calyr.aí INFRASTRUCTURE

Workstation Specification & WP Integration

€10k Hardware + M1–M24 DRG Timeline

Threadripper + RTX 5090 · ESM-2 / cVAE / BO · Parvotec DRG Work Packages

Hardware Procurement — €10k Investment

Bestellte Systeme

GerätSpezifikationGPUKosten
ASUS Ascent GX10
(Storage + Orchestration)
GB10 Prozessor (20c), 128GB RAM, 1TB NVMe €3.415,83
Dell Pro Max Tower T2
(Master Training)
Core Ultra 7-265 (20c, 5.2GHz), 32GB RAM, 1TB NVMe RTX 2000 16GB €3.192,50
Dell Pro Max Tower T2
(Secondary Training)
Core Ultra 7-265 (20c, 5.2GHz), 32GB RAM, 1TB NVMe RTX A1000 8GB €2.199,17
Total System €8.807,50
(€10,5k inkl. 20% MwSt)

GPU Compute Capacity

ParameterRTX 2000 16GBRTX A1000 8GBCombined
CUDA Cores256017924352
Memory16GB GDDR68GB GDDR624GB total
Memory Bandwidth288 GB/s192 GB/s480 GB/s
FP32 Perf~13 TFLOPS~7 TFLOPS~20 TFLOPS
Max Power70W50W120W
Best ForESM-2, cVAE TrainingNN Training, BOSequential pipeline

Key Difference vs. A100: A100 = 312 TFLOPS; RTX 2000+A1000 = 20 TFLOPS (~15× slower). Compensated by sequential, not parallel training.

Work Package Alignment (DRG + Aorta)

WP1: Target Definition (M1–M3)

Hardware Requirement: ASUS + 1× Dell (RTX 2000)

Task: Sequence Library Assembly, Initial Screening, ESM-2 Embedding Pilot

DMS SetupRAW FASTQ/BAM import → HDF5 conversion (CPU-bound on ASUS 128GB)
ESM-2 Pilot100–500 sequences on RTX 2000 (test embedding quality)
GPU Hours~20–40 GPU-h (RTX 2000)
Duration~1 week (sequential, includes validation)

WP2: DMS Library (M2–M8)

Hardware Requirement: ASUS (Cache) + Dell1 (RTX 2000, 24/7)

Task: Full ESM-2 650M Embedding (500k sequences)

Input500k AAV sequences (DRG + Aorta combined)
ESM-2 Model650M parameters on RTX 2000 (BatchSize=8, FP32)
Output~600GB HDF5 (500k × 1280-dim embeddings)
GPU Hours~1200 GPU-h (RTX 2000) = 72h continuous
TimelineM2–M8 (6 months for library; embedding = first 3 weeks)
StorageASUS 128GB RAM (embedding cache) + 1TB NVMe (output)

✓ Fits within M2–M8 window

WP3: Oracle v1.0 (M4–M10)

Hardware Requirement: Dell1 (RTX 2000) + Dell2 (RTX A1000)

Task: Multi-Task Neural Network + cVAE Architecture Design

Model Training4-Task NN (Tropismus, Yield, Thermostability, Immunogenicity) on RTX 2000
cVAE TrainingConditional VAE on RTX A1000 (lower memory footprint)
GPU Hours~150 GPU-h (RTX 2000) + ~100 GPU-h (RTX A1000) = 250 GPU-h total
TimelineM4–M10 (6 months design + validation; training = 4–6 weeks)
Parallel WorkBoth Dells used simultaneously: DDP over Ethernet (limited by network)

✓ Fits within M4–M10 window (slack for hyperparameter tuning)

WP4: BO Runden (M8–M18)

Hardware Requirement: Both Dells (RTX 2000 + RTX A1000) + ASUS (Orchestration)

Task: 3 Rounds of Bayesian Optimization · Library Screening · Ranking

Round 1 (M8–M11)Embedding: 72h (RTX 2000) → NN: 18h (RTX A1000) → BO: 48h (CPU Gaussian Process) = 8–9 days
Round 2 (M11–M14)Same as R1 (validated 2nd candidates)
Round 3 (M14–M18)Same as R1 (refined final ranking)
Total GPU Hours~2400 GPU-h (72h × 3) on RTX 2000 + 54h × 3 on RTX A1000 = 378 GPU-h A1000
Continuous Runtime~30 calendar days for 3 rounds (with sequential GPU scheduling)
Storage PressureCumulative: 3 × 600GB embeddings + 3 × checkpoints = ~2TB (need external NAS)

✓ CRITICAL: Requires external 4TB NAS by M8 (€1.5–2k additional)

WP5: Lead Validation (M14–M24)

Hardware Requirement: Both Dells (parallel wet lab + simulation)

Task: Experimental Validation, Refinement, Final Lead Selection

SimulationMolecular dynamics, structure prediction (CPU on ASUS or Dell CPUs)
ML SupportTransfer learning from Oracle (retrain final NN on 50–100 new candidates per month)
GPU Hours~300 GPU-h (spread over 10 months, ~30 GPU-h/month)
StorageExperimental results, wet lab data, logs (~100GB/month)
AvailabilityBoth Dells in intermittent use (not 24/7 like WP4)

✓ Sufficient capacity; Dells can be repurposed for other tasks in parallel

M1–M24 Integration Timeline

Monthly Hardware Utilization

MonthWP PhaseRTX 2000 (Dell1)RTX A1000 (Dell2)ASUS (Storage)GPU-h / Month
M1–M3WP1: Target Def30% (Pilot ESM-2)100% (Data Prep)~30
M2–M8WP2: DMS Library95% (500k embedding)100% (Cache + Staging)~900
M4–M10WP3: Oracle v1.050% (NN training)40% (cVAE training)100% (Coord.)~350
M8–M11WP4.1: BO Round 199% (ESM-2 + NN)80% (BO inference)100% (Orchest.)~450
M11–M14WP4.2: BO Round 299% (Sequential)80%100%~450
M14–M18WP4.3: BO Round 399%80%100%~450
M14–M24WP5: Validation20% (Transfer Learn.)10% (Inference)100% (Results Storage)~50
Total M1–M24~2730 GPU-h

Status: ✓ Feasible with continuous operation M2–M18 (8 months heavy, 16 months light/idle)

Critical Path

Kostenanalyse & Procurement (€13.5k Active Workbase)

Hauptinvestition: Hardware (€10.5k brutto)

SystemKomponentenNettoBrutto (+20% MwSt)
ASUS Ascent GX10GB10 20c, 128GB RAM, 1TB NVMe€3.416€4.100
Dell Pro Max T2 (RTX 2000)Core Ultra 7-265, 32GB RAM, RTX 2000 16GB€3.193€3.832
Dell Pro Max T2 (RTX A1000)Core Ultra 7-265, 32GB RAM, RTX A1000 8GB€2.199€2.639
HARDWARE SUBTOTAL€8.808€10.571

Zusätzliche Investitionen (Workstation-Ready)

KomponenteSpezifikationKosten (brutto)TimelineKritikalität
4TB NASSynology DS420+ oder QNAP TS-464 (RAID-1)€1.800–2.000Vor M8🔴 KRITISCH
10GbE Netzwerk2× 10GbE Card + 8-port Switch + Kabel€2.500–3.500Vor M2🟡 HOCH
UPS + PDUAPC Smart-UPS 6kVA + Eaton PDU€3.000–3.500Vor M1🟡 HOCH
Software LicensesWeights & Biases + Optuna Pro (optional)€200–300/MonatOngoing🟠 MEDIUM
INFRASTRUKTUR SUBTOTAL€7.500–9.500

📊 TOTAL "ACTIVE WORKBASE": €18.0–20.0k (Hardware €10.5k + Infrastruktur €7.5–9.5k)

RTX 2000+A1000 vs. Enterprise A100 — Kostenvergleich

MetrikRTX 2000+A1000 (€8.8k)4× A100 80GB (€320k)Speedup Faktor
GPU FLOPS (FP32)20 TFLOPS312 TFLOPS15.6×
GPU Memory24GB total320GB total13.3×
Memory Bandwidth480 GB/s2.4 TB/s
ESM-2 Embedding Time (500k seq)72 hours4 hours18×
Per-Round Total (3 batches)8–11 days3–4 days2.5–2.75×
Cost per GPU-Hour€0.36€3.208.9× cheaper
Power Consumption360W peak1.6kW4.4× less

✓ Fazit: €10.5k Setup ist 2.5–3× langsamer, aber 36× billiger und fits within 24-month project window.

Procurement & Shopping Links

Procurement Timeline

M0 (Jetzt)Planung, RFQ an 3 Anbieter, CIO-Genehmigung
M1Bestellung Dells + ASUS, 10GbE-Netzwerk, OS-Planung
M2Inbetriebnahme: BIOS, Ubuntu 24.04 LTS, NVIDIA Driver 555+
M3Pilot-Testing: ESM-2 Embedding (100 seq), NN Training
M8⚠️ NAS-Deployment Deadline (vor BO Round 1)

Kritische Entscheidungen

EntscheidungOption AOption BEmpfehlung
GPU PairRTX 2000+A10002× RTX 2000A (Spezialisierung)
NAS RAIDRAID-1 (2TB usable)RAID-6 (3TB usable)RAID-1 (Budget)
Netzwerk10GbE dedicatedShared 1GbE10GbE (M8 deadline)
OSUbuntu 24.04 LTSWindows 11 ProUbuntu (CUDA stability)
Backup StrategyiCloud + S3 GlacierLocal HDDiCloud + NAS

Storage Architecture (3TB → 4TB Required)

Tier 1: Local NVMe (Installed)

ASUS 1TBOS + Python env + cache layer~150GB headroom
Dell1 1TB (RTX 2000)Embeddings cache + Models + Checkpoints~250GB headroom
Dell2 1TB (RTX A1000)Training data staging + BO results~300GB headroom
Total Installed3TB✓ Sufficient for M1–M10 (single round)

⚠️ Bottleneck for WP4 (3 rounds): 3× 600GB embeddings = 1.8TB. Need external storage by M8.

Tier 2: External NAS (Required by M8)

Recommendation: Synology DS420+ or QNAP TS-464 (4TB RAID-1)

Cost: €1.5–2k (budget separately from €10k hardware)

PurposeWarm archive (Embeddings Rounds 1–3, Model Checkpoints, Logs)
Connection10GbE (or Gigabit Ethernet, acceptable for sequential access)
Capacity4TB RAID-1 (2TB usable, 2TB redundancy)
RetentionHistorical rounds (keep all for reproducibility)

Tier 3: Cloud S3 (Optional)

AWS S3 Glacier or Azure Blob (cold archive after M24) for long-term regulatory compliance (~€0.03/GB/month)

ML Software Stack (Adjusted for RTX 2000/A1000)

Core Stack

ComponentVersionNotes for RTX Hardware
Python3.12via uv venv (locked deps)
PyTorch2.4.0 + cu125Torch.compile() disabled (overhead). FP32 default (RTX A1000 stability).
transformers4.46.0ESM-2 650M (not 3B). gradient_checkpointing=True for 16GB VRAM
biotite0.45.0Sequence I/O, validation
scikit-optimize0.9.0BO, Gaussian Processes (CPU-bound OK)
botorch1.14.0Multi-objective BO (simplified GP kernels for CPU-bound execution)
optuna3.6.0Hyperparameter search (not critical; manual tuning OK on this budget)
pandas / numpy2.2.0 / 2.0.0Data manipulation (not GPU-dependent)

Development & Monitoring

OS & Drivers

Recommended: Ubuntu 24.04 LTS (Dual-boot from Windows 11 Pro)

NVIDIA DriverLatest (555+) for RTX 2000 / A1000 support
CUDA Toolkit12.5 (matches PyTorch cu125)
cuDNN9.0 (RTX 2000: SM 7.0; RTX A1000: SM 8.0 compatible)