Procurement decisions
Evidence on CPU vs GPU fit helps buy the right hardware for real workloads and budgets.
← Back to Home
Compute
Compute research helps BC Gov use AI in a cost-effective and safe way. We test what runs best on available hardware, how resources should be planned, and how to choose the right model setup for each application.
In plain terms: Procurement Performance Sovereign Sizing Model Fit Accuracy
Why It Matters
How compute research supports decisions
Resource planning
Capacity planning avoids over-provisioning and under-provisioning across on-prem and cloud environments.
Model-to-application fit
Different services need different model sizes, context lengths, and response speed targets.
Operational safety
Measured limits reduce outages and protect service quality when usage grows or model versions change.
What We Analyze
Compute factors in plain language
Prompt and context length
Longer prompts increase memory pressure and can raise latency; this affects concurrency and user experience.
Model type and architecture
Dense vs mixture models, parameter size, and architecture choices all change compute and memory behavior.
Quantization formats
Formats such as lower precision can reduce memory and cost, with trade-offs in quality and stability.
CPU and GPU utilization
Profiling shows whether workloads are compute-bound or memory-bound and where bottlenecks actually sit.
KV cache behavior
Context growth can expand KV cache quickly; we test limits so memory does not blow up during production use.
Throughput and latency
We measure tokens-per-second and response time to align technical setup with service-level expectations.
Sizing Guide
Sovereign AI compute sizing
This living draft connects partner workload discovery to compute sizing, placement policy, licensing questions, and platform readiness for BC Government sovereign AI infrastructure.
Sovereign AI Compute Guide
Current working draft covering heterogeneous compute, GPU/MIG sizing, CPU+AMX use, workload profiles, dynamic utilization, licensing boundaries, and readiness questions. The value of this document is to help partner teams describe what they need to run, and help platform teams translate that into schedulable compute profiles instead of defaulting to generic GPU requests.
PDF Snapshot
PDF snapshot of the latest draft for sharing and roadmap reference. The HTML version is the source tracked for ongoing edits. The PDF can be referenced in roadmap discussions where a stable snapshot is easier to share.
Compute Papers
Published compute research
| Date | Title | Stream | Type |
|---|
No publications match this filter.