Unified compute model
Reuse model and operator work across deployment targets.
HOLYCORES
AI ACCELERATOR
NOVA
An open acceleration platform for datacenter training and inference.
NOVA ACCELERATOR CARDS
Choose the card that fits the system.
N100 targets general development and inference nodes. N200 targets dual-chip throughput and multi-card scaling with planned active and passive variants.
COMPARE NOVA MODELS
One architecture. Different system boundaries.
The following targets support product planning and joint validation. They are not final production or certified specifications.| Planning dimension | NOVA N100A | NOVA N200A | NOVA N200P |
|---|---|---|---|
| Form factor | Single-chip active card | Dual-chip active card | Dual-chip passive card |
| Relative compute scale | 1× baseline | Approx. 2× N100A | Approx. 2× N100A |
| Target memory | 24 GB | 48 GB | 48 GB |
| Host interface | PCIe 5.0 ×16 | PCIe 5.0 ×16 | PCIe 5.0 ×16 |
| Card fabric | Host PCIe | 2× HC-Fabric | 2× HC-Fabric |
| Target board power | Up to 225 W | Up to 350 W | Up to 300 W |
| Cooling | Dual-slot active | 2.5-slot active | Dual-slot passive |
| Target deployment | Developer workstations, inference servers | High-throughput inference, multi-card nodes | High-density rack servers |
Planning targets may change following silicon, board, thermal, and system validation. Final availability, compatibility, and specifications will be defined in future product materials.
DESIGN PERSPECTIVE
One foundation. Three scales.
The portfolio is a continuous path from cloud systems to datacenter acceleration and edge inference—not a set of isolated devices.Composable systems
Co-design compute, fabric, cooling, and operations around real workloads.
Open evolution path
Start with validation, then expand to clusters and edge fleets.
FROM IDEA TO OPERATION
A platform is more than hardware.
Every engagement connects architecture, software, validation, and production operations.
01Characterize the workload
02Validate jointly
03Integrate the platform
04Scale and optimize
ENGAGEMENT MODEL
Validate deliberately. Scale with confidence.
01Characterize the workload
02Validate jointly
03Integrate the platform
04Scale and optimize
FAQ
Questions worth answering early.
Are all products generally available?+
These pages describe product direction. Availability, specifications, and schedules will be confirmed in formal product materials.
Why one architecture?+
It reduces repeated engineering when models move between deployment environments.
Can evaluation start small?+
The planned path starts with development and validation before scaling to production.
CHOOSE YOUR PERSPECTIVE
One platform. Different decisions.
Business leaders
Understand platform scope, deployment choices, and the long-term path.
Follow this path →System architects
Evaluate compute, fabric, software, operations, and facility constraints.
Follow this path →AI developers
Focus on model entry points, operators, tooling, and portability.
Follow this path →TECHNICAL EVALUATION
Ask the questions behind the specification.
Formal specifications and measured results will be published only after engineering validation.| Dimension | Question | Validation |
|---|---|---|
| Compute architecture | Adapt to different models and precision strategies | Build utilization and accuracy baselines with representative workloads |
| Memory hierarchy | Reduce avoidable movement and bandwidth stalls | Profile working sets and end-to-end bottlenecks |
| Scale-out fabric | Grow from device to node and cluster | Validate topology, communication, and scaling efficiency |
| Software stack | Map framework models reliably to hardware | Check coverage, correctness, and tunability |
| System engineering | Balance power, cooling, density, and serviceability | Validate against the target facility and workload |
| Operations | Deploy, observe, upgrade, and diagnose | Establish telemetry, health, version, and rollback workflows |
WORKLOAD MAP
Design around the work—not the label.
Training and fine-tuning
Throughput, communication, precision, and long-run stability.
Online and batch inference
Response time, generation rate, concurrency, and utilization.
Enterprise and sovereign AI
Data boundaries, operational control, auditability, and lifecycle.
Robotics and industrial vision
Local latency, power, interfaces, and environment.
ADOPTION PATH
Move from evidence to operations.
Development validation
Create a baseline with representative models and evaluation systems.
Private infrastructure
Integrate with customer security, datacenter, and operational practices.
Scale deployment
Expand clusters or edge fleets against capacity and service goals.
PRODUCTION READINESS
Performance is only one part of a system.
Reliability
Health, fault isolation, recovery, and long-run validation.
Security
Boot, identity, access, update, and data-boundary planning.
Observability
Telemetry that connects model behavior to system state.
Lifecycle
Versioning, compatibility, rollback, maintenance, and support.
RESOURCES & STATUS
Clarity before claims.
Platform overview
Product direction, architecture principles, workloads, and evaluation framework on this site.
Evaluation checklist
Workload definition, environment requirements, success metrics, and test planning.
Technical materials
Specifications, compatibility, SDKs, docs, release notes, and validated model status will follow maturity.
01 / ARCHITECTURE
Throughput without losing programmability
Scalable compute cores and on-chip fabric map flexibly to different models.
02 / FORM FACTOR
Fits the modern datacenter
A standard accelerator form factor supported by HOLYFLOW.
READY TO BUILD?