HOW TO DELIVER QUALITY OF SERVICE IN GPU VIRTUALISATION… WITH HYPERLANE

As embedded systems become more complex and performance-critical, particularly in markets like automotive, consumer electronics, and cloud gaming, ensuring predictable GPU behaviour is no longer optional. In virtualised environments, consistency, safety, and responsiveness matter just as much as raw power.

At Imagination, our hardware virtualisation solution, HyperLane, is engineered with one goal: to deliver Quality of Service (QoS) without compromise. This blog explores the core QoS mechanisms in our GPU architecture, and how they enable real-time, multi-tenant performance across diverse workloads.

WHY QoS MATTERS IN GPU VIRTUALISATION

In any virtualised system, multiple virtual machines (VMs) must share a single GPU. These VMs often serve vastly different applications, each with its own performance and safety requirements:

• A real-time automotive dashboard must maintain a locked 60 fps.
• An AI engine might need periodic bursts of GPU power.
• A navigation app can tolerate occasional latency.

Without built-in QoS features, contention between these tasks can lead to dropped frames, inconsistent behaviour, or even critical system failures.

THE CHALLENGE: SHARING GPU RESOURCES ACROSS DIVERSE WORKLOADS

Traditional GPU virtualisation techniques lack the isolation and scheduling precision required by embedded markets. Our approach solves this through a powerful combination of firmware control, real-time metrics, and robust hardware-enforced task separation.

HOW HYPERLANE ENSURES PREDICTABLE GPU PERFORMANCE

HyperLane is a key part of our architecture. It enables up to 16 VMs per core, each with dedicated control paths, eliminating hypervisor mediation at runtime. Combined with our QoS mechanisms, this provides a future-proof framework for real-time, mixed-criticality systems.

IMAGINATION’S 5 KEY QoS FEATURES

1. ROUND-ROBIN SCHEDULING

Our default firmware behaviour allocates GPU time in a fair, round-robin cycle—ideal when all VMs have equal priority. Example: A fragment-heavy app runs at 120 fps solo; when two VMs run concurrently, each maintains 60 fps.

2. PRIORITY-BASED SCHEDULING

VMs can be assigned priority levels. The GPU then pre-empts lower-priority workloads, guaranteeing critical tasks, such as a safety dashboard, never miss their deadlines.

3. DRIVER ISOLATION GROUPS

Workloads are grouped by safety/security levels. These groups do not execute in parallel, ensuring freedom from interference while still allowing concurrency within a trusted group.

4. DoS PROTECTION

To prevent GPU lock-up from faulty or malicious shaders, we enforce context-switch deadlines and pipeline auto-resets. This ensures continuity for high-priority workloads without developer intervention.

5. REAL-TIME MONITORING WITH PVRTUNE

Our PVRTune tool enables live performance tuning across VMs, showing GPU/CPU usage, bottlenecks, and per-VM metrics. It connects securely to a privileged VM to give deep visibility into real-time resource sharing.

HYPERLANE IN ACTION

REAL-WORLD USE CASES: FROM AUTOMOTIVE TO CLOUD GAMING

HyperLane’s QoS features are already in use across a range of markets:

• Automotive: Ensuring IVI, instrument clusters, and ADAS tasks run independently and predictably
• Consumer: Delivering glitch-free UIs in smart TVs while running AI workloads in the background
• Cloud gaming: Isolating player sessions while maintaining frame-rate parity under load

DELIVER PREDICTABLE PERFORMANCE WITH IMAGINATION

In virtualised GPU environments, performance without compromise means more than just speed. It’s about fairness, safety, and predictability, across every workload.

With HyperLane at the core of our GPU architecture, we offer a complete hardware virtualisation framework that meets the growing demands of edge computing.

Download our white paper on GPU Virtualisation to learn how you can implement scalable, secure, and real-time graphics solutions.