Scientific Node — Compute Bridge Architecture
Design Philosophy
NOT a gaming PC. NOT a replace-HPC machine. A bridge.
| Layer | Hardware | Purpose |
|---|---|---|
| Mobile Frontend | MacBook M5 Pro (Aorta) | Interactive coding, field research, remote control |
| Local Compute | Scientific Node (Parvotec) | Training, BO, simulations, overnight batches |
| Heavy Compute | HPC / VSC Vienna | Large-scale CFD, ensemble methods, production runs |
Why Threadripper instead of Intel/AMD EPYC?
- PCIe 5.0: Fast NVMe + multi-GPU expansion ready
- 24–32 cores: Good balance for CPU parallelism (CFD, molecular dynamics)
- Cost: €600–1.200 vs EPYC €2.500–8.000
- Expandability: TRX50 boards have 10GbE + dual GPU slots standard
Why RTX 5090 not RTX PRO 6000?
| RTX 5090 (32 GB) | ~€4.500–6.000 | CUDA 12.5, Tensor cores, fast; good for training + BO |
| RTX PRO 6000 (96 GB) | ~€13.000–16.700 | 96 GB VRAM; aber fast dein gesamtes Budget |
Decision Rule: Start with 32 GB. If you hit VRAM ceiling regularly → upgrade later. Mainboard is 2-GPU-ready.
10GbE is Non-Negotiable
Parvotec generates ~1.8 TB embeddings per BO round. Without 10GbE:
- 1GbE = 40+ hours per transfer → defeats overnight batch scheduling
- 10GbE = 4 hours per transfer → fits within M1–M8 window
Cost addition: €300–800 (cards + switch) — mandatory.
Scientific Node — Hardware Bill of Materials
Complete System (~€8–10k netto, €9.5–12k brutto)
| Komponente | Modell | Netto | Brutto | Link / Quelle |
|---|---|---|---|---|
| CPU | AMD Threadripper 9960X (24c, 5.7 GHz) | €800 | €960 | Geizhals.at |
| Mainboard | ASUS ProArt TRX50-CREATOR | €1.100 | €1.320 | 10GbE, dual PCIe x16, ECC support |
| RAM | Crucial CT2K64G56C46S5 (2×64 GB, DDR5-5600, ECC) | €1.800 | €2.160 | Alternate.de |
| GPU | NVIDIA RTX 5090 32 GB (Founder's Edition or AIB) | €4.500–5.500 | €5.400–6.600 | aktuell €5–6k, MSRP war $1.999 |
| System SSD | Samsung 990 Pro 2 TB (PCIe 4.0) | €180 | €216 | OS, containers, compilers |
| Data SSD | SK Hynix Platinum P41 4 TB (PCIe 4.0) | €280 | €336 | Embeddings cache, training data, models |
| PSU | CORSAIR AX1600i (1600 W, Titanium-rated) | €400 | €480 | RTX 5090 + Threadripper = ~700 W peak; 1600 W = 2.3× headroom |
| Case | NZXT H7 Flow RGB oder Corsair 5000T | €120 | €144 | Good airflow, quiet, full-size for later GPU expansion |
| Cooler (CPU) | Noctua NH-U14S TR5-SP6 oder be quiet! Dark Rock Pro TR4 | €150 | €180 | Threadripper runs hot; high-quality cooler essential |
| 10GbE Card | Mellanox ConnectX-6 (QSFP28, dual-port) | €250 | €300 | Or Cisco ENIC (cheaper used) |
| 10GbE Switch | NVIDIA Mellanox SN3700 (16-port managed) | €1.500 | €1.800 | High-performance, power-efficient, redundant PSU support |
| Cables + Misc | QSFP+ cables (2×3m), PCIe risers, thermal paste | €100 | €120 | |
| TOTAL HARDWARE | €11.180 | €13.416 | Without 10GbE switch: €9.5k netto | |
Optional Expansions (Für später)
- RTX 5090 #2: Add second GPU when VRAM becomes bottleneck (~€5.5k)
- RAM → 256 GB: For large molecular dynamics sims (~€3.6k for 4×64GB more)
- NVMe #3: 8 TB archive NVMe (€400–600)
Cost Analysis & Procurement Timeline
Budget Breakdown
| Kategorie | Komponenten | Netto | Brutto (20% MwSt) |
|---|---|---|---|
| Core Compute | CPU + Mainboard + RAM + GPU | €8.200 | €9.840 |
| Storage | 2×NVMe (2TB + 4TB) | €460 | €552 |
| Power & Cooling | PSU + Cooler | €550 | €660 |
| Gehäuse | Case + Cables + Misc | €220 | €264 |
| SYSTEM SUBTOTAL | €9.430 | €11.316 | |
| Network (Shared) | 10GbE Card + Switch + Cables | €1.850 | €2.220 |
| TOTAL WITH 10GbE | €11.280 | €13.536 | |
Procurement Strategy (M0–M2)
| M0 (Now) | Order CPU, Mainboard, RAM, PSU (longest lead time) |
| M0+2w | GPU pre-order (RTX 5090 supply constrained; check Geizhals daily) |
| M1 | Receive core components, assemble + BIOS setup |
| M1+2w | GPU arrives, 10GbE NIC + Switch (can order in parallel) |
| M2 | Full system operational: Ubuntu 24.04 LTS + CUDA 12.5 |
Austrian Retailers
- Geizhals.at — Price comparator, ships from DE/AT
- Alternate.de — B2B pricing, fast Vienna shipping
- Cyberport.de — Gaming + Pro hardware, occasional discounts
- NBB.com — Budget alternative (sometimes lower prices)
- IT&T Austria — Local distributor for ASUS/Corsair
Work Package Integration — What Do You Need When?
WP1: Target Definition (M1–M3)
| Task | Hardware Requirement | Why Workstation? | Fallback |
|---|---|---|---|
| Sequence assembly (FASTQ → HDF5) | CPU (16 cores) + 32 GB RAM | Threadripper 9960X handles bioinformatics pipelines faster than MacBook | MacBook M5 (slower) |
| ESM-2 pilot embedding (100–500 seqs) | GPU (RTX 5090, ~4h) | 5090 ~100× faster than M5 Metal; parallelize batches | MacBook M5 (~30h, offline only) |
| Result visualization + QC | CPU only | Local Jupyter, metadata inspection | MacBook (preferred, but workstation OK) |
WP1 Hardware Status: ✓ READY with Threadripper + RTX 5090
WP2: DMS Library (M2–M8)
| Task | Hardware Requirement | Timeline | Criticality |
|---|---|---|---|
| Full ESM-2 embedding (500k seqs) | GPU 24/7 (RTX 5090, ~72h continuous) | M2–M5 (first 3 weeks) | 🔴 CRITICAL PATH |
| Data staging (500k seq input) | CPU + 128 GB RAM + 4 TB SSD cache | M1–M2 | 🟡 High |
| Embedding validation sampling | CPU + metadata QC (no GPU) | M5–M8 | 🟠 Medium |
| External NAS deployment | 10GbE network + 4 TB RAID-1 NAS | Before M8 BO start | 🔴 CRITICAL (storage pressure) |
⚠️ KEY CONSTRAINT: Workstation must have 10GbE + external NAS by M8, else 3TB local NVMe becomes bottleneck for WP4 (1.8 TB embeddings × 3 rounds).
WP3: Oracle v1.0 (M4–M10)
| Task | Hardware | GPU Hours | Why Workstation |
|---|---|---|---|
| 4-Task NN training (4 epochs) | RTX 5090 + 128 GB RAM | ~150 GPU-h | 5090 = 10–20× faster than M5 Metal; memory bandwidth critical |
| cVAE training (conditional VAE) | RTX 5090 (lower footprint) | ~100 GPU-h | Generative task; M5 Metal insufficient |
| Hyperparameter sweep (Optuna) | CPU (Threadripper 24c) + GPU | ~50 GPU-h (BO search) | Parallelizable; Threadripper >> M5 cores |
| Model checkpointing + validation | Storage + CPU | — | Both systems OK; prefer workstation for automation |
WP3 Status: RTX 5090 + Threadripper ✓ SUFFICIENT for M4–M10 timeline.
WP4: BO Runden (M8–M18)
| Round | Workload | Hardware Bottleneck | Estimated Time |
|---|---|---|---|
| Round 1 (M8–M11) | Embed (72h) → NN Infer (18h) → BO GP (48h CPU) | GPU embedding (RTX 5090) | ~8–9 calendar days (sequential) |
| Round 2 (M11–M14) | Repeat Round 1 with 2nd candidate set | GPU + 4 TB NAS (1.2 TB embeddings now) | ~8–9 days (may slow with NAS I/O) |
| Round 3 (M14–M18) | Repeat Round 1 refined final set | Storage pressure: ~1.8 TB cumulative | ~8–9 days (watch NAS bandwidth) |
| Total BO | 3 rounds sequential (no DDP parallelism) | GPU sustained 24/7 × ~24 days | ~30 calendar days (fits M8–M18) |
⚠️ CRITICAL: 10GbE + NAS Tier 2 are MANDATORY for WP4. Without them, 4TB local NVMe insufficient.
WP5: Lead Validation (M14–M24)
| Wet Lab Integration | MacBook M5 (field) + Workstation (overnight) | MacBook handles field data ingest; Workstation processes batch |
| Transfer Learning | RTX 5090 + Threadripper (re-training) | ~300 GPU-h spread over 10 months |
| Inference on New Candidates | Either system; prefer M5 for interactivity | Oracle inference = ~1s/seq on M5 (acceptable) |
Summary: Hardware Readiness by WP
| WP | Core Compute | Storage | Network | Status |
|---|---|---|---|---|
| WP1 | Threadripper + RTX 5090 | 6 TB local NVMe | 1GbE OK | ✓ Ready M2 |
| WP2 | ✓ (Threadripper 24c parallelism) | ⚠️ Need 4TB NAS by M8 | ⚠️ Need 10GbE before BO | ✓ M2–M8 feasible |
| WP3 | ✓ (RTX 5090 training) | ✓ (Tier 1 + Tier 2 warm) | ✓ 10GbE for Parvotec↔Mac sync | ✓ Ready M4 |
| WP4 | ✓ (RTX 5090 24/7) | 🔴 CRITICAL: 4TB NAS MANDATORY | 🔴 10GbE MANDATORY | ⚠️ M8 only if NAS+10GbE deployed |
| WP5 | Workstation + MacBook (hybrid) | ✓ Archive to S3 Glacier | ✓ | ✓ Ready M14 |
M1–M24 Timeline & Workstation Usage
Monthly GPU/CPU Utilization (Workstation)
| Month | WP | GPU Usage | CPU Usage | NAS Pressure | Notes |
|---|---|---|---|---|---|
| M1–M2 | Setup + WP1 start | ~20% (pilot embedding) | 50% (sequence prep) | — | OS setup, CUDA drivers, Python env |
| M2–M5 | WP2 main (DMS) | 95% (500k embedding) | 30% (I/O scheduling) | — | Continuous 72h ESM-2 run; local cache OK |
| M4–M8 | WP3 overlap | 60% (NN training) | 80% (hyperparameter BO) | — | Both systems busy; parallelizable |
| M8 | WP2 complete | — | — | 🔴 NAS deployment DEADLINE | Must have 4TB RAID-1 before WP4 start |
| M8–M11 | WP4.1 (Round 1) | 99% (BO embedding) | 70% (GP process) | ~600 GB written | Intensive; 10GbE critical for staging |
| M11–M14 | WP4.2 (Round 2) | 99% | 70% | ~1.2 TB total | Potential NAS bandwidth bottleneck; monitor |
| M14–M18 | WP4.3 + WP5 start | 99% (BO final) | 50% (validation) | ~1.8 TB final | Hottest period; M5 does WP5 field work |
| M18–M24 | WP5 main | 20% (transfer learn) | 40% (analysis) | Archive to S3 | Workstation in idle/standby most days |
Critical Dependencies
- 🔴 M8 Deadline: 4TB NAS + 10GbE operational, else WP4 BO blocked
- 🔴 M2 Deadline: Workstation fully operational (Ubuntu 24.04 + CUDA + Python env)
- 🟡 M5 Deadline: 500k embeddings complete (WP2 bottleneck for WP4 downstream)
Workstation Power Consumption
| Idle | ~150 W (mostly PSU fan) |
| CPU-only load | ~300 W (Threadripper + fans) |
| GPU streaming (RTX 5090) | ~700 W peak (Threadripper 200W + GPU 450W + PSU loss) |
| 24/7 continuous (M8–M18) | ~5,040 kWh/month (worst case, avg ~600W) |
Recommendation: UPS 6 kVA (€3k) highly recommended for WP4 BO rounds. Sudden power loss = 72h embedding lost.