CESARops × Lemonade

Why Lemonade Matters

• Affordable local AI: Executes critical reasoning locally without dependency on high-cost cloud instances.
• Heterogeneous hardware: Integrates seamlessly across diverse computing environments.
• AMD Vulkan execution: Maximizes throughput on available AMD silicon via Vulkan.
• Keeping older hardware useful: Extends the operational life of existing GPU deployments.
• Role-qualified workers: Ensures models only perform tasks they are empirically validated to handle.

Why Additional AMD Hardware Matters

Additional hardware directly scales the platform’s engineering capabilities by enabling:

• More concurrent specialist workers
• Larger qualification suites
• More simultaneous validation
• Faster SAR processing
• Broader testing across hardware generations

Project architecture

Forge is a contract compiler for AI systems. It converts nonlinear human intent into validated, provenance-aware machine contracts, then qualifies open-weight models for bounded execution roles. Mission orchestration and logic stay fixed. Inference backends—like Lemonade—function strictly as interchangeable execution engines underneath.

cesarops3 — permanent role

A dedicated Lemonade worker on Polaris GPUs. The server stays; certified roles evolve.

Primary backend

Lemonade Server · llamacpp:vulkan

Hardware

AMD RX 580 (8 GB) · AMD RX 570 (4 GB)

Primary role

Thinker

Initial model

Phi-3 Mini (GGUF, Vulkan)

Why this setup

• Fast enough for orchestration
• Good instruction following
• Efficient on Polaris hardware
• Leaves room to evaluate larger candidates later

Scorecard entry

Name the backend explicitly — not just “AMD” — so the fleet reads as interchangeable workers.

Worker name

cesarops3

Backend

Lemonade

Runtime

llamacpp:vulkan

Hardware

RX580 + RX570

Role

Thinker

Primary model

Phi-3 Mini

Status

Production Candidate

Qualification

Under Continuous Evaluation

Role certification (evolves on the same server)

• Thinker✓
• Mission Intake✓
• Historical Planning✓
• Rust Coding✓
• Drift PlanningTesting
• ReviewerPending

Why Phi-3 first?

For the competition, Phi-3 Mini is a sensible baseline on Polaris. Later comparisons use the same qualification framework — evidence-based, not preference-based:

• Phi-3
• Qwen3-0.6B
• Gemma
• Llama variants

Live worker dashboard

Real telemetry — not a mock. Cards lead with Backend, then expand to runtime, models, hardware, role, decode, and the latest hillshade classification metrics. Refreshed from /lemonade/metrics.json.

Loading metrics…

At a glance: which fleet a worker belongs to, what hardware it uses, what role it performs, and how classification has been measured.

Hardware Strategy

CESARops intentionally targets inexpensive and legacy hardware. Forge selects workers based on demonstrated capability rather than assuming the presence of a single monolithic GPU. This approach extracts significant utility from edge accelerators, CPU fallbacks, and older generations of silicon.

• Legacy AMD Polaris (RX 580, RX 570)
• Legacy NVIDIA Tesla P100
• CPU Fallback Older Xeons
• Not Required MI300, RTX 5090

Metrics

Measured

Sustained decode throughput on 512 generated tokens. Both Polaris GPUs were active during Vulkan inference.

In Progress

• Drift Planning Qualification
• Broader AMD Polaris hardware validation

Experimental

• Direct Vulkan Engine Rust Port

Planned

• Reviewer Role Certification

STEM

One long-term goal is to use CESARops in STEM education by having students build inexpensive GPS drifters, predict where they will travel, deploy them, and compare real-world observations against model predictions. Students learn programming, environmental science, physics, engineering, and scientific reasoning while helping improve Great Lakes drift understanding.

Gilcher drift map Andaste drift map Use Back to return

Open source

• CESARops
• Forge
• Mission Control
• Wayfinder
• NauticUVS
• Lemonade integration

Mission Control Public Forge hub Forge routes Use Back to return

Why Lemonade instead of writing directly against llama.cpp?

CESARops was intentionally designed around stable interfaces rather than specific inference engines. Lemonade allows AMD hardware to become another interchangeable worker in the CESARops ecosystem without changing mission logic, validation, provenance, or specialist planning. That separation of concerns lets CESARops evolve independently from the inference runtime while benefiting from improvements in Lemonade over time. Direct llama.cpp (Vulkan or CUDA) and NVIDIA NIM remain first-class peers under the same contract.

Looking ahead

• Additional worker qualification
• Expanded AMD hardware testing
• More local models
• Community contributions
• Educational deployments