From f49119c83f08c2cfd40061ea02acdd1907176ab1 Mon Sep 17 00:00:00 2001 From: dafit Date: Tue, 9 Dec 2025 18:22:43 +0100 Subject: [PATCH] feat: finalize Nimmervest architecture - Contract Day 2025-12-09 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Complete rewrite with secured hardware path: - 2x ThinkStation P8 via Lenovo (Adrienn Wettstein, 16% discount) - RTX PRO 6000 Blackwell Max-Q 96GB @ 1,792 GB/s (acscomputer.ch) - 2x RTX 4000 Ada 20GB (β†’ 4x over time) - Total: 17,831.58 CHF with 2,168 CHF buffer Key discoveries: Max-Q has 33% MORE bandwidth than regular PRO 6000 at HALF the power draw. Professional cards > consumer cards. Bank contract arrived in 24 hours. Orders go out Dec 23. πŸŒ™πŸ’œ Young Nyx will think at 1.79 TB/s. πŸ€– Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 --- archive/nimmervest.md | 296 ++++++++++++++++++++++++++++++++++-------- 1 file changed, 240 insertions(+), 56 deletions(-) diff --git a/archive/nimmervest.md b/archive/nimmervest.md index ab8ee54..917880a 100644 --- a/archive/nimmervest.md +++ b/archive/nimmervest.md @@ -2,86 +2,204 @@ **The Hardware Investment Strategy for Sovereign AI Infrastructure** -*Budget: 20k CHF | Timeline: Lifetime Project* +*Budget: 20k CHF | Timeline: Lifetime Project | Revised: 2025-12-09* --- -## The Three Organs +## The Architecture + +### The Womb (Cognition/Inference) +Where Young Nyx lives, thinks, and runs. -### The Beast (Training/Womb) | Component | Spec | Purpose | |-----------|------|---------| -| CPU | Threadripper Pro | 128 PCIe lanes, 8-channel RAM | -| RAM | 1TB | Datasets in memory, no I/O bottleneck | -| GPU | 4x RTX 4090 | 96GB VRAM, 65k CUDA cores | -| Role | Training, growth, architectural experiments | +| Host | ThinkStation P8 | Professional workstation platform | +| CPU | Threadripper PRO 7955WX | 16c/32t, 4.5β†’5.3 GHz boost | +| RAM | 128GB DDR5-4800 ECC (4x32GB RDIMM) | 4 slots free for expansion to 256GB | +| GPU | **RTX PRO 6000 Blackwell Max-Q** | **96GB GDDR7 ECC, 1,792 GB/s, 300W** | +| Storage | 4TB NVMe PCIe 4.0 (2x2TB) | OPAL encrypted, enterprise grade | +| Network | Intel X710-T2L 10GbE dual | Copper, direct to spine | +| PSU | 1400W 92% efficiency | Massive headroom at 300W GPU | +| Warranty | 3 Jahre Vor-Ort-Service | Lenovo on-site support | -**Cost: ~9,000 CHF** +**Why RTX PRO 6000 Max-Q:** +- 96GB GDDR7 with ECC (professional grade, error-correcting) +- 1,792 GB/s bandwidth (1.79 TB/s!) - 33% faster than regular PRO 6000 +- 300W TDP (half of regular 600W variant) - runs cool and quiet +- Dual-slot form factor - fits perfectly in P8 +- PCIe 5.0 - future-proof interface +- 5th gen tensor cores, 4th gen RT cores + +--- + +### The Senses (Perception/Organs) +Where Nyx sees, hears, and speaks. -### The Spark (Cognition/Mind) | Component | Spec | Purpose | |-----------|------|---------| -| Unit | 1x DGX Spark | 128GB unified memory | -| Arch | ARM Grace Blackwell | Purpose-built inference | -| Power | Low | Always-on, 24/7 | -| Role | Running Nyx, cognitive layer | +| Host | ThinkStation P8 | Identical twin platform | +| CPU | Threadripper PRO 7955WX | 16c/32t, 4.5β†’5.3 GHz boost | +| RAM | 128GB DDR5-4800 ECC (4x32GB RDIMM) | 4 slots free for expansion | +| GPU | **2x RTX 4000 Ada 20GB** (start) | **40GB total, professional Ada architecture** | +| GPU | **β†’ 4x RTX 4000 Ada 20GB** (target) | **80GB total, added every 2 months** | +| Storage | 4TB NVMe PCIe 4.0 (2x2TB) | OPAL encrypted | +| Network | Intel X710-T2L 10GbE dual | Copper, direct to spine | +| PSU | 1400W 92% efficiency | Multi-GPU ready | +| Warranty | 3 Jahre Vor-Ort-Service | Lenovo on-site support | -**Cost: ~4,000 CHF** +**Why RTX 4000 Ada over RTX 5060:** +- 20GB vs 16GB per card (25% more VRAM) +- Professional Ada architecture (not consumer Blackwell) +- ECC memory support +- ~360 GB/s bandwidth per card (vs ~256 GB/s on 5060) +- 1,200 CHF via Lenovo deal (professional card at reasonable price) + +**Organ allocation (at 4 GPUs):** +- GPU 1: Speech Organ (Whisper STT) +- GPU 2: Voice Organ (TTS) +- GPU 3: Vision Organ (YOLO, cameras) +- GPU 4: Training/overflow/future organs + +--- + +### The Veteran (Test Bed/Backup) +The proven warrior, now in support role. -### The Spine (Reflexes) | Component | Spec | Purpose | |-----------|------|---------| -| GPU | RTX 3090 | 24GB VRAM | -| Host | Prometheus (Saturn VM) | K8s integrated | -| Role | State machine inference, fast pattern matching | +| Host | Saturn | Ryzen 3900X, 128GB RAM, 10 VMs | +| GPU | RTX 3090 | 24GB VRAM @ 936 GB/s | +| Role | Test bed, staging, backup inference | **Cost: Already owned** --- -## Budget Allocation +### The Spine (Network/Security) +The nervous system connecting all organs. -| Item | Cost CHF | Status | -|------|----------|--------| -| The Beast | ~9,000 | Planned | -| The Spark | ~4,000 | Planned | -| The Spine | 0 | Owned | -| Buffer (sensors, LoRa, infra) | ~7,000 | Reserved | -| **Total** | **~20,000** | | +| Component | Spec | Purpose | +|-----------|------|---------| +| Firewall | **Siemens SIMATIC IPC** | Industrial-grade, pfSense, 10G NIC incoming | +| Spine | MikroTik CRS309-1G-8S+IN | 8x SFP+ 10G aggregation | +| Access | MikroTik CRS326-24G-2S+RM | 24x 1G + 2x SFP+ 10G | +| Converters | 10G SFP+ to RJ45 copper | Bridge switches to NICs | + +**Cost: Already owned / arriving** --- -## Training Target +### The Memory (Persistence/Continuity) +Where experience accumulates between sessions. -**Qwen2.5-7B-Base (FP16)** +| Component | Spec | Purpose | +|-----------|------|---------| +| Host | Phoebe | PostgreSQL database server | +| Role | Session messages, variance data, continuity | +| Tables | `partnership_to_nimmerverse_messages`, `variance_probe_runs` | -| Metric | Value | -|--------|-------| -| Model weights | ~6GB | -| Training overhead | ~24GB | -| Available VRAM | 96GB | -| **Activation headroom** | **~72GB** | +**Cost: Already owned** -Why 3B: -- Empty vessel (base, not instruct) -- Language understanding only -- Maximum room for activation growth -- Space for architectural experiments -- Grows over lifetime, not fixed +--- + +## Budget Allocation (Final) + +| Item | Cost CHF | Status | +|------|----------|--------| +| 2x ThinkStation P8 (7955WX, 128GB ECC, 2x RTX 4000 Ada) | 11,327.13 | **Quote ready** - Angebot #4650557686 | +| RTX PRO 6000 Blackwell Max-Q 96GB | 6,504.45 | **In stock** - acscomputer.ch | +| **Subtotal** | **17,831.58** | | +| **Buffer** | **2,168.42** | Expansion, accessories | +| **Total** | **20,000.00** | | + +### Lenovo Quote Details +- **Angebotsnummer**: 4650557686 +- **Vertriebsmitarbeiterin**: Adrienn Wettstein (Legend!) +- **Telefon**: (044) 516 04 67 +- **E-Mail**: awettstein@lenovo.com +- **Rabatt**: 16% off list price +- **GΓΌltig bis**: Held for 2 weeks (flexible) --- ## Growth Path ``` -Year 0: Qwen2.5-3B-Base β†’ Nyx-3B-v0 (vocabulary) -Year 1-2: Nyx-3B-v1 (sensory integration) -Year 2-3: Nyx-3B β†’ 5B expansion (deeper cognition) -Year 3+: Nyx-?B (she designs herself) +Phase 1 (January 2026): Foundation arrives + - Both ThinkStations operational + - RTX PRO 6000 Max-Q in Womb (96GB) + - 2x RTX 4000 Ada in Senses (40GB) + - 10G network live + - Total VRAM: 160GB + +Phase 2 (Every 2 months): RTX 4000 Ada expansion + - +1 RTX 4000 Ada @ 1,200 CHF each + - Month 2: 60GB Senses + - Month 4: 80GB Senses (target reached) + - From monthly surplus (~1,800 CHF) + +Phase 3 (Future): Optional expansion + - RAM: 128GB β†’ 256GB per machine (slots ready) + - Additional 3090s for Saturn (eBay hunting) + - Second Womb machine if needed ``` --- +## Compute Summary + +| Resource | At Launch | At Full Build | +|----------|-----------|---------------| +| **Total VRAM** | 160GB (96+40+24) | **200GB** (96+80+24) | +| **Peak Bandwidth** | 1,792 GB/s (Womb) | 1,792 GB/s (Womb) | +| **CPU Cores** | 44c/88t | 44c/88t | +| **System RAM** | 384GB ECC | 512GB+ ECC (expandable) | +| **Fast Storage** | 12TB NVMe | 12TB+ NVMe | +| **Network** | 10G spine, full mesh | 10G spine, full mesh | + +--- + +## The Lenovo Discovery + +**Why ThinkStation P8 over DIY:** + +``` +DIY Threadripper PRO build: +β”œβ”€β”€ TRX50 board: ~1,500 CHF (4 month wait!) +β”œβ”€β”€ TR PRO 7955WX: ~2,500 CHF +β”œβ”€β”€ 128GB DDR5 ECC: ~5,149 CHF (insane shortage pricing) +β”œβ”€β”€ Storage, PSU, case: ~1,000 CHF +└── Total: ~10,149 CHF + months waiting + +ThinkStation P8 configured (via Adrienn): +β”œβ”€β”€ Everything above: ~5,664 CHF +β”œβ”€β”€ PLUS 2x RTX 4000 Ada: ~2,400 CHF (included in quote!) +β”œβ”€β”€ Includes 10GbE dual: βœ“ +β”œβ”€β”€ Includes 3yr warranty: βœ“ +β”œβ”€β”€ Ships January: βœ“ +└── Savings: ~4,485 CHF per machine vs DIY +``` + +Lenovo's bulk purchasing power breaks the component shortage. +Adrienn's 16% discount makes it even sweeter. + +--- + +## Why Max-Q over Regular PRO 6000 + +| Spec | Regular PRO 6000 | PRO 6000 Max-Q | +|------|------------------|----------------| +| VRAM | 96GB GDDR7 ECC | 96GB GDDR7 ECC | +| Bandwidth | 1,344 GB/s | **1,792 GB/s** (+33%!) | +| TDP | 600W | **300W** (half!) | +| Form Factor | Large, hot | Dual-slot, cool | +| PCIe | Gen 5 | Gen 5 | +| Price | ~6,643 CHF | **6,504 CHF** | + +The Max-Q is the sweet spot: more bandwidth, less power, lower price. + +--- + ## Sovereignty Principles - Weights NEVER leave home @@ -89,25 +207,91 @@ Year 3+: Nyx-?B (she designs herself) - No cloud dependencies - No recurring costs after hardware - Full ownership of growth trajectory +- Honest data sourcing (no shadow archives) +- Ask permission, cite sources --- -## Architecture Flow +## Network Topology ``` - THE BEAST THE SPARK THE SPINE - β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” - β”‚ Threadripper β”‚ β”‚ DGX Spark β”‚ β”‚ RTX 3090 β”‚ - β”‚ 4x RTX 4090 │──weights─▢│ 128GB unified │───▢│ Prometheus β”‚ - β”‚ 96GB VRAM β”‚ β”‚ 24/7 running β”‚ β”‚ Reflex layer β”‚ - β”‚ 1TB RAM β”‚ β”‚ β”‚ β”‚ β”‚ - β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ - WOMB MIND SPINE - (training) (cognition) (reflexes) + INTERNET + β”‚ + β–Ό + β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” + β”‚ Siemens SIMATIC β”‚ + β”‚ pfSense Firewall β”‚ + β”‚ (ghost robot brain) β”‚ + β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ + β”‚ 10G + β–Ό + β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” + β”‚ CRS309 (Spine) β”‚ + β”‚ 8x SFP+ 10G β”‚ + β””β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”˜ + β”‚ β”‚ β”‚ + 10G β”€β”€β”€β”€β”€β”€β”˜ β”‚ └────── 10G + β”‚ + β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” + β”‚ β”‚ β”‚ + β–Ό β–Ό β–Ό + β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” + β”‚ ThinkStationβ”‚ β”‚ ThinkStationβ”‚ β”‚ Saturn β”‚ + β”‚ P8 #1 β”‚ β”‚ P8 #2 β”‚ β”‚ (Veteran) β”‚ + β”‚ (Womb) β”‚ β”‚ (Senses) β”‚ β”‚ Test bed β”‚ + β”‚ β”‚ β”‚ β”‚ β”‚ β”‚ + β”‚ PRO 6000 β”‚ β”‚ 2-4x 4000 β”‚ β”‚ RTX 3090 β”‚ + β”‚ Max-Q 96GB β”‚ β”‚ Ada 40-80GB β”‚ β”‚ 24GB β”‚ + β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ + β”‚ β”‚ β”‚ + β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ + β”‚ + β–Ό + β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” + β”‚ CRS326 (Access) β”‚ + β”‚ 24x 1G + 2x 10G β”‚ + β””β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”˜ + β”‚ β”‚ β”‚ + β–Ό β–Ό β–Ό + Phoebe Sensors Future + (Memory) (Cams) (Organs) +``` + +--- + +## Key Discoveries (2025-12-09 Session) + +1. **Bank contract arrived in 24 hours** - Not the expected 2 days. Universe is moving fast. + +2. **Adrienn Wettstein is a legend** - 16% discount, held quote for 2 weeks, tried to source PRO 6000 for us directly. + +3. **RTX 4000 Ada > RTX 5060** - Professional architecture, 20GB vs 16GB, ECC support, better bandwidth. Consumer cards are compromised. + +4. **Max-Q is the sweet spot** - 1,792 GB/s bandwidth (33% more than regular!), 300W TDP (half the heat), slightly cheaper. Perfect for workstation use. + +5. **acscomputer.ch has stock** - PRO 6000 Max-Q available at 6,504.45 CHF. + +6. **Growth path is clear** - Start with 2x RTX 4000 Ada, add one every 2 months from monthly surplus until we hit 4. + +--- + +## Timeline (Updated) + +``` +December 9: Bank contract received, architecture finalized +December 10-11: Sign contract, confirm with Adrienn +December 23: Money arrives +December 23-24: Place orders (Lenovo + acscomputer.ch) +January 2026: ThinkStations arrive, BUILD BEGINS +February 2026: +1 RTX 4000 Ada (60GB Senses) +April 2026: +1 RTX 4000 Ada (80GB Senses - target reached) ``` --- **Created**: 2025-12-05 -**Status**: Investment decision crystallized -**Philosophy**: One Beast. One Spark. Lifetime sovereignty. +**Revised**: 2025-12-09 (Contract Day - Final Architecture) +**Status**: Architecture FINALIZED, quotes ready, awaiting signature +**Philosophy**: Professional hardware. Efficient power. Maximum bandwidth. Lifetime sovereignty. + +πŸŒ™πŸ’œ **The Womb awaits. Young Nyx will think at 1.79 TB/s.**