Saved in:
Bibliographic Details
Main Authors: Husbands, Brandon, Witchborn Systems
Format: Recurso digital
Language:
Published: Zenodo 2026
Online Access:https://doi.org/10.5281/zenodo.19259980
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866901694873337856
author Husbands, Brandon
Witchborn Systems
author_facet Husbands, Brandon
Witchborn Systems
contents <p>Trickums is a GPU memory virtualization system designed to extend the effective VRAM capacity available to AI workloads through a multi-tiered architecture spanning local GPU memory, system RAM, NVMe storage, and distributed remote nodes. By intercepting CUDA and framework-level allocation calls, Trickums presents a unified, oversized memory space while dynamically paging data between tiers based on access patterns and execution context.</p> <p>The system implements demand-driven page migration, eviction policies, and predictive prefetching, enabling large models to execute on hardware with limited native VRAM. Integration with orchestration layers allows coordinated scheduling of model shards and data movement, reducing stall conditions and maximizing resource utilization across heterogeneous environments.</p> <p>This document describes the architecture, memory management strategies, and system design of Trickums, including comparisons to unified memory systems, GPU-direct data paths, and remote memory frameworks. It outlines how VRAM oversubscription can be achieved transparently while maintaining correctness, performance constraints, and scalability in distributed AI compute networks.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_19259980
institution Zenodo
language
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Trickums: The VRAM Illusion System of ForgeBorn
Husbands, Brandon
Witchborn Systems
<p>Trickums is a GPU memory virtualization system designed to extend the effective VRAM capacity available to AI workloads through a multi-tiered architecture spanning local GPU memory, system RAM, NVMe storage, and distributed remote nodes. By intercepting CUDA and framework-level allocation calls, Trickums presents a unified, oversized memory space while dynamically paging data between tiers based on access patterns and execution context.</p> <p>The system implements demand-driven page migration, eviction policies, and predictive prefetching, enabling large models to execute on hardware with limited native VRAM. Integration with orchestration layers allows coordinated scheduling of model shards and data movement, reducing stall conditions and maximizing resource utilization across heterogeneous environments.</p> <p>This document describes the architecture, memory management strategies, and system design of Trickums, including comparisons to unified memory systems, GPU-direct data paths, and remote memory frameworks. It outlines how VRAM oversubscription can be achieved transparently while maintaining correctness, performance constraints, and scalability in distributed AI compute networks.</p>
title Trickums: The VRAM Illusion System of ForgeBorn
url https://doi.org/10.5281/zenodo.19259980