Compute Capacity Planning
Know What Your Infrastructure Needs — Before It Becomes a Problem
Kern Labs works alongside engineering leads and platform owners to size, review, and plan accelerator compute — with transparent assumptions and numbers that hold up under scrutiny.
What We Offer
Three Ways to Work With Kern Labs
Each engagement is scoped to where a team actually stands — early sketching, existing infrastructure, or multi-horizon growth planning.
Capacity Forecast Session
A half-day working session to estimate near-term accelerator capacity needs based on your team's expected workloads. The maths stays transparent throughout, and every assumption is named.
- Forecast worksheet included
- Variable monitoring list
- Suited to leads at the sketch stage
Utilisation Review & Right-Sizing
A two-session review of current compute usage to identify where resources sit idle and where they run tight. Recommendations are grounded in your own figures, not generic benchmarks.
- Written utilisation report
- Prioritised adjustment list
- For platform owners with existing infra
Multi-Phase Capacity Roadmap
A three-month engagement building a phased capacity roadmap across several planning horizons, aligning growth expectations with budget constraints and procurement lead times.
- Roadmap document + revision support
- Milestone reviews included
- For organisations scaling steadily
Plan With Confidence
Where Does Your Infrastructure Planning Stand Today?
Many teams defer capacity conversations until a project is already underway. A short planning session at the right moment can reduce last-minute scrambles and keep procurement timelines realistic.
Why Kern Labs
How Planning Ahead Changes the Picture
Structured capacity work gives teams a foundation that improvised estimates rarely provide.
Assumptions Made Visible
Every forecast we produce names its inputs. Teams see exactly which variables drive the numbers, making it straightforward to update as conditions shift.
Scope Matched to Need
A single-session forecast suits a team sketching early plans. A multi-phase roadmap suits one managing sustained growth. We work at the scale your situation calls for.
Independent Perspective
Kern Labs does not sell hardware or cloud credits. Our recommendations have no stake in which platform you choose — only in whether the plan holds up.
Lead Time Awareness
Accelerator procurement often takes longer than teams expect. Planning horizons that account for lead times prevent gaps between demand and available capacity.
Deliverables You Can Use
Sessions produce worksheets, reports, and roadmap documents that teams can hand to finance, procurement, and leadership — not slide decks that live on one person's laptop.
Based in Penang
Kern Labs operates from George Town with an understanding of the Malaysian technology and research landscape, including regional procurement channels and institutional context.
Accelerator Infrastructure
Planning Around NVIDIA AI Hardware
Most of the teams we work with are building on NVIDIA's accelerator stack. Understanding how these platforms behave under load — and how they are acquired — is central to any capacity plan that holds up in practice.
Why NVIDIA Hardware Shapes the Planning Problem
NVIDIA's data centre GPUs — H100, H200, and the newer Blackwell series — are the dominant compute substrate for large model training and inference. Their architecture dictates memory bandwidth constraints, interconnect topology (NVLink, InfiniBand), and the way batches must be sized for efficient throughput. Capacity planning without accounting for these characteristics produces estimates that look correct on paper but break under real workloads.
Procurement lead times for NVIDIA accelerators vary considerably. Depending on tier, region, and supply cycle, lead times can run from a few weeks for spot allocations to six months or more for dedicated node orders. A capacity plan that does not build in this buffer leaves teams scrambling when a project timeline accelerates.
NVIDIA's software stack — CUDA, cuDNN, TensorRT, and the NeMo and Triton frameworks — also influences utilisation figures. How well a workload is optimised for the hardware directly affects how much compute is actually available versus how much is nominally provisioned. Right-sizing work at Kern Labs accounts for this gap between theoretical and realised throughput.
The current workhorses for training runs and high-throughput inference. Hopper architecture introduces Transformer Engine and FP8 precision, which change utilisation calculations significantly compared to earlier generations.
The next generation introduces significantly higher memory capacity and a new NVLink Switch architecture. For teams planning 18–36 month horizons, Blackwell availability and pricing should be factored into phased roadmaps now.
Integrated multi-GPU systems that simplify deployment but carry different procurement dynamics than individual cards. Understanding which form factor suits your scale and budget is part of the planning work we cover.
Workloads that are not well-optimised for NVIDIA's CUDA stack often achieve 40–60% of theoretical GPU utilisation. Identifying this gap is frequently where right-sizing work finds the most headroom.
What Our Planning Work Covers in an NVIDIA Context
GPU memory sizing relative to model parameter counts and batch requirements
Interconnect bandwidth planning for multi-node training (NVLink, InfiniBand)
Mapping NVIDIA hardware generations to a phased procurement roadmap
Identifying underutilised capacity through DCGM and profiling data
Cloud vs. on-premise trade-off analysis for NVIDIA instances (A100, H100 spot/reserved)
Lead time buffers and budget staging for Blackwell-generation transitions
Kern Labs is an independent planning practice. We are not affiliated with or endorsed by NVIDIA Corporation.
Common Questions
Frequently Asked Questions
A few questions that come up regularly before teams reach out.
What kinds of teams does Kern Labs typically work with?
Do I need to share sensitive internal data to start a session?
How long does it take to see a deliverable?
Are sessions conducted in person or remotely?
What is the pricing structure — is it fixed or variable?
Does Kern Labs recommend specific hardware vendors or cloud platforms?
Our Location
Find Kern Labs in George Town
15A, Lebuh Farquhar, 10200 George Town, Penang
Get in Touch
Start a Conversation
Share a bit about what you're working through and we'll come back with a straightforward reply about how we can help.