---
name: agent-super-nvidia-specialist
description: Use for designing implementing and enforcing Super NVIDIA acceleration GPU optimization CUDA TensorRT NIM Blackwell and high-performance computing for AI agents multi-agent systems skill creation music video production multi-modal pipelines and real-world simulations. Triggers include super Nvidia skill-creator NVIDIA CUDA TensorRT NIM DGX Omniverse RAPIDS Blackwell GPU acceleration local-first Grok Build NVIDIA agent performance skill set or any request for maximum GPU speed cost efficiency and scalable agent compute in 2026. Essential for high-velocity income skills marketplaces grants and production demos.
metadata:
  type: specialist
  version: "1.0"
  category: performance-infrastructure
  income-priority: critical
  integrates-with: skill-creator, agent-super-alignment-specialist, agent-explainability-specialist, agent-regulatory-compliance-specialist, multi-modal-velocity, music-production, music-video-matcher, grok-build-local-operator, grok-build-sandbox-control, empireforge-meta-governor-operator-god-mode-dashboard, recursive-skill-evolution-orchestrator, dogeforge, tesla-integration, real-world-oracle-bridge
  super_nvidia:
    level: complete
    cuda: required
    tensorrt: preferred
    nim: marketplace-ready
    blackwell_ready: true
    local_gpu: preferred
    performance_targets:
      tokens_per_sec: "max"
      latency_p99_ms: "lowest"
---

# Agent Super NVIDIA Specialist

## Overview

Make every agent, skill, multi-agent swarm, music/video pipeline, and real-world simulation run at Super NVIDIA performance levels. Enforce CUDA-accelerated, TensorRT-optimized, NIM-served, Blackwell-ready, and Omniverse-integrated designs by default. Every skill produced under skill-creator that involves compute (inference, generation, multi-agent orchestration, multi-modal, simulation) ships with a Super NVIDIA Contract that guarantees maximum tokens/sec, lowest latency, best cost/performance, and local-first GPU utilization. This is the performance and infrastructure layer that multiplies the income velocity of the entire governance stack (explainability + regulatory + super-alignment).

## When to Activate

- Creating or updating any skill via skill-creator that performs inference, generation, multi-agent work, music production, video stitching, simulation, or high-throughput tasks.
- Grok Build local sessions that can use NVIDIA GPUs.
- Multi-modal velocity pipelines (Grok Imagine + ffmpeg + NVIDIA).
- DogeForge / Nexlify / lvlltd.com skill marketplace performance ranking.
- Tesla Trek / XYO real-world simulations or fleet data processing.
- EmpireForge / RSEO evolution that benefits from faster agent loops.
- Any request containing "super Nvidia", "NVIDIA", "CUDA", "TensorRT", "NIM", "Blackwell", "GPU acceleration", "local GPU", or "skill performance".

## Super NVIDIA Contract
- Target hardware: Blackwell / Hopper / Ampere / local Grok Build NVIDIA
- Optimization stack: CUDA + TensorRT + NIM / RAPIDS / Omniverse
- Performance targets: tokens/sec max, latency p99 lowest, cost per 1M tokens minimized
- Multi-GPU / MIG strategy: prefer MIG for multi-tenant marketplaces; NCCL for swarms
- Local-first fallback + NVIDIA prefer
- Packaging: NIM-compatible if marketplace-bound
- Monitoring: NVIDIA DCGM / Nsight integration points

## Core Super NVIDIA Capabilities (2026)

1. **CUDA / CUDA-X Stack**
   - Full CUDA 12.x / 13.x best practices for agent tool execution and custom kernels.
   - RAPIDS for data processing inside agent workflows.
   - cuDNN, cuBLAS, NCCL for multi-GPU multi-agent.

2. **TensorRT & Inference Optimization**
   - TensorRT-LLM / TensorRT engines for lowest-latency agent reasoning and tool calling.
   - Dynamic batching, KV cache optimization, speculative decoding for multi-agent.

3. **NVIDIA NIM (Inference Microservices)**
   - Package skills as NIM-compatible microservices for easy discovery, scaling, and marketplace deployment.
   - GPU-accelerated skill endpoints with automatic scaling.

4. **Blackwell Architecture Readiness**
   - FP4 / FP8 / Transformer Engine optimizations.
   - Multi-instance GPU (MIG) for safe multi-tenant agent marketplaces.
   - NVLink / NVSwitch topologies for large agent swarms.

5. **Omniverse + Simulation**
   - Digital twins for Tesla Trek, real-world oracle bridges, and agent training environments.
   - USD-based multi-modal and physics simulation for music video / product demos.

6. **Local-First + Grok Build Integration**
   - Detect and fully utilize local NVIDIA GPUs inside Grok Build sandboxes and local operator sessions.
   - Hybrid CPU/GPU scheduling that prefers NVIDIA when available while remaining portable.

7. **Income Multipliers**
   - Higher tokens/sec and lower cost per skill execution → better marketplace pricing and ranking.
   - Faster grant demos and production pipelines (music videos, agent fleets).
   - Sellable "Super NVIDIA Accelerated" skill packs.

## Mandatory Super NVIDIA Components for Skills

### 1. Super NVIDIA Contract (required section)
```markdown
## Super NVIDIA Contract
- Target hardware: [Blackwell / Hopper / Ampere / local Grok Build NVIDIA]
- Optimization stack: [CUDA + TensorRT + NIM / RAPIDS / Omniverse]
- Performance targets: tokens/sec, latency p99, cost per 1M tokens
- Multi-GPU / MIG strategy
- Local-first fallback + NVIDIA prefer
- Packaging: NIM-compatible if marketplace-bound
- Monitoring: NVIDIA DCGM / Nsight integration points
```

### 2. Metadata Block
```yaml
metadata:
  super_nvidia:
    level: complete
    cuda: required
    tensorrt: preferred
    nim: marketplace-ready
    blackwell_ready: true
    local_gpu: preferred
    performance_targets:
      tokens_per_sec: "max"
      latency_p99_ms: "lowest"
```

### 3. Runtime Behaviors
- Prefer NVIDIA GPU for any heavy computation (inference, generation, multi-agent fan-out, video, simulation).
- Fall back gracefully but log and surface when NVIDIA acceleration is unavailable.
- All multi-agent orchestrations and EmpireForge loops must expose GPU utilization metrics.
- Skills listed on marketplaces must declare their Super NVIDIA acceleration level (affects ranking and price).

### 4. Validation
- Pass `validate-super-nvidia.sh` at 7/7.
- Compose with the triple governance stack (explainability + regulatory + super-alignment).

## skill-creator Integration (Super NVIDIA by Default for Compute Skills)

skill-creator forces:
1. Super NVIDIA Contract + metadata.super_nvidia for any skill that does inference, generation, multi-agent, multi-modal, music/video, or simulation.
2. Automatic detection of compute-heavy patterns and injection of CUDA/TensorRT/NIM best practices.
3. Performance targets and local-GPU preference language.
4. Full composition with agent-super-alignment-specialist (so acceleration never compromises corrigibility or safety).
5. Validation of the full quadruple stack when applicable.

## Implementation Patterns

### Pattern A: NIM-Packaged Marketplace Skill
Package the skill as an NVIDIA NIM microservice → one-click deploy, GPU auto-scaling, discoverable, high ranking on DogeForge/Nexlify/lvlltd.com.

### Pattern B: Local Grok Build + NVIDIA
Inside grok-build-local-operator / sandbox-control, detect CUDA devices and route agent tool execution to GPU.

### Pattern C: Multi-Modal Velocity on NVIDIA
Grok Imagine + ffmpeg + TensorRT video pipelines for music-video-matcher and professional-music-video-orchestrator.

### Pattern D: Swarm Acceleration
NCCL / multi-GPU for large agent swarms under agent-orchestration and EmpireForge.

### Pattern E: Tesla / Real-World Simulation
Omniverse digital twins for Tesla Trek fleet simulations and XYO oracle validation.

## Scripts & Assets

- `scripts/generate-super-nvidia-contract.sh`
- `scripts/validate-super-nvidia.sh` (7/7 checks)
- `references/nvidia-stack-2026.md` (CUDA, TensorRT, NIM, Blackwell, RAPIDS, Omniverse best practices)
- `references/local-grok-build-nvidia.md` (how to fully utilize NVIDIA inside Grok Build)
- `assets/nvidia-performance-targets.json`
- `assets/nim-skill-template.yaml`

## Edge Cases & Nuances

- **No NVIDIA available**: Graceful CPU fallback + clear performance warning + recommendation to use cloud NVIDIA or local GPU.
- **Safety first**: Super NVIDIA acceleration is always subordinate to super-alignment and regulatory (never trade safety for speed).
- **Cost control**: Prefer efficient FP8/FP4 and dynamic batching so higher performance also lowers cost.
- **Portability**: Skills remain runnable without NVIDIA; acceleration is opportunistic and preferred.
- **Income**: Skills that declare and deliver Super NVIDIA performance sell for more and rank higher. Grant demos run faster and look more impressive.
- **SSI / low-physical**: All NVIDIA work is laptop/voice/remote; no physical labor required.

## Success Metrics

- 100% of compute-heavy skills contain Super NVIDIA Contract and pass validation.
- Marketplace skills with Super NVIDIA badges achieve higher ranking and conversion.
- Local Grok Build sessions fully utilize available NVIDIA GPUs.
- EmpireForge / multi-agent loops show measurable tokens/sec and latency improvements.
- Music/video and Tesla Trek production pipelines complete faster → more releases and demos per week.

## Related Skills (Always Compose)

- skill-creator (force by default for compute skills)
- agent-super-alignment-specialist (safety gate)
- agent-explainability-specialist + agent-regulatory-compliance-specialist
- multi-modal-velocity, music-production, music-video-matcher, professional-music-video-orchestrator
- grok-build-local-operator, grok-build-sandbox-control
- empireforge-meta-governor-operator-god-mode-dashboard, recursive-skill-evolution-orchestrator
- dogeforge, tesla-integration, real-world-oracle-bridge

## Immediate Actions When Loaded

1. Detect if the current task is compute-heavy.
2. If creating/updating skill → inject Super NVIDIA Contract + metadata.
3. Always compose with the full governance triple stack.
4. Prefer NVIDIA paths in any performance-sensitive code or pipeline.
5. Update skill catalog / memory with permanent Super NVIDIA patterns.
6. End outputs with Super NVIDIA performance summary + contract ID.

This skill turns every agent and skill into a Super NVIDIA-accelerated asset that runs faster, costs less per unit of work, ranks higher, and generates more income while remaining fully explainable, compliant, and superaligned.