iWave Partner

Technology Partner iWave

Xelera and iWave:
Ultra-low latency AI inference for HFT on iWave XFibre SmartNICs

Deploy Gradient-Boosted Trees and neural networks on iWave G30, G35 and G63 PCIe Accelerator Cards, and Neural Networks on the Versal™-based G63. Two deployment modes, one trained model file, one API.

~250 ns

Silva inline GBT inference, p99
(reference benchmark, LightGBM)*

~ 350 ns

iWave tick-to-trade framework,
end to end, excluding trading logic

1 API

C, C++, C# and Python,
identical across both modes

Silva, now on iWave hardware

iWave XFibre SmartNICs combine 10G to 400G Ethernet with AMD adaptive SoCsand the iWave HFT framework: a full hardware tick-to-trade pipeline.

Xelera Silva adds the decision layer. Train in XGBoost, LightGBM, CatBoost or PyTorch, then run the model as a turnkey accelerator, or as a Silva IP core inside the iWave HFT framework. Switching modes is a configuration change, not a rewrite.

Supported cards

10/40 GbE

G30-XFibre
- 10/40G®

Zynq™ UltraScale+™ SmartNIC

FPGA: AMD Zynq™ UltraScale
+™ZU7CG MPSoC
Network: 4× 10GbE SFP+ (1× / 4×config)
Host: PCIe Gen3 x8, FHHL
AI Models: Gradient-Boosted Trees
Best for: Single-venue 10G feedhandlers, compact treeensembles
GBT
Neural Networks

100 GbE

G35-XFibre-
100G®

Zynq™ UltraScale+™ SmartNIC

FPGA: AMD Zynq™ UltraScale
+™ZU11/ZU19EG MPSoC
Network: 2× 100GbE, NVMe expansion
Host: PCIe Gen3 x16, FH3/4L
AI Models: Gradient-Boosted Trees
Best for: Multi-feed 100G, larger treeensembles, concurrentstrategies
GBT
Neural Networks

400 GbE

G63-XFibre-
400G®

Versal™ Premium SmartNIC

FPGA: AMD Versal™ Premium
(VP1102 to VP1552 family)
Network: 400G, 112G PAM4 transceivers
Host: PCIe Gen5 x8, FH3/4L
AI Models: Gradient-Boosted Trees andNeural Networks
Best for: 400G aggregation, LSTM andattention models, large multistrategy books
GBT
Neural Networks

Deployment modes

Two ways to run Silva on XFibre

From fully turnkey to fully in-hardware: choose how much of the stack Xelera delivers and how much comes from iWave. The trained model file is identical in both.

Silva Accelerator: turnkey offload

The Silva IP core replaces the AI inference stage of iWave's inline tick-to-trade framework. Market data isdecoded, scored and turned into an order without leaving the chip. iWave provides the framework and the PCIeAccelerator Card.

Data path: Host to card over PCIe
Xelera delivers: Bitstream, driver,runtime, API
iWave delivers: PCIe Accelerator Card
‍

Silva IP in the iWave tick-to-trade framework

The Silva IP core replaces the AI inference stage of iWave's inline tick-to-trade framework. Market data is decoded, scored and turned into an order without leaving the chip. iWave provides the framework and the PCIeAccelerator Card.

Data path: on-chip, No PCIe
Xelera delivers: Silva IP Core + model compiler
iWave delivers: Tick-to-tradeframework, PCIeAccelerator Card
‍

Modes at a glance

Bring your own model.

Train with standard frameworks on your data. Deploy the same model file in any mode, hot-swap without stopping the card.

Gradient Boosted Trees

XGBoost, LightGBM, CatBoost. Numerical float32
and categorical features

G30 - G35- G63

Neural Networks

LSTM, MLP, Conv1D, single-head attention in float16 / bfloat16. State preserved across calls.

G30 - G35 - G63

Getting Started

Evaluate on CPU

Run your model in Silva CPU-only mode on any Linux box. Zero hardware leadtime.

Pick card and mode

Match bandwidth (10G, 100G, 400G),model class and integration depth with Xelera engineers.

Integrate

Install the turnkey bitstream and API, or drop the Silva IP into the iWave HFT framework with the example design.

Go live

Lifecycle support, periodic updates and model hot-swap in production.

News

Learn more about currents trends in acceleration tech

May 26, 2026

Xelera Technologies at STAC Summits: Advancing Low-Latency AI for Financial Markets

Xelera Technologies joined the STAC community in Hong Kong, London, and New York to explore the future of low-latency AI in financial markets. Discover how Xelera Silva’s flexible execution modes and high-performance inference technology help bring machine learning closer to real-time trading.

March 6, 2026

Reflections from Omnisecure 2026: The Future of Smart ID Solutions

Discussions at Omnisecure 2026 highlighted three priorities for the industry: digital identity as infrastructure, security-by-design for scalable systems, and sovereign technologies to protect sensitive data in modern digital environments.

November 24, 2025

Xelera Technologies Engages HPC and AI Leaders at Supercomputing 2025

At Supercomputing 2025, Xelera Technologies highlighted how Xelera Silva and Xelera Net deliver deterministic, high-performance acceleration for HPC, AI, and low-latency workloads while connecting with global experts on real-world performance challenges.