NXAI's xLSTM Powered by AMD Instinct GPUs: European-Developed AI Architecture that Outperforms Transformers for Industrial Applications

Author: NXAI

_{The information contained in this blog represents the opinion of NXAI as of the date presented. AMD and/or NXAI have no obligation to update any forward-looking content in the blog. AMD is not responsible for the content of this third-party blog and does not necessarily endorse the comments/claims/information presented therein.} 

Revolutionary xLSTM technology offers an efficient alternative to transformers for edge AI applications and suggests a path to increased European technological sovereignty.

NXAI has successfully demonstrated that their groundbreaking xLSTM (Long Short Term Memory) architecture achieves exceptional performance on AMD Instinct™ GPUs, marking a significant advancement in recurrent neural network technology for edge computing applications.

The collaboration with AMD provides the computing foundation needed to maximize xLSTM's potential for industrial applications. As edge AI deployment—computing that happens directly on devices rather than in the cloud—accelerates across sectors, the combination of hardware expertise from AMD and NXAI's algorithmic innovation offers compelling advantages over transformer-based approaches that typically require substantial computing resources.

Good Performance for Edge Computing

NXAI reveals impressive results with xLSTM for some specific tasks when compared to transformer models and state space models in language modeling tasks. More importantly, xLSTM delivers these results with significantly lower computational requirements—a critical factor for edge implementation. In benchmark tests, xLSTM achieved comparable accuracy to leading transformer models.

Key benefits of the xLSTM architecture include:

• xLSTM can efficiently handle time series data, vision inputs, or be used as a Bio-xLSTM for molecular biology with unlimited context or as a large action model in robotics.
• xLSTM is energy efficient during both training and operation.
• The architecture's efficiency enables implementation on edge devices.
• xLSTM demonstrates fast performance in both inference and training phases. 

Optimizing xLSTM for AMD ROCm™ Open Software Ecosystem

As part of NXAI's ongoing collaboration with AMD, xLSTM was optimized for execution on AMD Instinct™ MI300X GPUs using PyTorch's TunableOp interface. After recording and tuning the core GEMM kernels, xLSTM achieved up to a 2.16× improvement in generation throughput compared to previously reported performance numbers on NVIDIA's H100 GPU.

These results demonstrate not only the efficiency of xLSTM's architecture, but also its compatibility with the AMD ROCm™ open software stack and MI300X GPUs. This positions xLSTM as a compelling solution for high-performance inference on AMD platforms, with implications for data center deployments.

Strengthening European Technological Sovereignty

This breakthrough suggests a clear roadmap for implementing AI across diverse industrial settings while bolstering Europe's technological independence. By enabling advanced AI capabilities on edge devices, this technology helps reduce dependency on external technology providers.

The technology is particularly significant for manufacturing, automotive, and healthcare sectors, where processing sensitive data locally rather than in cloud services offers both performance and regulatory advantages.

Next-Generation Neural Network Design

Developed by Prof. Dr. Sepp Hochreiter, the renowned creator of the original LSTM technology that powers popular applications like Siri, xLSTM represents the next evolution in neural network design. This innovative architecture overcomes traditional LSTM limitations through Exponential Gating with specialized normalization and stabilization techniques, paired with a new Matrix Memory system.

NXAI is already collaborating with international industrial partners on specific time series models and using the Vision-xLSTM approach in drones.  

"This breakthrough represents more than just technical innovation— it's a strategic advancement for European industrial competitiveness.  The ability to run sophisticated AI on existing industrial edge systems without requiring specialized hardware creates unprecedented opportunities for automation and intelligence at the edge." — Albert Ortig.
 

NXAI:  Founded in 2024 with foundational research at core, NXAI is focused on developing cutting-edge technologies that bridge the gap between academia and market-ready innovations. The unique approach to AI industrialization is led by a proven team of experts, including AI pioneer Sepp Hochreiter, who brings his visionary leadership and research expertise to the forefront of AI's next wave.

AMD Silo AI:  Pushes the Frontier of AI on AMD Compute Platforms.  AMD Silo AI is a leading AI lab that helps customers develop and deploy advanced AI models and solutions optimized for leadership compute platforms.  As a global AI center of excellence with over 300 AI scientists and 125+ PhDs, they combine deep scientific knowledge with practical technology understanding to help organizations integrate, deploy and scale AI effectively. Our work spans from groundbreaking research to enterprise-ready AI solutions.