LIQID UltraStack with AMD Instinct™ MI350P.Run every model your teams need in parallel, one model per GPU, on a single AMD EPYC™ server. Available now from 4 to 30 GPUs.
Buy another server for every new job, or run every model your teams need on one.
Each new job adds models to route, extract, draft and check.
Agents use 10–100× more tokens per task.
Each new job means another server to buy, rack and run.
Another OS, more NICs, ports and licenses.
Regulated and confidential data can’t go to a hosted API.
Self-hosted models don’t change under you.
LIQID GPU Pooling attaches 4 to 30 MI350P GPUs to one AMD EPYC host over PCIe Gen 5. Each GPU runs its own model: 20 GPUs, 20 models at once.
Each model runs on its own GPU, so capacity grows in a straight line.
144 GB of HBM3E holds a 70B model at FP8 or gpt-oss-120b.
Standard PCIe cards, air-cooled, on a server you know.
Every configuration is available now, and every one runs one model per GPU.
Add GPUs and the next job’s models, not a new server.
Ten models side by side, one per GPU.
4 × Llama 3.1 8B triage · 4 × Llama 3.3 70B drafting · 2 × gpt-oss-120b QA
Financial crimes investigation · clinical documentation · practice-group review · analyst Q&A
Twenty models running in parallel.
Five agents × 4 GPUs: prior auth, appeals, eligibility, claims triage, denials
Revenue cycle suite · defensible e-discovery · FOIA and redaction · dedicated tenant models
Every department’s models on one host.
10 × team jobs · 14 × department suite · 6 × shared reasoning and Q&A
Bank-wide model hub · health-system hub · firm-wide hub · sovereign model catalog
Standard parts: your server, LIQID fabric and chassis, AMD GPUs.
Two questions settle the size: which jobs do you need to run, and how many users or agents run them at once?
*Efficiency targets compare against multi-server deployments of the same GPUs; they are LIQID internal projections using AMD published specifications, and results vary by configuration and workload.
Performance figures for the AMD Instinct MI350P are based on AMD published specifications and LIQID Phase-3 testing from 1 to 20 GPUs, and are subject to change. 30-GPU figures are projected from measured 20-GPU scaling. Every configuration runs one model per GPU. Models shown are US and EU open-weight models. 7B-class: 4,653 tok/s on 1 GPU, 18,526 on 4, 44,383 on 10, ~89,000 on 20 (measured); ~133,000 on 30 (projected). AMD specifications from the AMD Instinct MI350P product brochure.