QuantaraCore writing

Blog

Long-form writing from QuantaraCore. Memoria is the company's primary commercial platform. The archive also includes NeuroCUDA research notes and FinInsight experiments. Posts cite primary sources.

August 14, 2026

Reproduce NeuroCUDA Results

Clone github.com/Krishnav1/neurocuda, run python reproduce.py --quick, and match the published N-MNIST table.

August 14, 2026

Export PyTorch to NIR

Checkpoint to NIR graph to round-trip check. Residual graphs, not a definition of NIR.

August 14, 2026

NeuroCUDA vs SpikingJelly

Conversion compiler vs train-from-scratch SNN library. When each wins.

August 14, 2026

N-MNIST SNN Conversion

Event-based dataset path: 99.88% SNN vs 99.70% ANN. Not the CIFAR ResNet tutorial.

August 14, 2026

NeuroCUDA on Google Colab

GPU runtime, pip install, convert a checkpoint in the browser. Not a second local install guide.

August 14, 2026

QCFS ANN to SNN Explained

How calibration thresholds work. Bug-fix for frozen lambda lives on a different URL.

August 14, 2026

SNN Timesteps T Explained

Choose T=8, 16, or 32. Accuracy vs latency. GPU time is not Loihi energy.

August 14, 2026

BatchNorm Folding SNN

Fold BN into conv weights before IF replace, or activations shift and accuracy collapses.

August 14, 2026

Cite NeuroCUDA

BibTeX, PDF URL, GitHub, and honest backend labels for papers.

August 14, 2026

NeuroCUDA Docker

ROS2 image vs slim CI pip container. Not a Colab notebook.

August 14, 2026

NeuroCUDA Akida Status

Honest: physical Akida is not supported. No job ID. Use GPU/CPU plus NIR.

August 14, 2026

NeuroCUDA FPGA HLS

HLS C++ proof of concept. Not a synthesized bitstream. Not "runs on FPGA".

July 17, 2026

SpiNNaker Physical Silicon Confirmed

Jobs #420148 and #420186 - two boards, identical spike counts, EBRAINS Manchester. Honest limits vs ResNet GPU numbers.

July 4, 2026

PyTorch to Loihi 2 Without Lava

Intel archived Lava. Convert and validate a PyTorch model against Loihi 2 neuron equations with NeuroCUDA - no Lava SDK and no INRC for simulation.

July 4, 2026

How to Convert PyTorch to SNN (Step by Step)

QCFS calibration, BatchNorm folding, IF replacement, BPTT fine-tuning. Full pipeline with code and verified CIFAR-10 numbers.

July 4, 2026

pip install neurocuda: Complete Setup Guide

CUDA requirements, first conversion, backends (GPU, CPU, Loihi 2 sim), troubleshooting, and links to every tutorial.

July 4, 2026

snntorch vs NeuroCUDA: Which SNN Tool to Use?

Training from scratch vs converting a trained ANN. When snntorch wins, when NeuroCUDA wins, and how to combine them.

July 4, 2026

ResNet-18 SNN Conversion Tutorial (Real Code)

End-to-end ResNet-18 on CIFAR-10: load checkpoint, convert, fine-tune, verify 94.61% SNN accuracy with copy-paste code blocks.

June 29, 2026

NeuroCUDA vs GeNN vs Brian2

Three neuromorphic CUDA names, three different jobs. When each wins and real-world scenarios.

June 29, 2026

Neuromorphic CUDA: Complete Guide

Definitive reference: GeNN, Brian2, GPU-RANC, DVS pipelines, and the PyTorch deployment path.

June 28, 2026

Why Your SNN Accuracy Drops After ANN-to-SNN Conversion

The four most common causes, in debugging order: membrane-state reset, insufficient timesteps, uncalibrated QCFS thresholds, and unfolded BatchNorm. Real before/after data and a 10-minute checklist.

June 28, 2026

QCFS Threshold (Lambda) Not Learning During Conversion

If your QCFS clipping threshold freezes near its initial value, it's a gradient-flow and learning-rate bug, not a property of the method. The exact mechanism, and the parameter-group fix.

June 28, 2026

Loihi 2 vs GPU Energy: Real Joule Numbers

The per-spike vs per-MAC energy model behind neuromorphic energy claims, the real published constants, and why most public comparisons are modeled, not measured on silicon.

June 28, 2026

SNN Training Not Converging: The Dead Neuron Problem

A neuron that never spikes never gets gradient. How dead neurons cascade through a directly-trained SNN, how to detect them early, and how to fix the surrogate gradient and initialization.

June 28, 2026

Rockpool vs Sinabs vs Norse vs Brian2 vs Nengo

An honest comparison by what each framework is actually built for: hardware deployment, PyTorch research, biological simulation, or cognitive modeling - not a single winner.

June 28, 2026

NeuroBench Report Format Explained

What the Algorithm Track and System Track actually require, what counts as valid evidence, and how to format a report that doesn't conflate modeled and measured energy.

June 22, 2026

What Is ROS2? The Robot Operating System Explained

ROS2 isn't an operating system - it's robotics middleware. Nodes, topics, services, actions, DDS, and why ROS2 replaced ROS1, explained from the ground up.

June 22, 2026

ROS2 and Spiking Neural Networks: What Exists, What Doesn't

A survey of every public ROS2 + SNN project, Gridbot, the Loihi oculomotor head, ROS-MUSIC, and why almost none of it runs on ROS2 or installs with pip.

June 22, 2026

ROS2 Jazzy vs Humble: Which Distro Should You Use?

Ubuntu version, support timeline, and what actually changed in Jazzy, plus which distro NeuroCUDA ROS2 targets and why.

June 22, 2026

Event Cameras and ROS2: DVS Drivers and SNN Pipelines

The existing ROS2 event-camera drivers, the event_camera_msgs format, and how to feed event data straight into a spiking neural network.

June 22, 2026

Neuromorphic Robotics with Loihi 2 and ROS2: The Research So Far

The Astrobee RL-to-SNN pipeline, hexapod locomotion, TurtleBot3 multi-robot control, and what's still missing for production use of Loihi 2 in ROS2.

June 22, 2026

How to Deploy a Spiking Neural Network on a ROS2 Robot

Nodes, topics, message types, and a Docker shortcut for running NeuroCUDA's compiled SNN on a robot's camera feed - the first pip-installable ANN-to-SNN compiler with a ROS2 package.

June 29, 2026

What Is Memoria? Private Social Memory vs Social Media

Memoria is a private social memory app for family circles - not social media, not a photo backup. How it works and who it is for.

June 21, 2026

The AI Energy Crisis Explained

Why inference, not training, is the real power problem - and where neuromorphic computing fits.

June 22, 2026

The SNN Software Barrier: Neuromorphic's Real Bottleneck

Loihi 2 and SpiNNaker 2 already exist. What's missing is a way to get a PyTorch model onto one without learning an entirely new framework first.

June 22, 2026

NeuroCUDA vs. NIRTorch: Comparing NIR Conversion

NIRTorch and NeuroCUDA both move models between PyTorch and NIR. How their approaches differ, including a verified bit-exact ResNet-18 result.

June 22, 2026

Intel Lava Is Archived: What to Use Instead

The lava-nc repositories are now archived with no successor SDK. What that means for Loihi 2 developers, and the alternatives that still work today.

June 22, 2026

Best Compilers for Neuromorphic Chips in 2026

Nengo, Intel's archived Lava, Hailo's Dataflow Compiler, and NeuroCUDA compared. What each actually does, and which fits a PyTorch workflow.

June 22, 2026

ANN to SNN Conversion Tools Compared

snnTorch, SpikingJelly, SNNToolbox, and NeuroCUDA, side by side: training-from-scratch tools versus after-the-fact conversion tools.

June 22, 2026

PyTorch Spiking Neural Network Tutorial

Shorter four-call quickstart: convert a trained PyTorch CNN, measure sparsity, and validate accuracy.

June 22, 2026

What Is NIR? Neuromorphic Intermediate Representation Explained

A vendor-neutral graph format for spiking neural networks, supporting 8 simulators and 5 hardware platforms, and where it still struggles with residual networks.

June 21, 2026

What Is a Neuromorphic Chip, and How Do You Run a Model On One?

A practical guide to neuromorphic chips and the compiler problem standing between a trained PyTorch model and real hardware, and how NeuroCUDA closes that gap.

June 20, 2026

Neuromorphic Chips in 2026: The Complete Field Guide

Loihi 2, SpiNNaker-2, BrainChip Akida, IBM NorthPole, and Innatera T1, compared head to head. Spiking vs non-spiking, digital vs analog, and where each chip actually fits the market.

June 20, 2026

Why Neuromorphic Computing Needs Its CUDA Moment

Before 2007, every GPU vendor had an incompatible toolchain. Neuromorphic computing is in that exact position today. The history, the NIR format, and why infrastructure always wins.

June 20, 2026

NeuroCUDA Launch: Inside the Benchmarks

94.49% SNN accuracy, zero deviations across 256,000 Loihi 2 comparisons, 93.67% sparsity. How each number was actually measured, and why the methodology matters as much as the result.

June 20, 2026

Building from Amravati

Why QuantaraCore is based in Maharashtra, with Memoria as the primary commercial platform and research work alongside it.