OpenNUMA presents prototype for shared memory management for CPU and integrated GPU

Summary

OpenNUMA is a prototype custom memory allocator that enables the integrated GPU (iGPU) to share system memory for neural network workloads via PyTorch's XPU backend. While the research reports an approximately 34% reduction in transfer time from CPU to iGPU, it observes only marginal improvements (or even slight regressions) in overall computational performance.

Why it matters

Most machine learning deployments assume discrete GPUs with tensor cores. This leaves the iGPU computational capacity in mainstream laptops and older desktop systems underutilized. A functional shared memory allocator could activate this dormant capacity without requiring hardware upgrades, which is significant for cost-sensitive and legacy environments. However, this contribution is a proof-of-concept sketch rather than a production-ready system, meaning its practical impact remains speculative.

Highlights

• A 34.076% reduction in transfer time from CPU to iGPU, but a 5.258% increase in the reverse direction
• Overall workload performance is 'nearly identical' to the default allocator, failing to demonstrate meaningful computational speedup
• Fundamental shared tensor access between CPU and iGPU could not be verified through the existing PyTorch interface

Critical take

Although the paper claims platform independence, it is hard-dependent on PyTorch XPU, Intel SYCL, and Windows. Furthermore, the primary metric is merely the difference in transfer time, not data throughput or latency gains in actual model inference. More concerning is that the one feature that would truly make the allocator functional—verified shared tensor access between CPU and iGPU—has not been explicitly validated. This presents a prototype that reconfigures memory through an additional layer without proving it changes outcomes.