Farce Systems jhickson.us
F.A.R.C.E. Flash

Press

Press kit — F.A.R.C.E. Flash

TierPriceCap
Reader$1uncapped
Single Lane$34100
Founder's Array (8-lane)$22440
Datacenter (24-lane)$64820
Full Array (40-lane)$1,07410
Bare Array (40-lane, media not included)$31525
Validation Rig (40-lane, USB)$1,4505
Host Interface Pack$65420
Endurance Pack$15950
Thermal Pack$2050

Boilerplate, 50 words

Farce Systems builds reference kits for memory-constrained machine learning. Its design, the F.A.R.C.E. Flash — Flash Array for Resource-Constrained Environments — attaches to a Linux computer the owner already has and presents 1.28 TB of system-addressable memory from a 40-lane array of consumer MicroSD cards. That is ten times the capacity of an NVIDIA DGX Spark, which is a complete machine; the kit is a peripheral and the host is not included in its price.

(49 words)

Boilerplate, 100 words

Farce Systems builds open reference kits for memory-constrained machine learning. Its first design, the F.A.R.C.E. Flash — Flash Array for Resource-Constrained Environments — attaches to a computer the owner already has and presents 1.28 TB of system-addressable memory from a 40-lane array — ten times the capacity of an NVIDIA DGX Spark, which is a complete machine where this is a peripheral.

That capacity comes from forty consumer MicroSD cards. Farce Systems ships reference kits, not products: every component is an off-the-shelf part with a public datasheet, and the PCB design, assembly guide, and OpenFARCE software are open source. Before designing the array, Farce Systems commissioned an independent feasibility analysis of a 128 GB configuration and published it unedited, conclusion included.

(100 words)

Four headlines

  1. Farce Systems Announces 1.28 TB Developer Kit, Ten Times the Memory of NVIDIA's DGX Spark
  2. This startup commissioned a study of its own product. The study says it won't work. They published it anyway.
  3. Forty cards, 240 traces, and a 64-day lifespan: inside the F.A.R.C.E. Flash
  4. Ten times the capacity, a twentieth of the speed: the F.A.R.C.E. Flash arrives

Fact sheet

Product F.A.R.C.E. Flash Developer Kit
Expansion Flash Array for Resource-Constrained Environments
Company Farce Systems
Tagline Memory, at scale, at retail.
Secondary 1.28 terabytes of system-addressable memory. Forty consumer MicroSD cards. One hand-soldered perfboard.
Software OpenFARCE (open source, pre-alpha, not included in kits)
Community Array Builders

Architecture

Fact Value
Headline capacity, system-addressable 1.28 TB
Capacity against the NVIDIA DGX Spark 10x
Working set analysed by the Feasibility Report 128 GB
Array size, reference design 40 MicroSD slots
Card capacity shipped in kits 32 GB, UHS-I, Class 10
Raw array capacity, 40 x 32 GB 1.28 TB
SDIO signals per card 6 (CLK, CMD, DAT0-3)
Total high-speed traces, 40 cards 240
Trace spacing rule 3W rule
Peak 3.3 V rail draw, 40 cards writing ~6 A / ~20 W
Per-card write current 100-150 mA

Performance

Fact Value
Per-card UHS-I ceiling 104 MB/s
Aggregate sequential read, 40 cards, ideal ~4.1 GB/s
Projected decode speed 1.5-3.0 tok/s
Projected TTFT at 32K context 45-60 s
70B model at 1.58-bit ~13.8 GB
70B model at 8-bit ~70 GB
KV cache, 128K context, FP8 ~20 GB

Thermal and endurance

Fact Value
Thermal throttle threshold 70-85 °C
Throttled speed <10 MB/s
Consumer card P/E cycles 1,000-3,000
pSLC card P/E cycles up to 60,000
Assumed WAF under LLM load 5.0 conservative, 10.0+ realistic
Array lifespan under continuous load ~64 days

Reference: NVIDIA DGX Spark

Fact Value
Price $4,700
Memory 128 GB unified
Bandwidth >500 GB/s
70B 8-bit decode 40-60 tok/s

The NVIDIA DGX Spark is a product of NVIDIA Corporation. Farce Systems has no affiliation with NVIDIA.

Kits

No kit contains a fabricated PCB, an enclosure, assembled hardware, or software. Full contents lists are published per tier.

Report

Feasibility Analysis of MicroSD-Backed LLM Inference Architectures. Commissioned by Farce Systems, published unedited, including its conclusion that the architecture is impractical. Prepared with AI research tooling; reviewed for consistency, not corrected.

Full text, with section anchors, at /farce-flash/report/.

Quotable, verbatim:

a monument to the lengths to which engineers will go to avoid purchasing enterprise hardware

It is functionally the equivalent of watching a human type.

machine learning can indeed run in a flash

Assets

Wordmark, icon, board renders, architecture diagrams, and 1200x630 social cards: /farce-flash/press/

Contact

hello@jhickson.us