eRacks/AIRLIE

eRacks/AIRLIE is a 2-to-4-GPU AI inference and serving platform on an open-frame chassis that puts your budget into GPUs and VRAM, not an enclosure. Powered by a single-socket AMD EPYC 9005-series (Turin) processor with 128 PCIe 5.0 lanes natively - enough for 6 full-speed x16 GPU connections without any PCIe switch overhead.

CPU Options

• EPYC 9354 – 32 cores / 64 threads, 256MB L3, 280W – entry EPYC, strong single-thread
• EPYC 9554 – 64 cores / 128 threads, 256MB L3, 360W – mid-range, excellent throughput
• EPYC 9654 – 96 cores / 192 threads, 384MB L3, 360W – top single-socket, max parallelism

Memory

Up to 1.5 TB DDR5 ECC RDIMM across 12 channels. Configure from 128 GB to 1.5 TB. Large RAM pools enable CPU-based inference on 70B+ models alongside GPU acceleration.

GPU Configurations (1 to 6)

• 1-2 GPUs: LLM development, fine-tuning smaller models (7B-13B)
• 3-4 GPUs: Multi-user inference, RAG pipelines, concurrent workloads
• 4 GPUs: High-throughput serving, 70B-class quantized inference (need more GPUs? see eRacks/AIRA)

Open-Frame Chassis Advantage

The open-frame design provides unrestricted airflow for high-TDP cards (300-600W), eliminating the thermal throttling common in dense rack chassis. Rack-mountable with a standard 4U shelf adapter.

Use Cases

• Local LLM inference with Ollama, vLLM, or llama.cpp
• Multi-user AI serving for teams and research groups
• Fine-tuning and LoRA training on 7B-34B models
• RAG pipelines with local vector search
• On-premise AI with full data privacy (no cloud egress)

All components sourced from standard distribution channels - no vendor chassis lock-in, short lead times, field-replaceable parts. Ships with Ubuntu 26.04 LTS, Ollama, and Open WebUI pre-installed.

Power: every 2-4 GPU build with Arc Pro B70-class cards runs from a single power supply on a standard office circuit. Higher-wattage GPU picks (RTX PRO 6000-class) get a power check at order.