds, CPU-based IPsec processing becomes a performance bottleneck, consuming compute resources and reducing application performance. Xelera Net addresses this by offloading IPsec encryption and decryption to SmartNIC hardware, enabling up to 2×100 Gbps throughput with low latency while freeing server CPUs for application workloads. 🚀

Xelera Silva: Redefining Ultra-Low Latency AI Inference

Xelera Silva is an ultra-low latency FPGA inference accelerator designed for time-critical systems like HFT and in-path network analytics, delivering deterministic microsecond to sub-microsecond ML inference without requiring model rewrites. The blog outlines how Silva’s offload and inline deployment options combine speed, low jitter, and ease of integration to enable real-time, production-ready decision making where every microsecond matters.

Xelera: Building an Ecosystem of SmartNIC Accelerated Applications

Xelera AP is a high-performance software platform that unlocks the full potential of SmartNICs and DPUs by enabling accelerated network functionality, security, and AI inference. Built on top-tier hardware, it helps data centers boost speed, reduce latency, and offload critical tasks efficiently.

Machine Learning Inference for HFT: How Xelera Silva and ICC Deliver Ultra-Low Latency Trading Decisions

Traditional CPU-based ML inference can’t meet the ultra-low latency demands of high-frequency trading, but Xelera Silva on ICC VEGA servers delivers consistent sub-2µs performance—over 10x faster and more reliable for real-time decision-making.

Napatech SmartNICs and Xelera Silva Software Accelerate AI Inference for High-Frequency Trading

Xelera Silva and Napatech SmartNICs redefine speed and precision in AI-driven trading! Discover how this powerful pairing achieves inference speeds up to 30x faster than traditional CPU-optimized solutions — even at the 99th percentile.

Recap 2024

We are thrilled to share the highlights and milestones that have defined an extraordinary year for Xelera Technologies. From innovations to strategic partnerships, this year has been transformative. Join us as we look back on the moments that defined 2024 and set the stage for an even brighter future. Thank you for being part of our story. Your support has been crucial to our success, Here's to another year of excellence and innovation! ✨

Recap 2023

As 2023 comes to a close, it's time to reflect on a year filled with remarkable achievements and groundbreaking developments at Xelera Technologies. Join us in this recap while we explore the major milestones, technical advancements and team and community engagement that defined Xelera's dynamic journey throughout the year.

Ultra-low Latency XGBoost with Xelera Silva

In the era of data-driven decisions, every microsecond counts. In this article "Ultra-low Latency XGBoost with Xelera Silva", we discuss the optimization of xgboost, lightgbm and catboost for lightning-fast machine learning inference.

Low-latency Machine Learning Inference for High-Frequency Trading

As electronic trading strategies that support machine learning proliferate, the speed at which machine learning algorithms can make decisions is once again becoming one of the critical factors in differentiating oneself from the competition.

High-Frequency Trading with Machine Learning Algorithms

As ML-enabled trading strategies proliferate, companies must find ways to differentiate themselves. One trend we've seen over the past two years is trading companies moving machine learning into the hot path of the ultra-low latency trading cycle - a task that remains challenging as most machine learning does not operate at microsecond latency. Let's explore the transformative power of AI in trading together!

What is a SmartNIC?

Since phones became smartphones, we've seen more and more smart devices. In the data center, a similar evolution is taking place as network interface cards become SmartNICs. Cloud architects, builders and operators see them as an opportunity to build better data centers. Analysts predict the SmartNIC market will become a $2B market by 2027. So, what is a SmartNIC? In this article, we dive into this technology, explore its features and shed light on its potential.

Datacenter Networks

This week, the SmartNIC Summi takes place in San Jose, California. It brings together industry experts, users and vendors of SmartNICs, Data Processing Units (DPUs) and Infrastructure Processing Units (IPUs). The common thread among these technologies is that they allow us to move data processing tasks into the NIC in addition to basic network functions. But why are network developers doing this in the first place? And why are they doing it now?

Accelerating Decision Tree-Based Predictive Analytics

Gradient boosting frameworks such as XGBoost, LightGBM and CatBoost, as well as Random Forest algorithms are often a part of winning machine learning models in Kaggle competitions...

Recognizing Voices With AI

Voice-based digital assistants are on the rise. Xelera Technologies provides an AI module for speech processing systems that distinguishes voices and gives...

Do you want your data center to be greener?

The European Union wants data centres to be greener. Following the EU publication “Shaping Europe’s Digital Future”, Telecommunications and data centres have a significant environmental footprint...

One SmartNIC Foundation for AI Data Movement and 400G Network Security

One SmartNIC Foundation for AI Data Movement and 400G Network Security

Three Execution Modes, One API: How Xelera Silva Matches ML Inference to Your Latency Constraint

Xelera Sets the First STAC-ML El Popo World Record

Where Does the Time Go? A Systems View of GBT Inference Latency

Gradient Boosted Trees: Two Types of Numbers, Two Types of Failure

Gradient Boosted Tree Inference is not what you think it is

Introducing Xelera Silva CPU-Only: Sub-Microsecond GBT Inference on Any Machine

IPsec Performance at 400GbE: Eliminating CPU Bottlenecks with SmartNIC Offload Using Xelera Net

Xelera Silva: Redefining Ultra-Low Latency AI Inference

Xelera: Building an Ecosystem of SmartNIC Accelerated Applications

Machine Learning Inference for HFT: How Xelera Silva and ICC Deliver Ultra-Low Latency Trading Decisions

Napatech SmartNICs and Xelera Silva Software Accelerate AI Inference for High-Frequency Trading

Recap 2024

Recap 2023

Ultra-low Latency XGBoost with Xelera Silva

Low-latency Machine Learning Inference for High-Frequency Trading

High-Frequency Trading with Machine Learning Algorithms

What is a SmartNIC?

Datacenter Networks

Accelerating Decision Tree-Based Predictive Analytics

Recognizing Voices With AI

Do you want your data center to be greener?

Products

Solutions

Company

Resources