Brain-inspired computing

We put the mechanisms
of the human brain
into today's silicon.

Selective memory for datacenters. Associative recall for knowledge. Instant reflexes for machines.

Three engines  ·  one brain  ·  running on silicon that ships today

Engine 01 — Remember selectively

Zero storage I/O during decode.

KV-cache residency policy · idle-driven demotion · V-only 4-bit · NIXL-native · endurance-aware QLC write-shaping

Inference engines run the model. NVIDIA's transport layer — NIXL / KVBM, RDMA with CPU bypass — lays the road. XHBM is the policy layer on top: it decides what stays resident, when to move it, and how to demote it.

The discipline is what makes it safe. XHBM never touches the decode path. It acts only inside idle windows, where the GPU is already waiting. It attaches to the NIXL standard with no code changes on the adopter's side, and it is fully reversible — switch it off and the stack is exactly what it was.

Idle-driven demotion — decode path untouched
idle windowidle windowtime →GPU decodetoken loopstorage I/OdemotionV-only 4-bit demote → SSDwrite-shaped for QLC (CSAL)0 bytes0 bytes0 bytes
While tokens are being generated, storage stays completely silent (measured on B200, 32K–256K context). Demotion happens only inside idle windows — so token latency never wobbles.
0 bytes
decode-path I/O · B200 measured
−37.5 %
KV bytes · V-only compression
184 GB
peak @ 256K context
Industry impact
The same HBM, serving twice the users.

HBM is the most expensive and most scarce resource in the AI datacenter. XHBM lets the same HBM serve roughly twice the concurrent users without adding a single GPU. At industry scale, that is a software layer standing in for billions of dollars of memory and GPU buildout — capacity bought back from hardware that is already installed.

Measured — same H100, same SSD
Baseline · vLLM stock
4 users → out of memory
×
UmpaRumpa · same hardware
8 users · stable
Latency unchanged. Output preserved bit-exact.
2 ×
Why no one solved this before
Approach
Measured outcome
Result
Naive SSD offload
20.04× slowdown
FAIL
KIVI / KVQuant
Attention output corrupted
FAIL
SAGE-KV / H2O
+5–10% GPU overhead
FAIL
Always prefetch
5× slowdown
FAIL
UmpaRumpa
2× capacity · bit-exact · 0 GPU cost
PASS
All prior approaches added GPU work to relieve memory pressure — and broke either speed or accuracy. We don't touch the GPU.
Works with — NVIDIA NIXL / Dynamo · Solidigm SSD + CSAL
Engine 02 — Recall associatively

Zero hallucination, 10.45 milliseconds.

parallel cued retrieval · associative memory · one-shot readout · bit-exact decode preservation

Hear “red round fruit” — and “apple” arrives before the sentence finishes. The brain has not searched a database or walked a graph. One cue radiates across everything stored, and only the memories that match fire back.

Recall is that mechanism as an algorithm. The mathematics turns out to be elegant: retrieved = softmax(β · Q · KT) · V — the same operation inside every transformer attention head, mapped onto the right memory hierarchy. HBM becomes the hippocampus, holding the cue. NVMe becomes the cortex, holding the patterns. A query no longer walks the memory bus N times; it fires across the whole content store at once.

Output is preserved bit-exact — twice the concurrent users on the same hardware, and not a single bit of the answer changes.

Multi-hop traversal vs one-shot parallel readout
graph ANN (HNSW class) — sequential hopsqueryhit @ hop 4latency ∝ hops × cache missesparallel cued recall — one shotcue← bestlatency = one parallel readout · 10.45 ms measured
Graph traversal slows and jitters with every hop, because each hop must finish before the next can start. Parallel cued recall touches all of memory at once — one readout, even at six-thousand-book scale.
Two layers, one parallel mechanism
HIPPOCAMPUS · the cue≈ HBM · μs · smallparallel cue · only the matching neurons fireCORTEX · the content≈ SSD · ms · vast
The mathematics of one-shot recall
retrieved = softmax(β · Q · KT) · V
The same operation that drives attention inside every transformer model — extracted, scaled, and run on the hardware hierarchy that mirrors the brain's own.
124×
retrieval vs FAISS baseline
0 %
hallucination · 97% accuracy
concurrent users · bit-exact
Industry impact
One session's knowledge: from 19 books to 6,000.

Today's LLM session holds roughly nineteen books' worth of context. Recall handles six thousand books in the same session with zero hallucination — the entire English Wikipedia, 6.4 million articles, on a single GPU node, with no measured accuracy loss. An enterprise's whole knowledge base can sit inside one conversation instead of being paged through it.

The proof — Wikipedia, in one place
Corpus
6.4 M
Wikipedia articles
≈ 6,000 books
8 languages · 5 categories
Environment
NVIDIA GH200
480 GB · 96 GB HBM3
HBM · hippocampus
NVMe · cortex
BGE-M3 embeddings
Result
0 %
Hallucination
97 %
Accuracy
124 ×
vs FAISS · 10.45 ms recall
Live · publicly accessible
brain.umparumpa.com
Open in browser. Ask anything.
Runs on — standard GPUs. No new hardware.
Engine 03 — React instantly

It doesn't watch. It reflexes.

event-based vision (DVS) · spiking inference @ Akida · Kalman CA tracking · looming TTC τ=θ/θ̇ · constant-bearing miss estimate

An event camera sees only change. A static background produces no data at all, and latency is measured in microseconds rather than frames. A neuromorphic chip then decides at milliwatts.

On top of those two, React detects a fast-approaching object early, predicts where it is going with Kalman tracking and a looming time-to-contact estimate, and moves the machine before impact. Same perception, swap the policy: the robot that dodges becomes the robot that catches.

Reflex pipeline & latency budget
DVS eventssparse · μs latencypolarity histogram2ch · Δt=5 ms · int8SNN @ Akida91 KB · int8 · <1 WKalman CA + TTCτ=θ/θ̇ · miss estimateEVADE /INTERCEPTlatency budget — first reaction to an approaching objectframe camera + GPU≥33 ms frame wait + inferenceevent + Akida path<10 ms perception → decisionclosed loop validated with latency injection · 94%+ detection at speeds outside the training range
A frame camera has to wait for the next frame. An event has already arrived the instant something changed. The reflex is fast not because of compute — but because of structure.
94 %+
detection @ unseen speeds (sim)
91 KB
on-chip model weights
< 1 W
target inference power
Demo — reflex on a walking robot

What happens when you throw a ball at a robot?

Most robots simply get hit — their eyes are too slow. We took a commercially sold robot dog (Unitree Go2), changed not a single line of its walking AI, and attached only the reflex. All we send is one velocity command — “step aside” — in the exact format the Unitree SDK already accepts, so it drops straight onto the real robot.

01Cinematic view — large, left

The outside movie camera. The line trailing the ball traces its path.

02Top view — upper right

The bird's-eye view. The clearest way to read how far the robot actually side-stepped.

03Event view — middle right · the heart of it

The world as the robot's eye sees it. Only motion is visible; the static background is zero data. Less to see means faster thinking and almost no power. The box is a detection, the arrow is the predicted position 0.1s ahead.

04Status panel — lower right

The robot's state of mind: IDLE (all clear) → TRACK (something incoming) → EVADE (move). TTC is the robot's own estimate of time to collision, computed before a person could say "huh?".

Scene 1
Dodging from standstill
Decision within 0.06 s of release. It side-steps on its own legs.
Scene 2
Dodging while walking
Its eye shakes with every footfall. We cancel that shake using only the IMU already inside the robot — no added parts — and it side-steps mid-stride. The hardest scene in the clip.
Scene 3
Engine off
Same ball, same robot, our engine switched off. It takes the hit. That difference is the technology.

Physics simulation (MuJoCo). Robot model, mass and locomotion policy are all real Go2 spec. TTC is shown as estimate and ground truth side by side, because we would rather measure honestly. Next step: the physical robot.

Industry impact
Machines that move on the power of a reflex.

Because the reflex runs under a watt — brain power, not a tens-of-watts GPU — a battery-powered drone or a walking robot can keep evasion and safety switched on all the time instead of only when the budget allows. Reflexes as standard equipment on every edge machine: that is the market this engine opens.

Works with — event cameras (Prophesee class) · BrainChip Akida · your robot
Where it goes

Anything that moves — and anything that moves toward it.

The demo is a robot dog and a ball, because that is the cleanest way to show a reflex. But nothing in the engine is about dogs, or about balls. What it does is far more general: detect something on a collision course, estimate when it arrives, and act before it does. Every machine that shares space with moving things needs that, and today almost none of them can afford it.

It is a module, not a product line — an event sensor, a sub-watt chip, and a policy. Adding it does not mean redesigning the machine, retraining its control stack, or waiting for the next hardware generation. That is why the list of places it belongs keeps growing rather than narrowing.

01
Drones in the air

Birds, branches, cables, another drone in the same corridor. Inspection under bridges and inside tunnels where GPS is gone. Disaster response and search is one obvious case — but it is an example, not the boundary.

02
Conveyors and production lines

At line speed the decisive moment falls between two frames of an ordinary camera. A reflex that fires inside that gap can reject a foreign object, catch a jam as it starts, or divert a part without slowing the belt.

03
People and robots in one space

Same perception, opposite policy. Evade the person who steps into the cell; intercept the part that is handed over. A safety layer that runs continuously instead of a light curtain that stops the line.

04
Warehouses and logistics

Mobile robots in aisles shared with people and forklifts. Falling boxes, dropped pallets, a cart pushed around a blind corner — the events that a fixed schedule can never anticipate.

05
Field, construction and inspection

Swinging loads, falling debris, machinery in motion, dust and vibration that blind a frame camera. Change-only vision is exactly the sensor for scenes where most of the picture is irrelevant.

06
Anything battery-powered

The reason this list is long is the power budget. Under a watt, a reflex is always on. At tens of watts, it is a feature someone eventually switches off — and then it may as well not exist.

We are deliberately not choosing one vertical to name as the market. A reflex is infrastructure — the way a brake or a shock absorber is infrastructure — and the machines that need one are not a category, they are a majority.

A datacenter that remembers selectively. A knowledge system that recalls associatively. A machine that reacts instantly. Three engines, one brain — and in every case the silicon is someone else's, already shipping, suddenly doing more.

Why this works

A fly, with a speck of a brain, evades the fastest human hand.

It is not winning on compute. A fruit fly has about a hundred thousand neurons and a power budget measured in microwatts. It wins on structure — it keeps only what it needs, recalls by cue, and reacts to change rather than to pictures.

Von Neumann separated memory from compute, and every chip built since — CPU, GPU, TPU — inherited that boundary. Frontier scale is now colliding with the wall it implies: the cost of moving data exceeds the cost of computing on it, and memory costs more than the silicon that uses it. The next watt of efficiency is no longer hiding in the transistor. It is hiding in the structure.

So we don't build hardware. We implant that structure, and the hardware you already own does more. GPU, SSD, neuromorphic silicon, event sensors — the best silicon at every layer, with UmpaRumpa's brain layer on top of it.

Applications
LLM serving · knowledge systems · robots & drones
UmpaRumpa — the brain layer
XHBM · Recall · React
Silicon
NVIDIA GPU · Solidigm SSD · BrainChip Akida · event sensors
Today — Von Neumann
Memory
HBM
80 GB
PCIe bus
1 % utilized · bottleneck
Compute
GPU
700 W
Every operation moves data across this boundary. Most of the bus sits idle.