Frequently asked
Questions
Will it work?
We commissioned an independent analysis to answer exactly that. It evaluates three configurations. Two are marked Impossible. One — 1.58-bit weights with the array carrying KV cache only — is marked Theoretically Possible. We published the whole thing unedited, conclusion included. Read it before you back anything.
Is it really 1.28 TB?
Forty cards, thirty-two gigabytes each. Yes.
Is it really memory?
The kernel reports it as available memory. See the disclaimer for what that means.
How is this different from a big swap file?
It is a big swap file. It is also forty cards on one board.
Ten times the memory — so is it ten times faster?
No. Please see the performance table.
What's actually in the box?
Parts. Breakout boards, perfboard, header strips, jumper wire, ceramic capacitors, a microcontroller dev board, and — at most tiers — MicroSD cards, a heatsink, fans, thermal pads, and sandpaper. Each tier's contents list on the kit page is complete and exact. Nothing arrives assembled.
Is there a PCB in the kit?
No. The reference PCB design is open and published, but no kit contains a fabricated board. You build on perfboard, from the reference design and the assembly card. If you want the 8-layer board, take the files to a fab house yourself.
Why MicroSD?
Per gigabyte, it is the cheapest solid-state storage that exists at retail scale. Every other tier of the memory hierarchy is priced against datacenter demand. MicroSD is priced against dashcams and handheld consoles, and it is stocked everywhere on earth.
Why not just buy a DGX Spark?
If you can, buy one. It is an excellent machine and it will outperform this architecture by every measure we publish. It also costs $4,700. The F.A.R.C.E. Flash exists for people for whom that sentence ends the conversation.
How fast is it?
Our analysis projects 1.5 to 3.0 tokens per second on a ternary 70B model, with a first token at 45 to 60 seconds on a 32K context. The DGX Spark does 40 to 60 tokens per second. We describe ours as human-readable generation speed.
How long will the cards last?
Under continuous inference load, the analysis projects roughly 64 days before an array exhausts its program/erase budget. This is a consumable media architecture. Cards are replaced, not repaired. The Endurance Pack exists for that reason.
Do I really have to sand the cards?
If you want sustained throughput, yes. Plastic is an insulator, and a throttled card drops from 104 MB/s to under 10 MB/s and stalls every other lane waiting on it. The direct-die procedure is documented and the abrasives are included. It voids the card's warranty.
What software runs on it?
OpenFARCE, when OpenFARCE exists. It is a user-space asynchronous I/O layer that bypasses the kernel VFS. It is pre-alpha, it is open source, and it does not ship with any kit. No software is included in anything you buy.
Can I run Llama 3.3 70B?
Not on arrival, and not on anything in the box. The architecture targets a ternary-quantized 70B model at roughly 13.8 GB of weights, held in host DRAM, with the array carrying KV cache. That is a target, not a demonstration. No performance is promised.
Is this a real company?
Farce Systems is a project, not a corporation. It is not registered, incorporated, or capitalized. One person sources the parts and packs the boxes.
Can I get a refund?
Yes, in full, any time before your kit ships. After it ships, we cover anything that arrives missing, wrong, or dead — send a photo and we replace it or refund it. Parts are ordered only after payment clears, so nothing is in motion until then.
Why 32 GB cards if the Feasibility Report analyses 128 GB?
Those numbers describe different things. 128 GB is the working set the report analyses — weights plus KV cache. The kit contains 32 GB cards, and forty of them is 1.28 TB of raw capacity. Card capacity was never the constraint; bandwidth and endurance were.
Why are tiers capped?
Our media allocation is finite, and so is the parts budget behind it. Plainly: one person fronts the money for every box, and the caps are the number of boxes that person can afford to be wrong about.
Why doesn't the Bare Array include cards?
Some builders have media preferences, and card prices move faster than we can reprice a tier. Plainly: it is the cheap way in. You get every part of the reference design except the forty cards, and you source those wherever they are cheapest that week.
Where do you ship?
United States, Canada, EU and UK, and most of the rest of the world. Shipping is quoted separately at checkout by zone and charged at cost. We do not absorb it into the tier price, and we do not mark it up.