Research & Publications

xPU.IO projects are rooted in rigorous academic research. Our work has been published at top-tier conferences and contributes to the advancement of GPU-centric storage systems for AI workloads.

Preprint

Tutti: GPU-Centric KV Cache Storage for Long-Context LLM Inference

(Authors TBD)
arXiv preprint β€’ 2026

Abstract

Long-context AI agents require storing and retrieving massive KV caches. Tutti proposes a GPU-centric multi-tier caching architecture that leverages NVMe SSDs as the primary storage tier, achieving 256K+ context support with intelligent prefix caching.

Core Contributions

  • Multi-tier KV cache architecture (HBM/DRAM/NVMe)
  • Prefix caching optimization for Agent workloads
  • 94.2% cache hit rate on SWE-bench Pro traces
  • 131:1 input/output token ratio utilization

Key Results

  • Supports 256K+ token contexts
  • 94.2% cache hit rate in Agent applications
  • Efficient storage of massive conversation histories
Read PaperView CodeProject Page
Published

Phoenix: A Refactored I/O Stack for GPU Direct Storage without Phony Buffers

Jianqin Yan, Shi Qiu, et al.
SC'25 - International Conference for High Performance Computing, Networking, Storage, and Analysis β€’ 2025

Abstract

Traditional GPU-Storage data paths require data to traverse CPU and host memory, acting as "phony buffers" that waste bandwidth and CPU cycles. Phoenix refactors the I/O stack by leveraging Linux ZONE_DEVICE to make GPU memory a legal I/O buffer, enabling direct DMA from NVMe to GPU without application or filesystem modifications.

Core Contributions

  • Eliminates phony buffers in GPU-Storage path
  • Achieves 24 GB/s throughput (96%+ bandwidth utilization)
  • Supports multiple GPU vendors without kernel I/O changes
  • Integrates with io_uring for high-concurrency I/O
  • Maintains POSIX interface compatibility

Key Results

  • 2.2Γ— faster than CPU-mediated path
  • 5.6Γ— TTFT improvement in LLM inference (with LMCache)
  • Zero application code changes required
Read PaperView CodeProject Page

Our Research Focus

GPU Direct Storage

Eliminating CPU bottlenecks in GPU-Storage data paths through direct DMA transfers and ZONE_DEVICE mechanisms.

KV Cache Management

Efficient storage and retrieval of KV caches for long-context LLM inference with multi-tier architectures.

AI Infrastructure

System-level optimizations for AI workloads focusing on performance, scalability, and vendor neutrality.