The vision
Silicon,
within reach.
A wave doesn't need permission to carry energy — it needs a medium. Taranga exists to make graphics-and-inference silicon a medium anyone can work in: designs that are open to the last gate, hardware that costs pocket money, and arithmetic so small it prices parallelism like wire.
And one discipline above all: hardware claims wait for hardware. We publish the measurements that go against us at the same size as the ones that don't.
What we're up against
The most important machines became unreadable.
None of this is a complaint about the GPU giants — their silicon is a wonder, and our own benchmarks print its victories at full size. It is an observation about access: the gap between using compute and being allowed to understand and make it has never been wider.
Stochastic computing doesn't close that gap by beating binary — it closes it by being small enough to hold. A multiplier you can point at. A pipeline a student can read in an afternoon. A GPU that fits a $25 FPGA and a weekend.
Silicon became a fortress
GPU design concentrated into a handful of companies, behind proprietary toolchains, NDAs, and reticle-scale economics. A student can read every line of a compiler; almost no one alive has read every line of the chip it runs on.
Entry is priced in millions
Between EDA licenses, IP fees, and mask costs, the ticket to making real graphics silicon excludes nearly every classroom, lab, and small country on Earth. The knowledge isn't secret — the path is gated.
Claims outran measurements
Hardware marketing routinely promises what silicon hasn't shown. Our own first draft did it too — the reviewers caught it, and the correction became our founding discipline.
What we hold to be true
When you let probability do the arithmetic, the hardware almost disappears — and what little remains is small enough to be owned, read, taught, and trusted by anyone. That is worth more than benchmarks to the people benchmarks forgot.
The discipline beneath it: the wave pays for its one-gate multiply in cycles and joules, and our evidence page prints that bill in full. A vision that hides its costs is a pitch.
The long arc
From ripple to ocean.
Four swells, each funding the next. The first is shipped and reproducible; the second is in progress and says so; the last is a research bet we state plainly as one.
The peer-revised StochastiCore design is open and reproducible: every cell count regenerates from one Yosys flow, the Signal Lab runs the physics in any browser, and a 68-page guide takes a beginner from a bag of parts to a glowing VGA monitor for $80–150. The ripple is the proof that the wave is real.
Physical validation: timing closure on real FPGAs, measured power, on-device frame rates. The measurements go on this site whatever they say — and only then does the Dev Kit ship as a product: kitted parts, printed guide, a GPU you assemble, probe, and own.
The SC fabric as licensable IP for the corners where it honestly wins — instrument-cluster displays, radiation-tolerant space and industrial controllers, voltage-overscaled wearables — and Taranga-1: a small ASIC tape-out on a multi-project shuttle. A chip measured in square millimeters and rupees, not reticles and rounds.
Beyond graphics: can deterministic stochastic fabrics carry inference, sensor fusion, and in-memory compute on the same one-gate economics? The deterministic generator already computes exact products — the open question is how far the niche widens before the L-cycle price closes it. We may find the ocean is a lake. We will publish the soundings either way.
If we are right
What the world looks like when this works.
Not a frames-per-second chart — a change in who gets to make silicon, and where it survives.
Every student builds a working GPU in a semester — wires the DAC, flashes the open RTL, watches their first triangle land. Graphics stops being a black box the day you've soldered one.
Radiation flips bits; the stochastic fabric shrugs — every upset is worth 1/L, never the most significant bit. The display computer degrades like film grain instead of dying like a register.
A ventilator's pressure gradient renders smooth on 3-bit color with no dithering hardware at all, in a corner of the cheapest FPGA on the bill of materials.
A sovereign silicon curriculum starts from a GPU whose every gate is public, whose toolchain is free, and whose claims were stress-tested by four anonymous reviewers before any minister saw a slide.
Every vignette extrapolates a measured property — gate counts, the fault crossover, the dithering theorem, the $80–150 BOM — never a capability the record doesn't hold.
The vow that makes it credible
Hardware claims wait for hardware.
Our energy numbers favor our competitors and we publish them. Our throughput numbers favor our competitors and we publish them. Our timing closure is unfinished and the hardware page says "in progress" instead of a promise.
This is not modesty — it is strategy. In silicon, trust compounds slower than transistors and evaporates faster. A company whose smallest claims always hold is the only kind that gets to make big ones later.