Agentic AI
Multi-step agentworkloads
Agents accumulate working state across multi-turn conversations and tool calls. That state stays warm between calls without being evicted to storage. X-Mem™ holds the working set at memory-tier latency.
Memory Systems
Purpose-Built for AI
X-Mem™ Technology — Co-Designed from Silicon to Token
The world's first memory processing unit (MPU) chip that enables AI infrastructure with composable memory-to-compute ratios.
Animated hero concept: a unit of two MPU™ memory nodes and two XPU compute nodes meshed through a scale-up fabric, with data streaming between them; the view then pulls back to show the same unit tiling outward across the rack.
What would you do differently if you had orders of magnitude more memory?
- XPU compute is starved for memory.
X-Mem™: the Warm Memory Tier™
AI workloads generates a lot of warm data that needs to be kept resident at memory speed. A new tier of memory is needed.
HBM is too small to hold an agent’s working set. Storage is too slow to serve it. And the working set is not one session’s problem: it grows with every concurrent agent and user, and it outlives the sessions that created it. The working set is a shared resource on the rack.
Netpreme builds the Warm Memory Tier™ that keeps it resident at memory speed.
Faster · Higher BW
G1
HBM
On-Package · Active Data
PCIe
Scale-up Fabric
G2
System DRAM
CPU-Dedicated
G1.5 & G2.Exp
Warm Memory Tier™
GPU-Dedicated · Unified Memory · Load/Store
Big Latency Cliff
Scale-out Network
G3
Flash / NVMe-oF
Block I/O · Cold Data
G4
Elastic Storage
Cross-Rack
More Capacity
Engineered forAI workloads
Join Our Team
We’re hiring across Silicon, Systems, and ML with offices in Santa Clara, CA and Cambridge, MA.