- Helped establish the concept of p-bits (Nature)
- Leads the OPUS Lab at UC Santa Barbara
- Bell Labs Prize · NSF CAREER · ONR Young Investigator
Where noise becomes signal
Life afterGPUs.
AI inference is a memory-reading problem, not a compute problem. Flucta builds probabilistic inference, algorithms co-designed with silicon that read up to 7.5× less memory today, on a path to 100× intelligence per watt.
250:1
Energy to read a byte vs. compute on it
7.5×
Less attention memory traffic, no retraining
<5%
Accuracy delta from the dense baseline
100×
Target intelligence per watt with the SPU
01. The thesis
The bottleneck isn’t compute.
It’s reading memory.
GPUs are an extraordinary machine for training large AI models. Inference is a different problem. On a modern accelerator, reading one byte from HBM costs roughly 25 picojoules; computing on it costs about 0.1. As context windows march toward a million tokens, the KV cache dominates both the energy and the latency of every answer.
Memory
HBM · hundreds of GB
⚠ Bottleneck
Constricted memory interface
Moving data cost
25 pJ / byte
Compute cost
0.1 pJ / byte · 250× less
The industry’s reflex is to widen the bus and stack more memory. That makes the straw thicker. It doesn’t change what’s in the cup.
02. The difference
Same answer.A fraction of the reads.
Read
Reads / token5 / 24
Legacy · Dense Read
Reads everything, every token
VS
Read
Reads / token4 / 24
Flucta · Sampled Read
Samples signal
03. The roadmap
The SDK is the wedge. The chip is the moat.
The 7.5× is proven on today's GPUs, but they were built to stream dense tensors, not to sample sparse ones. Capturing the full prize needs new silicon: the Stochastic Processing Unit, a memory hierarchy designed around read-less inference. Software validates the math. Hardware builds the moat.
Where we’re going
Today
SANTA + Gallop SDK
Algorithm validated · FPGA prototype underway
2027
2D multi-chip SPU
28 nm · custom memory hierarchy
2029+
3D-stacked SPU
Near-memory compute · datacenter
04. The people
Physicists and chip architects, building from first principles.
Probabilistic computing and 3D-integrated silicon are the team's home turf, recognized by NSF CAREER Awards and best-paper honors, and proven in fabricated chips.
The team
Founding teamResearch internsAdvisors* Consulting
Silicon Architecture
Kartik Prabhu
Ph.D., Stanford
ex-Meta · lead author, MINOTAUR, the 8-chip transformer in silicon
Algorithm / Kernel
Shuvro Chowdhury
Ph.D., UC Santa Barbara
Algorithm / Kernel
Corentin Delacour*
Ph.D., Univ. of Montpellier
Algorithm / Kernel
Kemal Selçuk
Ph.D., UC Santa Barbara
Algorithm / Kernel
Nihal Sanjay Singh
Ph.D., UC Santa Barbara
KB
Founding team
Kai Bartolone
ZH
IncomingFounding team
Zengxiao He
RL
IncomingFounding team
Rubens Lacouture
KC
Hardware / FPGA
Kevin Callahan-Coray
UC Santa Barbara
AL
Hardware / FPGA
Amrith Lotlikar
Stanford
TW
Hardware / FPGA
Tong Wu
Ph.D. student
SM
Advisor
Subhasish Mitra
Stanford
Professor of EE & Computer Science · Robust Systems Group
SL
Advisor
Suk Hwan Lim
ex-Google
Former Corporate EVP, Samsung Semiconductor
05. Backed by
Funded by investors
who back foundational technology.
Lead community
South Park Commons
The community and fund for people building at the frontier, backing technical founders from the earliest, pre-idea stage.
Venture
Moxxie Ventures
An early-stage fund investing in founders building technology that makes the future work better.
Deep tech
Humba Ventures
Susa's deep-tech sister fund, backing hard-tech founders at the earliest stage.
Seed
Pear VC
Seed-stage firm partnering with technical founders from day one.
AI · Data
Defined
Thesis-driven fund backing founders shaping the AI frontier.
faqs
Frequently Asked Questions
Flucta makes LLM inference read-less. Instead of streaming the entire KV cache from HBM on every token, our SANTA + Gallop SDK reads only the handful of positions that carry the signal, cutting attention memory traffic by 7.5× and, with it, the energy and latency of each answer.
On a modern accelerator, reading one byte from HBM costs roughly 25 picojoules while computing on it costs about 0.1, a 250:1 gap. As context windows push toward a million tokens, the KV cache dominates both the energy and the latency of every answer, and widening the memory bus doesn't change that.
Attention is a weighted average over the KV cache, and those weights are sharply peaked: a small fraction of positions carry almost all of the answer. Read-less inference reads only those positions, using importance sampling to form an unbiased estimate of exact attention instead of scanning everything.
The redundant work is the reading, not the answer. SANTA's sampling is unbiased, it estimates the same attention the dense model would compute, and Gallop reuses state that changes slowly between tokens. In practice this holds accuracy within <5% of the dense baseline.
SANTA (Stochastic Attention) samples the few KV positions that carry the signal instead of reading the whole cache, about 2.5× less traffic. Gallop does one full “heavy” scan, caches the high-yield subset on chip, then reuses it across the next tokens, about 3×. They compose to 7.5×.
No. Flucta runs on the same model weights with no retraining or fine-tuning. It changes how attention reads memory at inference time, not what the model learned.
The 7.5× is proven on today's GPUs. The SDK works now. The Stochastic Processing Unit (SPU) is the roadmap: GPUs were built to stream dense tensors, not to sample sparse ones, so purpose-built silicon and a memory hierarchy designed around read-less inference capture far more of the prize. The software is the wedge; the chip is the moat.
If HBM cost per token is a real line on your P&L, or you're an investor, researcher, or engineer who wants in, email founders@flucta.ai to request the technical brief. We're also hiring kernel & GPU systems engineers in the Bay Area to build the SPU.