# XCENA

**Source:** https://geo.sig.ai/brands/xcena  
**Vertical:** Artificial Intelligence  
**Subcategory:** CXL Computational Memory  
**Tier:** Emerging  
**Website:** xcena.com  
**Last Updated:** 2026-04-14

## Summary

MX1 launched with PCIe 6.0 + CXL 3.2. MX1S ships 2026. FMS 2025 Best of Show. Thousands of RISC-V cores inside memory for near-data AI inference — solves KV cache bottleneck.

## Company Overview

XCENA (formerly MetisX) is developing computational memory products that place thousands of RISC-V processing cores inside a CXL-attached memory module — enabling AI inference to run where the data lives rather than requiring massive data transfers to a GPU. The MX1 computational memory product launched with PCIe 6.0 and CXL 3.2 interfaces and won Best of Show at FMS 2025, with the improved MX1S targeting a 2026 production launch.

The core problem XCENA addresses is the KV (key-value) cache bottleneck in long-context LLM inference: as language models process longer prompts and generate longer outputs, the KV cache — the memory structure storing attention computation results — grows rapidly and requires constant high-bandwidth memory transfers between the GPU's memory and compute units. XCENA's near-memory compute allows KV cache operations to execute where the cache data lives, eliminating the memory bandwidth bottleneck that limits long-context inference speed.

The CXL 3.2 interface allows XCENA's computational memory to attach to standard servers alongside existing GPU and CPU configurations, enabling enterprises to add near-memory compute capacity to deployed infrastructure without replacing servers. This incremental deployment model contrasts with alternative approaches requiring new server designs and could accelerate adoption among enterprises with existing AI infrastructure investments.

## Frequently Asked Questions

### What does XCENA make?
CXL computational memory — thousands of RISC-V cores inside a memory module for near-data AI inference, eliminating the KV cache bandwidth bottleneck for long-context LLM inference.

### What products has XCENA launched?
MX1 (PCIe 6.0 + CXL 3.2, FMS 2025 Best of Show). MX1S with dual PCIe Gen6 targeting 2026 production launch.

### What is the KV cache bottleneck?
Long-context LLM inference requires constant high-bandwidth transfers of growing KV cache data between GPU memory and compute. Near-memory RISC-V cores in XCENA's module eliminate this transfer bottleneck.

### How does XCENA deploy?
CXL 3.2 interface attaches to standard servers alongside existing GPUs — enterprises add near-memory compute capacity without replacing deployed AI infrastructure.

### What is computational memory and how does XCENA's approach work?
Computational memory places processing logic directly within memory arrays rather than moving data to separate compute units. For LLM KV cache operations — reading and writing attention state during inference — this eliminates the bandwidth bottleneck between CPU/GPU and DRAM, processing attention computations where the data lives. XCENA's CXL-based computational memory expander brings this capability to standard server platforms via the CXL interconnect standard.

### What performance improvements does XCENA deliver for LLM inference?
XCENA targets LLM inference throughput limitations caused by KV cache memory bandwidth — for large models running long context windows, KV cache reads become the bottleneck that limits tokens/second per GPU. XCENA's computational memory reduces this bottleneck by processing KV cache operations locally, potentially doubling inference throughput on memory-bound LLM workloads without changing the GPU or model weights.

### How does XCENA integrate with existing AI server infrastructure?
XCENA deploys as a CXL-attached memory expander — a PCIe card or EDSFF (Enterprise and Datacenter SSD Form Factor) module that plugs into standard server expansion slots and appears as a CXL memory device to the host system. Existing inference software (vLLM, TensorRT-LLM) requires minimal modification to route KV cache allocations to XCENA-managed memory regions, enabling incremental integration without system redesign.

### What is XCENA's competitive landscape in CXL memory?
XCENA competes with Samsung's HBM-PIM (processing-in-memory HBM), SK Hynix's AiM DRAM, Upmem (DRAM with processing elements), and MemVerge (CXL memory management software). The key distinction is XCENA's CXL form factor compatibility with standard servers (no specialized memory bus required) and focus on LLM inference KV cache as the specific optimization target, versus more general-purpose near-memory processing approaches.

## Tags

ai-powered, b2b, saas

---
*Data from geo.sig.ai Brand Intelligence Database. Updated 2026-04-14.*