Origin
Built for our lab first
Born inside Holocyte's own materials-science computing.
Built for research teams with real constraints
Own the compute your research depends on.
ArrayMeld designs and deploys modular local clusters for large-model AI inference and scientific simulation — sized around your workload, installed in your lab or office, and validated before handover.
Local & private
Modular 2–8+ nodes
AI & scientific workloads
Adelaide engineered
A practical middle ground between one workstation and institutional HPC. Extreme-scale jobs may still belong in cloud or national facilities; ArrayMeld gives your team an accessible platform for daily development, inference and selected simulations.
Method
Workload-led design
We size the system to your model or simulation, not a catalogue.
Platform
Open & documented
Linux / ROCm / open runtimes you can inspect and keep.
Fit
Grant-friendly
Staged expansion designed around real research budgets.
01 · Two directions, one platform
Two ways research teams put ArrayMeld to work.
The same modular architecture serves two kinds of team — and we keep them distinct, because a coding-agent deployment and a molecular-dynamics workflow succeed on different criteria. Choose the path that fits, or ask us in a workload assessment.
For research & HPC teams
Scientific computing & research
The workloads ArrayMeld was born from inside Holocyte's materials-science work: simulation, modelling and method development, run locally and burst to HPC only when a job truly demands it.
- Full native AVX-512 on x86 — GROMACS, LAMMPS, NAMD and the standard HPC libraries run as shipped, no porting to ARM
- Molecular dynamics & enhanced sampling (e.g. OpenMM, GaMD)
- Parameter sweeps and embarrassingly parallel tasks
- Pre-/post-processing, method development, early- to medium-scale simulation
For AI-application teams
Private AI applications
Run large open-weight models on hardware you own — coding agents, research assistants, retrieval and evaluation — with no per-token billing and no prompt leaving your network.
- 400B-parameter-class capacity from ≈ AUD 20,000 of hardware
- Selected large open models across aggregate node memory
- OpenAI-compatible local endpoint for your IDE agents and tools
- GLM-5.2-class flagship, or Qwen 3.6 / Gemma 4 on request
02 · The gap we fill
More than a workstation. More accessible than enterprise HPC.
A single machine runs out of memory. The public cloud meters every hour and every prompt. Shared university HPC means queues and restrictions. ArrayMeld is the practical space in between — capacity your team owns, controls and can grow.
A practical research platform — not a hardware bundle.
Our position: use local compute for sustained, sensitive and development workloads; keep cloud and institutional HPC available for bursts and extreme scale. A good recommendation from us sometimes points you elsewhere.
Our position: use local compute for sustained, sensitive and development workloads; keep cloud and institutional HPC available for bursts and extreme scale. A good recommendation from us sometimes points you elsewhere.
The workstation ceiling
One high-end workstation eventually hits memory, expansion or availability limits — and the frontier-scale open models your field is adopting simply do not fit on it. ArrayMeld scales capacity across standardised nodes.The cloud meter
Metered cloud is convenient until sustained workloads, data sovereignty and predictable budgeting matter. ArrayMeld gives you owned, unmetered capacity that keeps prompts, code and data on your own network.The shared-HPC queue
Institutional clusters are excellent for the largest jobs, but queues, software restrictions and administrative dependencies slow daily iteration. ArrayMeld sits in your lab, available when you are.The enterprise-appliance wall
Enterprise AI appliances carry a high entry cost and an inflexible architecture. ArrayMeld starts small, is documented end to end, and expands node by node as funding allows.03 · Why ArrayMeld
Designed around the realities of research budgets and technical risk.
Our commercial promise is not "the fastest at any price." It is a useful, supportable system whose compromises are explained and whose target workload is tested.
ExpandableStart at two nodes, grow to four, then to a custom scale-out — capacity added in stages as budget allows.
Local controlData, models, scheduling and access stay on infrastructure your organisation owns.
ValidatedEvery system leaves with measured results against a representative workload — not a spec sheet.
Open & documentedAn open Linux/ROCm stack and a full operating record — you are never left with an undocumented black box.
Budget-consciousOwned capacity scoped to your workload, with no per-token metering and a credible upgrade path.
04 · How the options compare
Where each route leaves you — and what ArrayMeld addresses.
We do not compete on being the cheapest box. We compete on the limitation every alternative leaves on the table for a budget-constrained research team.
| Route to more compute | Where it is strong | The limitation for a research team |
|---|---|---|
| High-end workstationSingle machine | Simplicity; excellent for one user, one job | Hits memory, expansion and availability ceilings; largest models won't fit |
| Public cloudRented, metered | Elastic bursts; no capital outlay | Ongoing metered cost, data-control concerns and variable access for sustained work |
| Shared university HPCInstitutional | Outstanding for the largest, tightly-coupled jobs | Queues, software restrictions and administrative dependencies slow iteration |
| Enterprise AI applianceTurnkey vendor | Integrated support at the top end | High entry cost and a fixed architecture with limited upgrade flexibility |
| ArrayMeldModular local cluster | Owned, private, workload-scoped capacity in your lab | Complementary to cloud/HPC for extreme scale — sized to sustained, local, development workloads |
On economics, plainly: a lower component price only matters when the finished system reliably runs your workload. ArrayMeld prices the design, integration, validation and handover — and returns a total-cost comparison against your actual usage during the assessment, rather than a headline savings claim.
05 · Starting points, not rigid bundles
Choose a useful first step. Expand when the workload proves it.
Three configurations sized to how research capacity actually grows — from a validation pilot to a production-style team platform to a custom scale-out.
Entry & validation
ArrayMeld 2
A two-node pilot to prove fit, develop workflows and run selected smaller-scale work.
- Nodes2 × AI Max+ 395
- Aggregate memory256 GB (2 × 128)
- InterconnectDirect or Fabric
- Best forpilot & development
Recommended
Production-style
ArrayMeld 4
The primary platform: real headroom for large models, longer context and daily shared team use.
- Nodes4 × AI Max+ 395
- Aggregate memory512 GB (4 × 128)
- InterconnectDual 25GbE Fabric
- Best fordaily team platform
Custom scale-out
ArrayMeld Scale
Six, eight or more nodes engineered to a specific workload, with a formal site and support plan.
- Nodes6–8+ custom
- Aggregate memory768 GB+ (6–8+ × 128)
- InterconnectFabric or Hybrid
- Best forlargest open models
Aggregate memory figures are raw capacity summed across separate nodes — not one unified VRAM pool. Operating system, runtime and workload headroom reduce usable capacity. Final configuration, inclusions and price are fixed in the written quotation, with GST itemised separately.1
06 · A lower-risk buying process
We begin with the workload, not the parts list.
This process gives researchers, IT and procurement the same definition of success before any hardware is ordered — so there are no surprises after the handshake.
STEP 01
Workload fit callA short, no-obligation conversation about what you need to run, why it should be local, and whether ArrayMeld is even the right answer.
STEP 02
Design & quoteWe assess software, memory, site and budget, then return a scoped architecture, a bill of materials and a fixed written scope — GST itemised.
STEP 03
Build & validateWe assemble, configure and tune the system, then measure it against a representative workload before it ever leaves us.
STEP 04
Deliver & hand overInstalled on-site or shipped, with documentation, an acceptance report, administrator handover and a defined support boundary.
07 · Built for our lab. Productised for yours.
ArrayMeld grew out of Holocyte's own research-computing needs.
We did not begin with a catalogue of computers. We began with research that needed more capacity than a workstation could give and more control than a metered service allowed.
Our team works in materials science, where atomistic calculations, molecular modelling, polymer simulation and AI-assisted R&D create real pressure for local compute. We designed and operated a modular multi-node platform for our own work, standardising a maintainable, expandable node architecture that also proved ideal for local AI inference.
University collaborators and research peers then began asking us to deploy the same capability for them. ArrayMeld is the productisation of that internally proven system.
"The first customer was our own research team."
— Holocyte, Adelaide · South Australia
The need
Frontier-scale models and simulations that no single workstation could hold — and data too sensitive to send off-site.
The build
A compact, standardised multi-node node architecture, refined across repeated design and validation cycles.
The proof
It ran our own materials-science and AI-assisted workloads day after day, in our own lab.
The demand
Peers asked us to build the same system for their teams — so we productised it.
Today
ArrayMeld makes that practical research infrastructure available to other budget-constrained teams — operated by Holocyte Pty Ltd during the launch phase.
08 · A candid fit assessment
Strong where local ownership matters. Honest about where it does not fit.
An audience that reads spec sheets deserves a vendor that says when to buy something else. ArrayMeld will recommend another route when the workload is a poor fit.
Strong fit when…
- ✓Your workload needs more memory or parallel capacity than one workstation provides
- ✓Sensitive code, models or data should stay on infrastructure you control
- ✓You have sustained, repeatable local workloads rather than rare bursts
- ✓Budget arrives in stages and you value a credible expansion path
- ✓You want validation, documentation and support — not just the cheapest parts
Another platform may be better when…
- ✕The workload needs extreme low-latency, tightly-coupled accelerator interconnects
- ✕Your core software is CUDA-only or unsupported on an AMD/ROCm stack
- ✕You need very large-scale training or data-centre-class availability
- ✕Demand is rare and bursty enough that cloud rental stays cheaper
- ✕The site cannot accommodate the power, cooling, noise or network requirements
09 · What arrives with the system
A deployable platform, evidence and a clear handover.
The hardware is the cheap part. The product is the integrated environment and the record that proves it works — designed to remove ambiguity between engineering, procurement and end users.
Documented buildAn as-built bill of materials, network map, versioned software stack and configuration record — the whole system, written down.
Acceptance reportMeasured results against a representative workload: load, timing, throughput, memory, thermals and multi-hour stability.
Administration guideStart-up, shutdown, recovery and maintenance procedures for your nominated users and administrators.
Handover sessionA working walkthrough so your team can operate, monitor and troubleshoot the platform with confidence.
Transparent scopeBill of materials, substitutions, GST, delivery, warranty pathway and support boundaries — all stated in writing.
Expansion pathA credible plan for adding nodes, fabric or storage later, so the first purchase is never a dead end.
10 · Frequently asked
The questions to settle before buying.
Straight answers on the things that matter to a technical evaluator. Deeper engineering detail lives on the architecture page.
Is this one large shared-memory GPU?
No — and we will not sell it as one. It is a coordinated multi-node system: each node has its own local memory and compute, and the runtime distributes the model and data across them. That is why we describe capacity as "aggregate node memory," never "unified VRAM," and why interconnect and topology design matter.
Does it run AI, scientific computing, or both?
Both — that is the point. The same modular architecture serves private AI inference and selected scientific workloads. Each has its own fit assessment and acceptance test, because a coding-agent deployment and a molecular-dynamics workflow succeed on different criteria.
Can it run my specific model or solver?
We do not assume compatibility. During the assessment we review your exact model, framework or solver against the supported stack, and where practical we validate a representative workload before you commit. CUDA-only or specialised software may point to a different platform — we will tell you if so.
Why not just use the cloud?
Cloud and institutional HPC remain valuable, and ArrayMeld is designed to complement them. Local ownership is strongest when you have sustained suitable workloads, need data and code to stay in your control, and benefit from predictable, unmetered access. The assessment compares the options honestly rather than assuming local is always better.
What does it cost?
Price is set per configuration after the workload assessment, because node count, memory, fabric, storage, support and installation all depend on what you need to run. Every quotation states inclusions, exclusions and delivery, with GST itemised separately — we keep public "from" prices off the site until a configuration is stable enough to hold to.1
Who am I actually contracting with?
ArrayMeld is the brand; during the launch phase it is operated by Holocyte Pty Ltd, which handles quotations, contracts, payments, delivery, warranty and support. A future Holocyte subsidiary, Cluster Solution Pty Ltd, is intended to operate the business as it matures — the ArrayMeld brand and your support relationship stay the same throughout.
Where are you located — and do you work on-site?
We are based in Adelaide, South Australia. On-site delivery, installation and technical service are available across Australia and New Zealand; for other regions we support the full build-and-validate workflow through remote collaboration and online meetings.
Start with the workload
Tell us what you need to run — not which computer you think you need.
Share your target model or simulation, your context and data requirements, and your budget path. We will return a scoped configuration and the validation plan we would sign off against.
- →A short fit call first — no obligation, no hard sell
- →We assess AI, scientific or mixed workloads
- →Honest guidance, including when cloud or HPC suits you better
Prefer email? arraymeld@holocyte.com · Please don't send confidential datasets or unpublished enabling details in a first enquiry.