Modern hardware accelerators like GPUs offer a massive computational advantage over CPUs, but they are frequently hamstrung by a "memory wall." While a typical server might have terabytes of system RAM, a high-end GPU often has less than 100GB of High Bandwidth Memory (HBM). When datasets reach the billion-scale or exabyte-scale, they simply cannot fit on the device.
To bridge this gap, researchers have developed a suite of "out-of-core" (OOC) and hybrid processing strategies. The goal is to maintain the throughput of the GPU while treating its limited VRAM not as a final destination for data, but as a high-speed cache for a much larger storage tier.
The Partitioning Strategy
The most fundamental approach to overcoming memory limits is data partitioning. If a dataset is too large to be processed at once, it must be broken into chunks that fit within the available VRAM.
Early work by Vincent Garcia, Eric Debreuve, and Michel Barlaud in Fast k nearest neighbor search using GPU demonstrated that fundamental algorithms like k-Nearest Neighbor (kNN) could be scaled by dividing the reference set into manageable blocks. Similarly, Oliver Deussen and colleagues in A Graphics Hardware Accelerated Algorithm for Nearest Neighbor Search showed that by offloading distance calculations to the GPU and iterating through partitioned blocks, one could achieve significant speedups over CPUs even with the overhead of data movement.
However, simple partitioning introduces a new bottleneck: the latency of the peripheral bus (like PCIe). If the GPU finishes its work faster than the next chunk of data can arrive, the hardware sits idle.
Hiding Latency with Asynchronous Streaming
To solve the idle-hardware problem, modern frameworks use asynchronous data transfers to overlap computation with communication. While the GPU is processing "Chunk A," the system is already streaming "Chunk B" into a different part of the VRAM.
This "pipelined" approach is central to the work of Ismael Boureima, Manish Bhattarai, and their team at Los Alamos National Laboratory. In Distributed Out-of-Memory NMF on CPU/GPU Architectures, they introduced a framework that allows Non-negative Matrix Factorization (NMF) to scale to an 11-Exabyte sparse matrix. By combining batching with asynchronous transfers, they ensured that the GPU's computational engines were never starved for data.
Similarly, Mi Li, Bernhard Pfahringer, and Eibe Frank developed ASB K-means in Large scale K-means clustering using GPUs. Their algorithm uses geometric pruning to skip unnecessary calculations, further reducing the amount of data that needs to be moved and processed, effectively breaking the "VRAM barrier."
Hybrid CPU-GPU Cooperation
Rather than treating the CPU as a mere data loader, hybrid architectures leverage the unique strengths of both processors. The CPU is excellent at complex branching logic and managing massive memory pools, while the GPU excels at uniform, parallel math.
In FusionANNS: An Efficient CPU/GPU Cooperative Processing Architecture for Billion-scale Approximate Nearest Neighbor Search, Haikun Liu and Yuhang Tang describe a system that distributes tasks between the CPU and GPU while optimizing for SSD access. This allows a single entry-level GPU to search through billion-vector datasets that would otherwise require a massive cluster of high-end cards.
Joseph JaJa and Yu Jin applied this logic to spectral clustering in A High Performance Implementation of Spectral Clustering on CPU-GPU Platforms. By offloading heavy matrix multiplications to the GPU while using the CPU for sophisticated control logic (like the ARPACK library), they achieved 400x speedups over standard implementations.
Algorithmic Adaptations for Memory Efficiency
Beyond just moving data faster, researchers are redesigning algorithms to be inherently more memory-efficient.
- Direct Processing on Compressed Data: Zezhou Huang and Krystian Sakowski in GPU Acceleration of SQL Analytics on Compressed Data demonstrated that GPUs can perform SQL queries directly on compressed formats like Run-Length Encoding (RLE). This effectively increases the "logical" memory capacity of the GPU by allowing it to hold more information in the same physical space.
- Task Scheduling: In Accelerating Mixed-Precision Out-of-Core Cholesky Factorization with Static Task Scheduling, Jie Ren utilized static task scheduling and high-speed interconnects (NVLink-C2C) to manage massive matrices. By carefully ordering operations, the system minimizes the need to swap data back to the host.
- Memory-Efficient Sorting: Elias Stehle introduced a hybrid radix sort in A Memory Bandwidth-Efficient Hybrid Radix Sort on GPUs that combines global partitioning with on-chip local sorts. This approach reduces the pressure on global memory bandwidth, allowing the system to sort datasets up to 64 GB—far exceeding the VRAM of the GPUs used in the study.
The Future of Out-of-Core Computing
As datasets continue to grow in fields like genomics, finance, and LLM serving, the reliance on physical VRAM capacity is becoming a liability. The research direction is clear: the future of high-performance computing lies in heterogeneous memory management. By treating the GPU, CPU, and even SSDs as a unified hierarchy, systems like SpeedLoader by Yiqi Zhang and Yang You are proving that we can train and run massive models on modest hardware by rethinking how data flows through the machine.