Home/Products/Bud FCSP
Infrastructure · GPU virtualization

Turn every GPU into a flexible, governable resource pool.

GPUs are expensive and most of them sit underused. Bud FCSP - Fixed Capacity Spatial Partition - splits a single physical GPU into multiple secure virtual GPUs with real isolation, so small inference jobs stop consuming whole cards.

2–4×
Higher GPU utilisation from running multiple workloads per card
1000×
Faster context creation — 78µs against 84ms for vGPU
3600×
Faster memory tracking — 0.3µs against 1.1ms for HAMi
$14M
Projected savings for a 1000-GPU enterprise deployment
Why it exists

Six problems that show up at fleet scale.

Wasted resources

Small inference workloads occupying entire GPUs, leaving expensive hardware mostly idle.

Team contention

Teams competing for GPU access with no safe way to share between tenants.

Vendor lock-in

Vendor-specific solutions that limit hardware choice and create single-supplier dependency.

Poor isolation

Basic time-slicing gives inadequate isolation, so noisy neighbours degrade everyone.

Cost explosion

Costs rising faster than capability as the fleet grows to meet demand.

Scheduling chaos

No priority-based scheduling or fair allocation across teams and projects.

What it delivers

Five capabilities that change the economics.

True GPU resource isolationEach container gets a virtual GPU with complete isolation and guaranteed resources. If one workload spikes it cannot take another tenant’s memory or compute.
Higher GPU utilisationMultiple workloads per card instead of one job per GPU — 2–4× utilisation, which is millions a year at fleet scale.
Hardware-agnostic virtualizationWorks across heterogeneous accelerators rather than a single proprietary stack, giving one virtualization layer over mixed hardware.
Kubernetes-native integrationIntegrates into the cluster directly. Developers keep using CUDA, PyTorch, TensorFlow and vLLM unchanged.
Software or MIG modeFlexible partitioning, so the isolation level can match the workload rather than the other way round.
Performance

Where the speed comes from.

Lock-free shared memory and intelligent stream classification, measured against the alternatives.

Lock-free shared memory

Atomic operations allow concurrent access without mutex locks, removing the context-creation bottleneck.

Smart stream classification

NCCL bypass for distributed training, with kernel categorisation for attention and FFN workloads.

Zero-copy memory access

Direct memory operations with no intermediate copies.

OperationBud FCSPNVIDIA vGPUHAMi
Context creation78µs84ms~1ms
Memory tracking0.3µs~100µs1.1ms
Hardware supportHeterogeneous acceleratorsA100 / H100 onlyLimited
Scheduling overhead~40nsHigherHigher
Architecture

Four modules.

Memory tracker

Lock-free shared memory using atomic operations for concurrent context access.

Kernel rate limiter

Compute throttling at microsecond granularity.

Stream classifier

Workload categorisation for optimised GPU scheduling.

Process manager

Container-aware process isolation and lifecycle management.

Isolation modes

From zero overhead to strict quota.

Idle GPU resources flow automatically to active tenants under work-conserving scheduling, so utilisation does not depend on someone rebalancing by hand.

NoneZero enforcement, for trusted single-tenant environments or performance benchmarking.
BalancedIsolation against sharing, tuned for multi-tenant production.
StrictHard quota enforcement matching MIG behaviour, for compliance-sensitive workloads.
AdaptiveAdjusts the isolation level dynamically against real-time contention and workload patterns.
Capabilities

What you can set per container.

Fractional GPU allocation

Define compute percentage and memory limits per container.

Hard memory enforcement

Applications cannot exceed their assigned GPU memory.

Compute throttling

Time-shard scheduling prevents hogging while keeping concurrency.

Priority scheduling

Pause lower-priority work for urgent inference.

Real-time monitoring

Per-container GPU usage, with policies adjustable live.

Oversubscription

Safely run more workloads than physical GPUs when demand is bursty.

FCSP or MIG

They are not mutually exclusive.

FCSP can orchestrate MIG partitions on supported NVIDIA GPUs, combining hardware isolation with software governance. Where hardware-level partitioning is mandated, use MIG underneath; where you need heterogeneous hardware, finer granularity or oversubscription, FCSP covers what MIG cannot.

The business value

At 100+ GPUs, a 20% utilisation improvement usually justifies the layer on its own. It reduces fleet size, delays new hardware purchases, raises revenue per GPU, improves SLA predictability, and makes multi-tenant governance tractable.

Operating 100+ GPUs?

At that fleet size even a 20% utilisation improvement usually pays for the layer. We can model it against your workload mix.