Start from the workload
Users, peak requests, simultaneous turns, model family, context length and tools per turn drive the GPU count. A developer's setup is not a sizing guide.
Deployment · On premises
Install RatioCore inside your own data center on AMD, NVIDIA or a mix of both. Prompts, memory and audit records stay on infrastructure you operate.
What your users see
Your people work with agents in the messaging apps you approve, Microsoft Teams in this example. Where RatioCore runs is your choice. Each channel you enable is an outbound path, so review it like any other connector.
Ana RibeiroShift supervisor
RatioCore show me last week's downtime for line 3
RatioCore Plant assistantApp
Line 3 had 2 stops, 47 minutes in total. Both were conveyor jams on Tuesday.
When it fits
On premises suits organizations that must keep prompts, documents and decisions on infrastructure they control, or that already own GPU capacity. Models run locally, so no prompt leaves the network unless you route an agent to an external provider.
Reference stack
RatioCore keeps the agent layer separate from model serving. The same three tiers apply on every path; only where each one runs changes.
A dedicated Linux server for the control plane, agent runtimes and gateways. Today the Core scales vertically on one host, so give it ample CPU and memory.
On its own server or the application host for a pilot. For production, add WAL archiving, off-host backups and a rehearsed restore.
AMD and NVIDIA nodes can serve side by side. Each model family gets its own route; failover works within a family, so give each at least two nodes when availability matters.
Your monitoring, secrets, logging and backup tools wrap the platform. RatioCore exposes health and readiness endpoints and Prometheus metrics.
AMD hardware
Data-center platforms for production pools, and workstation-class cards for small sites and pilots.
| System or card | GPUs | Memory per GPU | Best for |
|---|---|---|---|
| Dell PowerEdge XE9785 | 8x Instinct MI355X, with 2x EPYC 9005 CPUs; air-cooled | 288 GB | Large production pools |
| Supermicro AS-4126GS-NMR-LCC | 8x Instinct MI355X; liquid-cooled, 4U | 288 GB | Dense production pools |
| Supermicro AS-8126GS-TNMR | 8x Instinct MI350X; air-cooled, 8U | 288 GB | Production pools |
| Instinct MI325X / MI300X servers | 8-GPU platforms from server vendors | 256 GB / 192 GB | Existing Instinct fleets |
| Radeon PRO W7900 | Workstation or server card | 48 GB | Pilots, small sites, embeddings |
| Radeon AI PRO R9700 | Workstation or server card | 32 GB | Pilots, small models, embeddings |
All listed AMD GPUs appear in ROCm's production support matrix. Radeon cards are supported on Ubuntu and RHEL. Sources: Dell XE9785, Supermicro AMD, ROCm GPU specs, ROCm system requirements
NVIDIA hardware
NVIDIA systems for every scale, served through open servers such as vLLM and SGLang or through NVIDIA's own serving software.
| System or card | GPUs | Memory | Best for |
|---|---|---|---|
| NVIDIA GB200 NVL72 | 72 Blackwell GPUs and 36 Grace CPUs; liquid-cooled | 13.4 TB per rack | Very large model fleets |
| DGX B300 / HGX B300 | 8 Blackwell Ultra GPUs | 2.1 TB per system | Large production pools |
| DGX B200 / HGX B200 | 8 Blackwell GPUs | 1.4 TB per system | Large production pools |
| H200 (SXM or NVL) | Hopper GPU | 141 GB | Production pools, existing Hopper fleets |
| RTX PRO 6000 Blackwell Server Edition | PCIe server GPU, up to 4 MIG instances | 96 GB | Mid-size pools, mixed workloads |
| L40S | PCIe server GPU | 48 GB | Pilots, smaller models, embeddings |
NVIDIA NIM serves models behind industry-standard APIs and needs an NVIDIA AI Enterprise license in production. NVIDIA Dynamo is open source and offers an OpenAI-compatible front end over vLLM, SGLang and TensorRT-LLM. Sources: GB200 NVL72, DGX B300, HGX, H200, RTX PRO 6000, L40S, NIM, Dynamo
Sizing
Users, peak requests, simultaneous turns, model family, context length and tools per turn drive the GPU count. A developer's setup is not a sizing guide.
A single-slot node shared by many agents is the most common bottleneck. Prefer multi-slot servers or more nodes per model family.
Large-memory nodes can serve the main reasoning model while smaller cards handle embeddings, summaries and pilot agents.
Installation is a guided procedure today. Hardware lists are examples confirmed on vendor pages in November 2026, not a certified compatibility matrix; validate drivers, models and serving software in your environment.
Other paths
Serve open-weight models on AMD Instinct GPUs from Azure, Oracle Cloud, DigitalOcean and other providers, with the Core in your cloud account.
Explore AMD cloud Your data centerInstall RatioCore on your own servers with AMD Instinct, NVIDIA, or both, for regulated and network-restricted environments.
Explore on premises Fully cloudRun every tier in your AWS, Azure, Google Cloud or Oracle account, with self-served models, managed model APIs, or both.
Explore fully cloudFAQ
Still deciding? Bring your questions to a walkthrough.
Yes. RatioCore routes each agent to a model endpoint, not to a GPU vendor. AMD and NVIDIA servers can sit in the same pool, each serving its own model family.
It can. With local models and no external channels or connectors enabled, nothing needs to leave your network. Each channel, connector or external model you enable adds its own outbound path.
For a pilot, yes: one application host with PostgreSQL in a container and one or two GPU nodes. Production separates the database, adds backups and sizes the GPU pool to peak concurrency.
See it on your business
Bring a process your team runs today. In a demo we show how agents could run it, where your people stay in charge, and what a focused pilot could prove.