> 3 minutes: One visit, start to finish
> 40 seconds: To paint you into a year of your choice
> 6: AI models working as a single agent
> 1 box: About 55 W, and the size of a book
> 0 bytes: Sent to the cloud
> 12,000 years: Twenty stops, from the Ice Age to a futuristic starship
Step into a green telephone box, pick up the handset, dial a year, and talk. Forty seconds after you choose your picture, you are standing in it: in the snow of the Ice Age, on the new Eiffel Tower, in zero gravity on a starship, with your own face.
Three minutes after you pick it up, the printed postcard is in your hand. Every step of it, listening, talking, painting, checking and printing, runs on one computer the size of a book that has never been connected to the internet.
We built it to show what generative AI looks like when it stays in the room: fast, private, and on hardware you can buy. This page tells how it works, what it runs on, and why that matters beyond a game.
Scanned your postcard at electronica 26? Welcome. Your photo never left the box on the Avnet Silica stand, and the card in your hand is the only copy. What happened in those three minutes is detailed below!
What is it?
A telephone box that paints you into any year and prints the postcard in about three minutes.
First shown
electronica 2026, Munich, 10th-13th November, on the Avnet Silica stand in Avnet City.
What does it run on?
One ASRock Industrial AI BOX-A395 with an AMD Ryzen™ AI Max+ 395 and 128 GB of memory; six open AI models working as one agent
Where does your data go?
Nowhere. It's not connected to the cloud, you don't make an account, and there is no internet connection.
Six steps, one conversation: your three minutes in the booth
The time conductor, who introduces herself as Momo, guides every visit. You only talk, dial and listen. Behind her, a painter and an inspector work, and the big screen next to the booth shows the three of them talking.
1) Travel papers
- You dial a year, or just say one. A small camera above the screen takes your travel photo. A shutter clicks in the handset.
2) The call
- The machine lands on the nearest of twenty stops, from the Ice Age to a starship, and the conductor offers three pictures for that time.
3) Thinking
- From your choice and the conversation, the conductor writes a brief for the painter: who you are, what you do, where you stand.
4) Making
- The painter paints you into that year in about forty seconds, in four passes, blurry first and sharper each time. You watch it happen.
5) Checking
- The inspector looks at the result: is it still you, and does it show what was asked? Below seven out of ten, the painter tries again.
6) The photocard
- You choose a line, or dictate your own. The card prints while you say goodbye and is in the tray before the handset is back on the hook.
A real call to the year 1690: second-by-second breakdown
Timed from the log of a test call:
- 0:00 - Handset picked up. The time conductor answers and asks for a year.
- 0:20 - You dial 1, 6, 9, 0. The machine lands in Shakespeare's London.
- 0:25 - A shutter click in the handset: travel photo taken, face found.
- 0:35 - Three pictures appear on the screen. You say "writing a play with a quill pen".
- 0:40 - The conductor briefs the painter and starts chatting: a fact about the Globe, a word about the box.
- 1:25 - The picture is painted: four passes, blurry to sharp, watched live.
- 1:32 - The inspector checks it: likeness 9 out of 10, quill present. Approved.
- 1:35 - "Take the same again." The painter paints a second version.
- 2:33 - Both appear on the screen. "I prefer the bottom one."
- 2:38 - Three lines to choose from. "A new play by tonight. The ending is still a mystery. To me too." The printer starts.
- 2:40 - Handset put down. The printed card lands in the tray as you step out.
To the right/below: What the crowd sees. The 75-inch screen next to the booth shows the machine at work: the conductor, the painter, and the inspector talking to each other in plain sentences; the six models lighting up as they take their turn; and the chip's live figures: load, clock, watts, temperature, memory. The pictures of the day join the gallery at the top.
One chip does it all
What's in the box under the big screen?
The whole demo, both screens, the telephone, the painter and the voice, runs on one small industrial computer the size of a thick book: an ASRock Industrial AI BOX-A395. Inside it is one chip, the AMD Ryzen™ AI Max+ 395, with sixteen Zen 5 processor cores, Radeon™ 8060S graphics with forty compute units, and 128 GB of memory shared between the processor and the graphics.
That shared memory is the trick. A 35-billion-parameter language model, a 20-billion-parameter image model, a vision model and a speech model all sit in memory at the same time, about 60 GB of them, so nothing is loaded while you wait. The processor cores and the graphics cores read the same bytes: no copies, no bus to cross. On a conventional PC, that would take a separate graphics card with more memory than most of them have.
| Computer |
ASRock Industrial AI BOX-A395, compact industrial mini PC; no network cable while it runs |
| Processor |
AMD Ryzen™ AI Max+ 395 ("Strix Halo"): 16 Zen 5 cores, up to 5.1 GHz |
| Graphics |
AMD Radeon™ 8060S, 40 compute units, on the same chip |
| Memory |
128 GB LPDDR5X-8000 on a 256-bit bus, shared by processor and graphics; about 60 GB taken by the models |
| Power |
About 55 watts for the whole chip while painting; quiet enough to stand next to a telephone box |
| Software |
Ubuntu Linux, llama.cpp (Vulkan) and ComfyUI (ROCm) serving the models |

The ASRock Industrial AI BOX-A395, a silver aluminium mini PC, next to an old red rotary telephone on a wooden table.

Your photo card, 10 x 15 cm: the picture, the year you dialled, your line, and the QR code that brought you here. Example with a visitor who does not exist.
Your photo card, the only copy
The picture fills the card. The band at the bottom carries "Greetings from", and the year you dialled, the line you chose or dictated, and a small QR code to this page. It is printed on a small dye-sublimation photo printer next to the box, in about twenty seconds.
- There is no download. The picture was never online, so there is nothing to fetch. The card is the only copy. Take a photo of it if you want to share it.
- Your face, another time. The painter keeps your face and changes everything else: the clothes, the place, the year. Glasses stay if you wore them.
- Sixty pictures, twenty stops. Every stop has three pictures, and every picture was tested on five different faces before the show.
- A small painter can err. A hand, a hat or something in the background is sometimes not quite right. That is the price of a painter that fits in a box on a table.
Meet the chip: AMD Ryzen™ AI Max+ 395
AMD's largest processor for mini PCs and thin workstations, introduced in 2025 under the code name "Strix Halo". It puts a sixteen-core processor, a workstation-class graphics unit, a neural processing unit and the memory controller for 128 GB on one package. That is why one small box can do what used to take a tower with a big graphics card.
Memory is the gate, not speed
A 22 GB language model, a 39 GB image model with its 9 GB text encoder, a vision model and a speech model: about 60 GB in all. No consumer graphics card holds that. On this chip they all sit in the one memory, resident, and the painter and the conductor work side by side without swapping anything in or out.
Bandwidth is the budget
All units share about 256 GB/s. When the painter is busy, the conductor slows from about sixty to under twenty words a second, and the big screen shows that number honestly. That is the trade-off of unified memory: one pool, one budget, no copying.
Two software stacks, one GPU
The language and vision models run through llama.cpp on Vulkan, the painter through PyTorch and ComfyUI on ROCm, at the same time, on the same graphics unit. Open drivers, open tools, nothing custom.
Processor cores
16 Zen 5 cores, 32 threads, up to 5.1 GHz, 64 MB L3 cache, 4 nm
Graphics
AMD Radeon™ 8060S: 40 RDNA 3.5 compute units, up to 2.9 GHz, on the same silicon
Neural processing unit
AMD XDNA™ 2, up to 50 TOPS (not used in this demo: everything runs on the graphics and processor cores; the NPU is headroom)
Memory
Up to 128 GB LPDDR5X-8000 on a 256-bit bus, about 256 GB/s, shared by all three units; on Linux up to 110 GB can be handed to the graphics unit
Power envelope
Configurable from 45 W to 120 W; this box runs at about 55 W while painting and stays cool and quiet
The device: ASRock Industrial AI BOX-A395
An industrial mini PC built around the chip, about the size of a thick book: 200 x 100 x 232 mm, 2.8 kg. The 128 GB of memory is soldered on the board; there are two M.2 slots for SSDs, two USB4 ports, 10 Gigabit and 2.5 Gigabit Ethernet, Wi-Fi 7 and a TPM 2.0 chip. In the Time Machine, none of the network ports is connected: the box runs the whole demo on its own.
It is a standard product, available through Avnet Silica, with Windows 11 or Linux. The Time Machine runs Ubuntu Linux.
Generative AI without the cloud
Why did we build it this way?
The Time Machine is a game, but the point behind it is serious. The same box, with the same open models, could stand on a desk in a factory, a hospital or a shop, and do real work with data that may never leave the building.
Size and weight
200 x 100 x 232 mm, 2.8 kg
Memory and storage
128 GB LPDDR5X onboard, two M.2 SSD slots
Connections
2 x USB4, 10 GbE + 2.5 GbE, Wi-Fi 7, Bluetooth 5.4, HDMI and DisplayPort
In this demo
Two screens, a webcam, a photo printer, a telephone and a USB sound card; no network
Inside the machine: six models, one agent
"Agentic" means that no single model does the job. A language model plans the picture from your conversation, hands the brief to the image model, lets a vision model check the result, and sends the painter back when the likeness is not good enough. On the big screen next to the booth, that exchange runs live, in plain sentences.
- The "Time Conductor" (Plans and talks): Listens to you, chooses the stop, offers three pictures, writes the brief for the painter, and keeps you company while it works.
"Painter, the choice is number 1: running away from the mammoths in the Ice Age. Version 1 from the travel photo, please."
- The "Painter" ("Paints" you in): Takes your travel photo and the brief, and paints you into that year in four passes on the graphics part of the chip, in about forty seconds.
"Version 1 ready in 39 seconds. Inspector, your turn: is it the same person, and are they running away from mammoths?"
- The "Inspector" (Checks the result): Looks at the picture. Is it still the same person (scored out of ten)? And does it show what was asked? Below seven, the painter gets another try.
"Checked: likeness 9 out of 10, and yes, it shows what was asked. Looks good to me, carry on."
| PART | MODEL | SIZE | RUNS ON | WHAT IT DOES |
|---|---|---|---|---|
| Ears | Whisper large-v3-turbo | 809 M parameters | Radeon graphics, whisper.cpp | Hears you through the handset, about half a second per sentence |
| Time conductor | Qwen3.6 35B-A3B (mixture of experts) | 35 B parameters, 3 B active per word, 4-bit, 22 GB | Radeon graphics, llama.cpp | The conversation, the brief for the painter, the lines on your card; about fifty words a second |
| Voice | Kokoro | 82 M parameters | Zen 5 cores | Speaks into the handset; the avatar in the booth moves its lips with the words |
| Camera | YuNet face detector | 230 kB | Zen 5 cores | Finds a face in the travel photo, and sees whether somebody is in the booth |
| Painter | Qwen-Image-Edit 2511 with a four-step Lightning adapter | 20 B parameters | Radeon graphics, ComfyUI | Paints you into the chosen year in four passes, about forty seconds per picture |
| Inspector | Qwen3.5 VL | 4 B parameters | Radeon graphics, llama.cpp | Scores the likeness out of ten and checks that the picture shows what was asked |
All six are open models under open licences (Apache 2.0 and MIT), served by two open-source programs, llama.cpp and ComfyUI. Numbers are measured on the stand box; speeds drop while the painter is busy because the models share the graphics cores.
From the ice age to a starship
Any year works. Dial 0, and you land in the ice age; dial 2999, and you are on a starship. The machine lands at the nearest of twenty stops, each with three pictures to choose from.
Year 100 - The Roman Empire: A racing chariot
Year 900 - The Viking Age: Rowing into the fjord
Year 1250 - The Age of the Castles: The tournament
Year 1500 - Florence: Testing a flying machine
Year 1600 - London: Shakespeare's Globe
Year 1889 - Paris: The waiters race
Year 1925 - The Jazz Years: Charleston
Year 1969 - The Space Race: Launch
Year 2050 - Munich: Boarding a flying taxi
Year 2100 - Moon Base: A giant leap
Year 2300 - Mars: Skiing down a red dune
Year 2500 - The Starship: Zero gravity
Where can I see it?
First shown in Munich, built to travel
The Time Machine made its first appearance at electronica 2026 in Munich, on the Avnet Silica stand within Avnet City. The next dates and locations to see the Time Machine will appear here. If you would like to have the Time Machine at one of your own events, or want to talk about what a local AI setup like this could do for your product, get in touch.
Last seen, and where next
electronica 2026
Munich, 10th-13th November, Avnet Silica stand at Avnet City
Next
Watch this space!
Frequently asked questions
| What happened to my photo? |
It stayed in the box on the stand, together with your pictures and the words from your conversation. The team looks at them only to see how the demo performed. They are not shared with anyone, and they are deleted after exhibitions. |
| Can I get a digital copy of my photo card? |
Unfortunately not. The picture was never technically online, so there is nothing to download. The printed photo card is the only copy. Taking a photo of the photo card with a smartphone might work well. |
| Why does a detail in my picture look wrong? |
The 'painter' is a 20-billion-parameter model that fits in a box on a table, not a cloud service with thousands of chips. It keeps your face well and gets the scene right nine times out of ten, but a hand and a hat or something in the background are sometimes slightly incorrect. The 'inspector' catches the worst cases and asks for a second try. |
| Who is Momo? |
Momo is the 'conductor', the voice in the handset. She is named after the girl in Michael Ende's book who gives people their time back. She is a program running on the box, not a person: she has no memory of earlier visitors and no access to the internet. |
| Is anything running in the cloud/online? |
Nothing. The box has no network cable while it runs. Speech recognition, the conversation, the painting, the check, and the printing all run on the one AMD Ryzen AI Max+ 395 chip in the box. |
| Could I run something like this myself? |
Yes. The box is a standard product from ASRock Industrial, the models are open models you can download, and the serving software is open source. Talk to Avnet Silica about the hardware and what a local AI setup could do for your application. |
| Where can I see the Time Machine? |
It was first revealed at electronica 2026 in Munich, from the 10th of November to the 13th of November, on the Avnet Silica stand. Where it goes next will be announced on this page. It fits in a van: one telephone box, one screen, one small box. |
| Where did the telephone box come from? |
It's the same green box that visitors used at electronica 2024 to phone a chatbot. This year it became a time machine. The old rotary telephone has a small Arduino board inside that reads the dial and hook. |
Looking for a non-cloud-reliant Edge AI solution for your application?
Avnet Silica supports engineers across Europe with AMD processors, industrial computers and design help, from the first idea to production.
Bring AI acceleration into embedded systems that demand responsive, predictable performance.
AMD Ryzen™ AI Embedded processors integrate CPU, GPU, and NPU processing in a single SoC, providing a flexible foundation for applications that combine real-time compute, graphics, AI, and control. Long-life platforms and support for deterministic, low-latency operation make the portfolio suited to demanding, always-on edge systems.




