by GNACODENothing leaves the machine
Models, prompts, files and results stay on the local drives. GNAXES needs no internet connection to work, and it can run fully air-gapped.
Point at the + for technical details
GNAXES · Local AI design workstation
Hardware-based iterative
AI-supported development
Real Time Artificial Intelligence Design with Intellectual Property Protection
GNAXES runs several AI models at once on one machine, coupled to hardware add-ons. Proceed rapidly from idea to prototype to final design. Using a local machine is the safest way to develop with AI.
Your designs, prompts, data and results stay local, so no data is handed over to an AI cloud provider.
128 GBGPU memory on four AI accelerators
2 TBsystem memory possible, for frontier-size models
4+models working in parallel on one design
Offlineruns with no cloud connection at all
01 / Why local
Your IP stays secured on the local machine
A cloud AI service sees every prompt, drawing and data file you send it, and its terms decide what happens to them. GNAXES runs the models on your own hardware, under your own roof. There is nothing to send to a provider, so no provider's data terms apply.
Your IP stays under your control
No cloud provider's terms apply to your prompts or results. The models are open-weight models under licences such as Apache 2.0, MIT and the NVIDIA Open Model License, each checked before it is supplied.
Never used for training
Your designs cannot end up inside someone else's model, because no one else ever receives them.
Controlled data stays controlled
Material under a non-disclosure agreement or export rules stays on your premises, inside your own access rules and networks.
An evidence trail for inventions
Every prompt, model version, tool run and file change can be logged with a time stamp in the open XES event-log format. The log shows who did what and when, which supports invention records and patent work.
A fixed cost
You pay for the machine, not per token. A design run that works through the night costs the same as a short question.
| Cloud AI service | GNAXES | |
|---|---|---|
| Where your data is processed | The provider's data centre | Your premises |
| Who can reach prompts and files | The provider, under its terms | The users you set up, on your own network |
| Works without internet | No | Yes, including air-gapped |
| Use of your data for training | Depends on the provider's terms | Never, it stays with you |
| Record of how a design came about | Whatever the service keeps for itself | Your own time-stamped event log |
| Cost | Per token or per seat, rising with use | Fixed hardware cost |
| Models | The provider's choice, changed at its pace | Open-weight models you choose and keep |
02 / Cooling
Cooling done the right way
Cooling that protects your hardware
GNAXES is a significant investment, and heat is what wears electronics out. Its lid pumps the processor's heat away with Peltier modules, so the processor runs colder than in a water-cooled PC, while the rest of the box stays close to room temperature. Colder parts age more slowly, so the machine stays stable for more years.
Processor die temperature
Full load, 350 W. Calculated for a 22 °C room with the same processor model for all three coolers.
Large air coolertower heatsink
95 °C limit
Water coolinggood block, 360 mm radiator
GNAXESPeltier lid, cold plate 15 °C
20 °C406080100 °C
1.8×slower thermal ageing of the processor than with good water cooling
10×slower than with a large air cooler
Three separate air paths
Front view. The processor's heat never passes through the box.
Colder than water cooling
Twelve Peltier modules actively pump heat out of the processor's cold plate, which sits at 15 °C instead of above room temperature. At full load the die stays about 8 K below a good water loop and about 35 K below a large air cooler (calculated).
Slower ageing
Wear-out in chips, such as electromigration and the breakdown of thin oxide layers, speeds up with temperature. The Arrhenius model used in semiconductor reliability halves the ageing rate for roughly every 10 K cooler at these temperatures.
Heat kept out of the box
The processor's heat leaves through the lid's own air path, and the four GPU blowers push theirs straight out of the back. The air around the board, memory, voltage regulators and drives stays about 3 K above room temperature at peak load (calculated).
Longer capacitor life
Capacitor makers rate life by temperature: an electrolytic capacitor's expected life doubles for every 10 K cooler, and a polymer capacitor's grows tenfold for every 20 K. Cool box air therefore lengthens the life of the board's power stages.
No pump, no coolant, no leaks
There is no liquid to pump, top up or leak. The fans are the only moving parts in the cooling, and each one is monitored and easy to replace.
Always above the dew point
The controller measures the room's humidity and holds the cold plate at 15 °C or 3 K above the dew point, whichever is higher, so the cold parts stay dry, even when the lid is opened. A cool die also keeps its boost clocks through long simulations.
Calculated with the GNAXES thermal model at 22 °C room temperature, using the same processor package model for all three coolers; water cooling means a good block on a 360 mm radiator, air cooling a large tower heatsink. Ageing ratio from the Arrhenius relation with an activation energy of 0.7 eV, the value chip makers use when the failure mechanism is not known (for example Analog Devices reliability reports); with 0.5 eV the advantage over water cooling is about 1.5×. Capacitor rules: Würth Elektronik SN008. Boost behaviour: AMD Precision Boost 2, which lists processor temperature among its limits.
03 / The machine
A server and a development instrument combined
GNAXES runs multiple models in parallel and can dissect a problem into the parameters that matter for optimisation. Coupled to the hardware add-ons, it measures the performance of prototypes, proposes optimisation routes and optimised designs, and steps the design iteratively towards a final product.
In this way GNAXES becomes an instrument that helps you find the critical parameters in your design and optimise them faster. This saves on your R&D budget and shortens the time to market.
- 01Peltier lid12 Peltier modules, nine heatsinks and nine fans cool the processor.
- 02Touch display14-inch screen with the running models, load, cooling and messages.
- 03Flush controlsPower button with a blue ring light and a status light, flush with the front.
- 04Side intakesFresh air for the box enters through grilles in both sides.
- 05Lid air pathRoom air enters through the top grilles and leaves through the lid's side and rear slots.
- 06Controller bayGNACODE's own controller boards sit behind the right end of the display.
- 07Deck and hingeThe 20 mm deck is the lid's backbone and turns on a hinge at the rear.
- 08Base plateA 15 mm machined base with air grilles and rubber feet.
04 / What it is made of
Computer configuration chosen to run many models in parallel
Large language models are limited by memory more than by arithmetic. GNAXES therefore combines four AI accelerators with 32 GB each, a processor with eight memory channels, and room for 2 TB of system memory.
With 256 GB of DDR5 as standard and up to 2 TB, tasks can be divided between the processor and the GPUs, so very large models can run. The frontier-size models need 512 GB or more.
AI accelerators
4 × AMD Radeon AI PRO R9700
32 GB GDDR6 · 640 GB/s each
64 compute units · 128 AI accelerators
95.7 TFLOPS FP16 · 766 TOPS INT4
(1,531 TOPS with sparsity)
64 compute units · 128 AI accelerators
95.7 TFLOPS FP16 · 766 TOPS INT4
(1,531 TOPS with sparsity)
128 GB of GPU memory in all. Each card can serve its own model, or several cards combine for one large model.
Processor
AMD Ryzen Threadripper PRO 9955WX
16 cores · 32 threads · up to 5.4 GHz
8-channel DDR5-6400 · 128 PCIe 5.0 lanes
8-channel DDR5-6400 · 128 PCIe 5.0 lanes
Runs simulations and tool chains, and the parts of frontier-size models that live in system memory.
System memory
256 GB DDR5 ECC, up to 2 TB
8 RDIMM slots, one per channel
Keeps the model library ready to swap, and holds the expert layers of the largest models in hybrid mode.
CPU cooling
Peltier lid, held above the dew point
12 modules · 6 regulated channels
cold plate 15 °C or dew point + 3 K
cold plate 15 °C or dew point + 3 K
The cold plate stays above the dew point, so it stays dry. At the full 350 W processor load the calculated die temperature stays near 60 °C, far below its 95 °C limit.
Motherboard
ASUS Pro WS WRX90E-SAGE SE
7 PCIe 5.0 slots (6 at x16, 1 at x8)
4 × M.2 PCIe 5.0 · 2 × 10 GbE · 2 × USB4 40 Gbps · AST2600 BMC
4 × M.2 PCIe 5.0 · 2 × 10 GbE · 2 × USB4 40 Gbps · AST2600 BMC
A workstation board with remote management on its own network port.
Storage
4 to 8 TB PCIe 5.0 NVMe
model library · projects · event log
Everything the machine knows and produces is stored on its own drives.
Controller and display
GNACODE GIOT CORE + 3 NODE-HMAX
Peltier current control · 9 fan channels
dew-point sensing · power governor
dew-point sensing · power governor
GNACODE's own control boards run the cooling and the 14-inch touch display, independent of the AI workload.
Power
2,200 W Platinum, or 1,600 W for 120 V
power profiles per outlet type
typical design load about 1.3 kW
typical design load about 1.3 kW
The controller keeps the machine inside the outlet's continuous rating, first through the cooling set point, then through processor and GPU limits.
Cabinet
Machined aluminium, 460 × 465 × 330 mm
6063 plates · 5052 covers · hinged lid
Solid base, deck and top plates with bent sheet covers. The lid lifts on a torsion-bar hinge for service.
Instruments
Connect over 10 GbE or USB4
data acquisition · cameras · test benches
Measuring instruments connect outside the box, so they can be changed at any time without opening it.
05 / Models and capacity
From fast agents to frontier-size models
GNAXES runs open-weight models: models whose weights are published, so they can run on your own machine, fully offline. The four GPUs hold several models at once or one large reasoning model, and system memory extends this to frontier-size models.
Example set-ups
GPU 132 GB
GPU 232 GB
GPU 332 GB
GPU 432 GB
System memory 256 GB to 2 TB DDR5Model library kept ready, simulations
| Model class | Examples (open weights) | Size at 4-bit | Runs on | Speed per user | Basis |
|---|---|---|---|---|---|
| Fast agentsMoE, 3 to 4 B active | Qwen3.6-35B-A3B Gemma 4 26B-A4B reads images and drawings | 16 to 25 GB | 1 GPU each, up to 4 at once | 75 to 110 tok/s | measured |
| Dense mid-size27 to 31 B | Qwen3.6-27B Gemma 4 31B | 16 to 22 GB | 1 GPU | 22 to 33 tok/sabout 45 with multi-token prediction | measured |
| Large reasoningMoE, 117 to 122 B | gpt-oss-120b Qwen3.5-122B-A10B Nemotron 3 Super 120B-A12B | 63 to 87 GB | 3 to 4 GPUs | 25 to 70 tok/s | estimated |
| Frontier, hybridMoE, 550 to 760 B | DeepSeek V4.1 Flash class GLM-5.3 class | 330 to 450 GB | 4 GPUs + 512 GB system memory | 4 to 11 tok/supper estimate, limited by the processor's memory bandwidth | estimated |
tok/s: tokens generated per second for one user; a token is about three quarters of an English word. With several requests at once the total output rises: two R9700 cards serving a fast agent delivered 140 to 290 tok/s at four requests, depending on the vLLM version (measured).
How long does a design task take?
A design task is a loop, not a single answer. The example is one PCB job: the first design, measurements on the prototype, analysis, sourcing, and the redesign that is ordered as the next prototype. The times are estimated AI time, from reading the input to the finished answer; simulation, CAD work and the measurements themselves come on top.
- 1First designabout 3.4 min AI time
From the requirements to a first schematic and layout, with test connectors at the points that will be measured.
Read a 60-page datasheet and draft the symbol, footprint table and pin mapabout 80 s estimatedCheck a 1,500-net schematic against a design-rule list and explain each findingabout 2.0 min estimated - 2Measure the prototypeabout 60 s AI time
The add-on records every test connector on the prototype, and the AI maps the data to the PCB components each signal belongs to.
Map 64 measured channels to the netlist and the components they belong toabout 60 s estimated - 3Optimisation and analysisabout 2.8 min AI time
Compare the measurements with the specification, find the critical parameters and rank the changes that move the board towards its target.
Analyse the deviations from the specification and propose ranked changesabout 2.8 min estimated - 4Source alternative componentsabout 70 s AI time
Find parts that meet the new values, with matching footprints and ratings.
Compare alternatives for 12 components from your parts library and datasheetsabout 70 s estimated - 5Redesign and order the next prototypeabout 22 min AI time
Apply the changes, review the whole design once more and send the files for the next prototype.
Update the schematic, list the layout changes and write the change logabout 3.2 min estimatedFull design review of the project folder with a frontier-class modelabout 19 min estimated - 6Wait for the new prototypeabout 2 weeks
Fabrication and delivery of the board. This is by far the longest step, and it is not AI time: GNAXES works on other jobs in the meantime.
AI time for one round (steps 2 to 5)27 minPrototype made and delivered2 weeks - Measure the new prototype and continue from step 2until the specification is met
Each round starts from real measurements, so the next proposal is based on how the board actually behaves, not only on the simulation.
about 31 minAI time in the first round, including the first design
about 27 minAI time in each further round, steps 2 to 5
about 2 weeksper round for the prototype, which sets the pace of the project
This is one job. GNAXES runs others in parallel while it waits for prototypes
The two weeks of waiting are free machine time. A second PCB job, or work from the other add-ons, runs alongside, for example:
Generate 24 lens-design variants as parameter sets for simulation
about 80 s estimated
Write a parametric CAD script for a part from a sketch and a requirement list
about 70 s estimated
Analyse 48 process runs and propose the next 8 experiments
about 1.8 min estimated
12 million
generated tokens in an 8-hour night, with fast agents serving four requests each on both GPU pairs (estimated at 200 tok/s per pair). Long design studies can run unattended while the office is closed.
Measured: community llama.cpp results on one Radeon AI PRO R9700 (llama.cpp discussion 21043, R9700 ROCm benchmarks) and vLLM results on a 4 × R9700 host (Level1Techs). Estimated: scaled from memory bandwidth with the efficiency of the measured models; for the hybrid class, from about 100 GB/s of usable system-memory bandwidth on the 16-core processor. Real speeds depend on the quantisation, the context length and the software version. Model names belong to their developers; check each model's licence for your use.
06 / GNAXES Design Server
Set up a design, not a prompt
A server sits in front of all the models and turns GNAXES into a design workbench. Each design becomes a workspace with its requirements, files, measurements and history, and repeatable design flows run on the right models and tools.
- Design workspacesRequirements, files, measurement data and decisions, kept together for each design.
- Design recipesRepeatable multi-step flows, for example from a datasheet to a symbol, a footprint and a review.
- Model router and schedulerSends each step to the model and GPU that suit it, and queues work across the four GPUs and the processor.
- Tool connectorsThe models call your engineering tools, such as electronics CAD, mechanical CAD, circuit simulation, Python and instrument drivers.
- Local knowledgeYour datasheets, standards and earlier designs, indexed on the machine and searched by the models.
- Familiar front end and an APIOpen WebUI in the browser for the team, and an OpenAI-compatible interface for your own scripts and CAD plug-ins.
- Access control and event logUsers, roles and closed workspaces for NDA projects; every step recorded in the XES event log.
Workspaces
Sensor front end, rev C
Collimator 25 mm
Gear housing, print 4
Process run 12
Recipes
Datasheet to footprint
Schematic review
Filter check
PCB recipe · fast agent + circuit toolsPCBoffline
You
Check the ADC input filter against the converter's datasheet and propose new values if needed.GNAXES
The present filter (1.0 kΩ, 3.3 nF) has its corner at 48 kHz. The datasheet in this workspace asks for 25 kHz or less at the chosen sample rate. Raising the resistor to 2.2 kΩ puts the corner at 21.9 kHz with the capacitor unchanged.
▶ circuit simulation · 3.2 s · settling within half an LSB: passed
| PART | NOW | PROPOSED |
|---|---|---|
| R12 | 1.0 kΩ | 2.2 kΩ |
| C7 | 3.3 nF | 3.3 nF |
Ask about this design, or run a recipe…
EVENT LOG (XES)
14:02:11 prompt · user eng01
14:02:11 model fast-agent · v3.6
14:02:13 read datasheet.pdf p. 14
14:02:16 tool circuit-sim
14:02:19 result passed
14:02:20 answer · 412 tokens
14:03:02 file schematic · rev 12
GPUS
GPU 1 fast agent · 41 %
GPU 2 to 4 reasoning · 63 %
14:02:11 prompt · user eng01
14:02:11 model fast-agent · v3.6
14:02:13 read datasheet.pdf p. 14
14:02:16 tool circuit-sim
14:02:19 result passed
14:02:20 answer · 412 tokens
14:03:02 file schematic · rev 12
GPUS
GPU 1 fast agent · 41 %
GPU 2 to 4 reasoning · 63 %
Illustration of a design workspace. The Design Server is in development and is configured for each customer; the content shown is an example.
07 / Measured Design add-ons
AI that learns from your measurements
A model is only as good as the data behind it. The Measured Design add-ons close the loop between AI proposals and the real world: the AI proposes, you build, instruments measure, and the measurements correct the models before the next proposal.
Each add-on brings the recipes, tool connectors and measurement interfaces for one field of design. They share the same loop, so a team can start with one field and add the next on the same machine.
We describe the add-ons in detail under a non-disclosure agreement. Ask for a briefing, and we send our NDA first.
Add-on
PCB design
From schematic to a board that meets its measured specification. The AI proposes component and layout changes, and measurements on real boards confirm or correct them.
Details under NDAAdd-on
Optics
Lens and illumination design checked against the measured light output, not only against the simulation.
Details under NDAAdd-on
Mechanics
Precise parts from ordinary fabrication, by measuring each iteration and compensating the next. Suited to mounts, gears and actuators.
Details under NDAAdd-on
Biological processes
Process development for reactions, syntheses and assays, where the AI plans the next runs from the measured results.
Details under NDA08 / Technical specifications
GNAXES at a glance
Compute
- Processor
- AMD Ryzen Threadripper PRO 9955WX, 16 cores, 32 threads, 4.5 GHz base, up to 5.4 GHz, 64 MB L3
- AI accelerators
- 4 × AMD Radeon AI PRO R9700 (RDNA 4), 64 compute units and 128 AI accelerators each
- GPU memory
- 4 × 32 GB GDDR6, 640 GB/s each (128 GB, 2.56 TB/s in all)
- Peak per card
- 95.7 TFLOPS FP16, 766 TOPS INT4 (1,531 TOPS with structured sparsity)
Memory and storage
- System memory
- 256 GB DDR5-6400 ECC RDIMM as standard, up to 2 TB in 8 slots
- Storage
- 4 to 8 TB PCIe 5.0 NVMe, 4 × M.2 slots
Expansion and connections
- Motherboard
- ASUS Pro WS WRX90E-SAGE SE, 7 PCIe 5.0 x16 slots (6 at x16, 1 at x8)
- Network
- 2 × 10 GbE (Intel X710-AT2), 1 GbE for remote management
- USB
- 2 × USB4 40 Gbps Type-C, USB 10 Gbps Type-A
- Management
- ASPEED AST2600 BMC, in-band and out-of-band
Software
- Operating system
- Ubuntu 26.04 LTS with AMD ROCm
- Inference
- vLLM and llama.cpp, hybrid CPU and GPU mode
- Front end
- GNAXES Design Server, Open WebUI, OpenAI-compatible API
- Models
- Selected open-weight models, licence-checked
Cooling and control
- CPU cooling
- 12 Laird UltraTEC Peltier modules on 6 regulated channels; cold plate at 15 °C or the dew point + 3 K
- Heat rejection
- 20 mm aluminium deck, 9 pin-fin heatsinks, 9 fans in the lid
- Controller
- GNACODE GIOT CORE and 3 × NODE-HMAX
- Display
- 14-inch touch display, 3840 × 1100
Power
- 230/240 V
- 2,200 W 80 PLUS Platinum power supply
- 120 V
- 1,600 W power supply, 100 to 240 V
- Power profiles
- Matched to the outlet: 230 V 10 A or 16 A, 240 V, 120 V 15 A or 20 A
Cabinet
- Construction
- CNC-machined 6063 aluminium plates, 1.5 mm 5052 aluminium covers, hinged lid
- Size
- 460 × 465 × 330 mm (W × D × H)
Planned standard configuration; specifications may change during development. GPU figures are AMD's peak values per card.
09 / Questions
Common questions
Does GNAXES need an internet connection?
No. Everything runs on the machine. It can be connected to your office network so the team can use it from their desks, or kept fully offline. Updates and new models arrive as packages you install yourself.
Which models come with it?
A selection of current open-weight models for coding, reasoning and reading drawings and documents, chosen for the work you do and checked for their licences. You can add or replace models at any time.
Can we use our own models or fine-tune them?
Yes. Any model in a common open format can be loaded, and models can be adapted to your own designs on the machine, so the training data stays with you too.
How many people can use it at once?
A design team can share one GNAXES. The Design Server queues and batches requests, and with several requests at once the total output is several times that of a single user.
What do the add-ons include?
Each add-on brings the recipes, tool connections and measurement interfaces for its field. We present them under a non-disclosure agreement.
10 / Contact
Book a talk
Pick a 30-minute slot that suits you. We will walk you through GNAXES with an example from your field.
30-minute video callOn Proton Meet: end-to-end encrypted, runs in your browser, no account needed.
Choose a time
- A walk-through on your kind of design task
- Our NDA first, if you want to discuss the add-ons
- A configuration matched to your models, team size and power supply
Prefer to write? Use the form, or email info@gnaxes.com