Canadian engineering. Real-world constraints.
More intelligence.
Less overhead.
Efficient AI execution. Passive airspace awareness. Maple Silicon builds systems for the hardware and environments that matter.
Built in Canada · NVIDIA compute · Core ML research
Make every GPU count.
Compiler-driven inference acceleration for NVIDIA workloads.
Smaller models. Real devices.
SparseFlow V2: model compaction and Core ML evaluation on iPhone.
See without transmitting.
Passive drone detection and operator-facing airspace awareness.
01 — The Problem
Inference cost grows
faster than model quality.
LLM teams are under pressure to ship lower latency and lower cost at the same time. The default answer is more GPUs, which makes margin worse and deployment slower.
SparseFlow focuses on the buyer question that matters: throughput per GPU. It improves execution efficiency on supported NVIDIA hardware without forcing a full model-stack redesign first.
02 - Systems
Two systems,
kept separate.
SparseFlow is the compute product. Maple Shield is the passive drone detection system. CAIRN is the detection and intelligence engine inside Maple Shield.
SparseFlow
Model inference acceleration for NVIDIA workloads. SparseFlow is for teams trying to cut latency and GPU cost without rebuilding the whole stack first.
- A100 validated; RTX 4090 benchmarking in progress
- Up to 1.6-1.7× on FFN-heavy inference paths
- No retraining required to evaluate value
- Built for throughput, latency, and cost-sensitive deployment
Maple Shield
Zero-RF passive drone detection, persistent tracking, and operator-facing airspace awareness for contested and remote environments.
- CAIRN 5-layer detection and intelligence engine
- Passive vision detection - zero RF emissions
- Integrates with existing Axis, Bosch, and Hanwha camera infrastructure
- 10 km vision target for fixed critical-site coverage
- Evaluation conversations available
Research / SparseFlow V2
AI efficiency,
closer to the device.
We are extending SparseFlow research into the Apple ecosystem: physically compacting vision models, exporting them to Core ML, and measuring how they behave on an actual iPhone.
Independent research by Maple Silicon. This work does not imply an Apple partnership, endorsement, or contract.
Model compaction
Compare recovered dense, channel-pruned, and physically compact vision models.
Core ML evaluation
Connect source-model validation to Core ML export and device-side benchmarking.
Evidence before claims
Short-run iPhone latency results are mixed. Sustained performance, energy use, and on-phone task accuracy remain open.
03 — Engineering Signals
Numbers behind the
product.
150% of cuBLAS baseline
benchmark platform
release, open source
(2026)
04 — Evaluation Path
Built to test fit before
you commit budget.
No Long Integration Project
Start from the existing workload or sensing geometry before anyone commits to a heavy deployment motion.
Use The Current Stack
SparseFlow benchmarks against existing PyTorch and Hugging Face style paths. Maple Shield evaluations start from current site coverage assumptions.
Baseline Versus Wedge
Compare dense baseline versus SparseFlow savings, or map passive coverage density versus traditional counter-UAS system economics.
Decision-Ready Readout
For the right setup, the goal is a fast answer on whether the system deserves a deeper benchmark review or deployment conversation.
05 — About
About Maple Silicon
Built by an engineering-led company for teams who need serious technical products, not presentation-layer promises.
Maple Silicon Inc. is a Canadian company building systems products for AI compute efficiency and edge awareness. We care about operational reality: constrained hardware, measurable gains, and deployment paths that stand up outside a demo.
SparseFlow reflects that mindset in GPU inference. Maple Shield reflects it at the sensing edge. Across both, the company bias is the same: make the system useful in the environment where customers actually operate.
Background in CUDA kernel development, GPU systems engineering, and product-minded technical execution. Built dense GEMM kernels reaching 31.47 TFLOPS on RTX 3090. Currently leading Maple Silicon product development and customer engagements end-to-end.
"The next performance win is not just a better model. It's better execution on the GPUs teams already pay for."
06 — Work With Us
Pick the system that matches the mission.
SparseFlow is for inference acceleration. Maple Shield is for passive drone detection and airspace awareness. Start with the right page, then request evaluation if the fit looks real.