High-capacity memory
Keep more close to compute.
256 GB of HBM3E per GPU provides space for model weights and active inference data. Model size, precision, context length, and batching determine how that capacity is used.
← Capacity reservations
AMD Instinct
The Infaio Agentic Grid · AMD Instinct
AMD Instinct
MI325X.
High-capacity GPU memory. Flexible compute for the work ahead.
Plan reserved MI325X capacity for model serving, batch inference, and model adaptation on The Infaio Agentic Grid. Match the configuration to your models, your demand, and your growth.
AMD Instinct MI325X · Architecture
Room for the model.
Power for the workload.
Model weights, working context, and concurrent requests all compete for memory. MI325X combines high-capacity HBM3E with GPU compute for training and inference.2
256 GBHBM3E memory per GPU
6 TB/sPeak memory bandwidth per GPU
8 GPUsAMD MI325X platform configuration
Connected accelerators
Scale beyond one GPU.
AMD’s eight-GPU MI325X platform connects accelerators through Infinity Fabric, with 2.048 TB of aggregate HBM3E memory across the platform.
Open software
Build on ROCm.
AMD’s ROCm software stack supports GPU-accelerated AI development. Plan the framework, model, and serving configuration around your workload.
On The Infaio Agentic Grid
A capable GPU.
A more complete architecture.
Infaio is designing MI325X capacity as part of a broader system. GridRunner coordinates model calls and compute placement, while CPU services handle the tools and execution that agents need.
For disaggregated deployments, discuss pairing GPU prefill with specialized decode capacity such as SambaNova SN50.3
Explore specialized decodeStart with the work
Capacity with
a purpose.
Bring the model, the demand, and the outcome you need. We’ll discuss the capacity plan around them.
Model serving & retrieval
Put your context to work.
Plan model serving for assistants, retrieval-augmented applications, and agents that work with company knowledge.
Batch inference
Move through the backlog.
Explore parallel document processing, extraction, classification, and other high-volume inference workloads.
Model adaptation
Make the model your own.
Discuss fine-tuning and evaluation workloads alongside inference, with capacity scoped to the model and method.
Start with the work.
Tell us about your models, applications, and expected demand so we can discuss a suitable capacity plan.
Plan where it runs.
Share your preferred deployment location and requirements for data control, networking, and proximity.
Align on the terms.
Discuss configuration, capacity, timing, and contract requirements before a reservation is confirmed.
Start the conversation
Let's plan
your capacity.
Tell us what you have in mind. Our team can discuss the right scope for your reservation.
Explore AMD Helios MI455X This inquiry does not secure capacity. Availability, configuration, and deployment schedules are confirmed through the contract process.Thank you for your inquiry.
Your reservation inquiry has been submitted to Infaio. We’ll follow up at the email address you provided. Capacity is not reserved until a separate agreement is confirmed.
Reservation rate terms
1. Starting rates are quoted in USD per GPU-hour and apply to long-term contracts with non-refundable reservation fees. Subject to contract terms and conditions. Contract duration, allocation size, configuration, included services, fees, and deployment schedules are confirmed in the applicable agreement. A starting hourly rate is not a quote for a complete node or rack and does not imply on-demand availability.
Submitting an inquiry does not secure capacity, charge a reservation fee, or accept a contract. The Infaio Agentic Grid is in development; architecture descriptions express design intent. Hardware-specific availability and service scope must be confirmed separately.
Hardware references
2. Hardware specifications are from AMD’s MI325X product page and MI325X platform specifications. 6 TB/s is peak theoretical per-GPU memory bandwidth, not application throughput. 2.048 TB is aggregate memory across eight GPUs, not a guarantee that every workload can use it as one transparent memory pool. The eight-GPU platform is an AMD hardware configuration, not a confirmed minimum reservation or Infaio allocation.
Deployment scope
3. Infaio’s GridRunner, GridVault, CPU-service, and GPU/SN50 integration descriptions express planned architecture. Model compatibility, licensing, precision, runtime support, networking, system size, and measured performance require deployment-specific qualification. Hardware capabilities do not establish Infaio benchmark performance, availability, or service guarantees. GPU-hour rates do not imply an all-inclusive heterogeneous system; additional processors and services are scoped in the agreement.
AMD, AMD Instinct, AMD EPYC, AMD Pensando, AMD ROCm, and the AMD logo are trademarks of Advanced Micro Devices, Inc. SambaNova and its logo are trademarks of SambaNova Systems, Inc. Manufacturer marks identify the hardware discussed and do not imply endorsement or an authorized-partner designation.