Purpose-built acceleration where it matters most
Powered by MPPA® architecture, Kalray's Data Processing Unit technologies deliver performance, efficiency, and scalability where it matters most, enabling smarter AI, accelerated storage and networking, and more secure infrastructure from AI factories and HPC to cloud and edge.
Across workloads
Six categories. One architecture
Throughput without the CPU tax
Data movement, encryption, and I/O processing are major bottlenecks in modern storage infrastructures. DPUs bring parallelism and dedicated processing to eliminate those bottlenecks, especially for AI, HPC, and cloud environments.
- Achieve deterministic, terabit throughputs without burning CPU cores
- Seamlessly offload virtual networking to dedicated hardware
- Reduce latency for AI and HPC fabrics
Storage Acceleration Patterns
- NVMe-oF, storage disaggregation and virtualization
- Data encryption and compression
- Data transfers between CPU, GPU, storage
- High performance RDMA
Free your CPUs from the data path
Packet processing and traffic flow management are placing increasing strain on CPU resources when using traditional Network Interface Cards. DPUs overcome these limitations by offloading and accelerating packet and flow processing directly on the DPU, freeing CPU cycles while delivering higher throughput, lower latency, and more efficient networking across AI and HPC fabrics, as well as cloud environments.
Network Acceleration Patterns
- High-performance virtual switching
- AI and HPC fabric optimization
- Packet processing acceleration
- Cloud native networking
Hardware-rooted isolation at wire speed
Fine-grained security processing, encryption, and isolation are becoming essential in modern public and multi-tenant infrastructures. DPUs offload critical security functions, encryption, policy enforcement, segmentation, inspection, and threat detection, onto dedicated hardware, delivering strong isolation and enabling zero-trust security at wire speed.
- Security with zero CPU tax · physical hardware separation
Security Acceleration Patterns
- Infrastructure services protection from host workloads
- Secure multi-tenant segmentation and isolation
- Hardware based policy enforcement
Keep GPUs fed, utilized, focused
As AI workloads grow more distributed and LLM architectures get more complex, network and data movements are now a first-class performance bottleneck. DPUs are becoming a foundational building block for modern AI infrastructure by offloading critical tasks such as RDMA communication, storage-to-GPU data paths, and KV cache transfers, keeping GPUs fed, utilized, and focused entirely on inference and training.
- Increase AI infrastructure performance and efficiency by removing current bottlenecks
- Accelerate and optimize data transfers
- Enable scalable distributed AI inferencing
Where DPUs Change AI Infrastructure Economics
- AI data movements optimization
- AI fabrics acceleration
- KV Cache Transfer for Prefill / Decode Disaggregation
- Secure multi-tenant AI infrastructure
Wire-speed processing for the cloud-native telco edge
5G and telco edge clouds demand wire-speed packet processing, ultra-low latency, and support for dense, virtualized network functions. DPUs offload and accelerate critical workloads, vRAN, UPF, directly in silicon, freeing CPUs for higher-value tasks. They deliver cloud-native efficiency with predictable performance for the most demanding 5G and edge environments.
5G & Telco Edge Patterns
- vRAN / Open RAN acceleration
- UPF packet processing offload
- Network slicing enforcement
- Edge security, encryption, and isolation
Industries with specific needs build their own
From aerospace to deep tech, not every workload fits a predefined category. DPUs can be custom-designed for virtually any compute-intensive or latency-sensitive task. Whether you're building specialized pipelines, protocol stacks, or unique AI flows, DPUs can be architected to give you full control over performance, power, and flexibility.
- Architecture tailored to customer-defined workloads
- Flexible software-defined pipelines and deterministic scheduling
- Suitable for industries with specific needs: aerospace, defense, fintech
Build your own
-
Industry-specific hardware acceleration
When standard cards don't fit, Kalray's IPs and design services bring custom DPU programs from idea to silicon.
From use case to deployed acceleration
Whichever way you want to engage, cards, IPs, services, talk to Kalray about your acceleration needs.