eRacks/AIRA

eRacks/AIRA delivers up-to-6-GPU AI compute from a single-socket AMD EPYC 9005-series (Turin) processor on an open-frame chassis. EPYC provides 128 PCIe 5.0 lanes natively - enough to give every GPU a full x16 link with lanes to spare, no PCIe switch needed. No backplane markup. No vendor lock-in.

CPU Options

• EPYC 9354 – 32 cores / 64 threads, 256MB L3 – cost-optimized build
• EPYC 9554 – 64 cores / 128 threads, 256MB L3 – balanced throughput
• EPYC 9654 – 96 cores / 192 threads, 384MB L3 – maximum parallelism

Memory

Up to 1.5 TB DDR5 ECC RDIMM across 12 channels. At 6 GPUs plus high RAM, AIRA can run 405B-class quantized models in hybrid CPU+GPU mode while simultaneously serving additional inference requests from GPU VRAM.

GPU Configurations (1 to 8)

• 4 GPUs: Cost-effective multi-user serving; 70B quantized models
• 5 GPUs: 70B full-precision inference, parallel fine-tuning
• 6 GPUs (by quote): Maximum throughput; 405B quantized, multi-model hosting

$/TFLOP Advantage

8x RTX 6000 Ada (48 GB each) = 384 GB total VRAM at roughly 1/4 the cost of an equivalent H100 SXM5 system. For inference workloads that do not require NVLink (most transformer serving), AIRA delivers competitive throughput at dramatically lower total cost.

Open-Frame Chassis

The open-frame design handles the thermal demands of 8 high-TDP GPUs without the airflow compromises of a closed 4U chassis. Rack-mountable with a standard 4U shelf adapter.

Use Cases

• High-throughput LLM serving for organizations or ISPs
• 70B-405B model inference (quantized or full-precision)
• Large-batch RAG and embedding pipelines
• Fine-tuning and training on datasets up to ~50B tokens
• Multi-model hosting: run several 13B-34B models concurrently
• On-premise AI infrastructure replacing $200K+ HGX configs

Same proven architecture as the eRacks/AIRLIE, scaled up to 6 GPUs. Ships with Ubuntu 26.04 LTS, Ollama, vLLM, and Open WebUI pre-installed.

Power: builds up to 4 Arc Pro B70-class GPUs run from a single power supply on a standard office circuit. 5-GPU builds ship with 2,000W-class or dual power supplies and want a dedicated 20A or 240V circuit. 6-GPU builds are quoted, so we can spec your site power (240V or dual 120V circuits) and PSU configuration correctly - a class of build eRacks has delivered for years.