Research

Four processor generations. Four open-access papers. FPGA validated on AWS F2 and Kria K26.

N4

Coming soon

512-core dual-chiplet architecture. 4.19 million physical neurons, 134 million virtual via TDM. FPGA validated on AWS F2 at 62.5 MHz. Edge variant validated on Kria K26 at 0.378 W total board power.

512 cores
Dual compute chiplets, each with 256 neuromorphic cores
134M virtual neurons
4.19M physical, up to 134M via 32-context TDM
2.5D packaging
2 compute dies (7 nm) + 1 I/O die (12 nm), CoWoS-S interposer
48 GB HBM3E
2.0 TB/s bandwidth
100–300 mW
ASIC power projection at 28 nm
8 neuron models
Programmable microcode, 50 learning opcodes
10 synapse formats
CSR, dense, population, convolutional, block-sparse
62.5 MHz
Validated clock on AWS F2 UltraScale+

N4-Edge

Low-power embedded variant. SRAM-only, no HBM. Targets $5-20 FPGA modules.

32 cores
262,144 neurons
0.378 W
Total board power, Kria K26
2.59% LUT
3,036 of 117,120
+3.301 ns WNS
Positive timing slack
15,668 ts/sec
Sustained throughput on F2

Architecture evolution

Each generation extends the ISA and adds hardware capabilities. N1 through N3 have dedicated product pages with full specifications.

N1 N2 N3 N4
Cores 128 128 128 (16 tiles) 512 (dual chiplet)
Neurons 131K 131K 524K–1M 4.19M (134M virtual)
Neuron models 1 (CUBA LIF) 5 7+ (programmable) 8 (50-opcode microcode)
Neuron engine Fixed datapath Microcode Microcode + ANN mode Microcode + ANN/SNN hybrid
Weight precision 8-bit fixed 1–16-bit variable 1–16-bit variable 1–16-bit variable
Spike formats Binary Binary + graded Binary + graded + compressed Binary + graded + compressed
Learning STDP + reward 3-factor + homeostatic Per-tile, 28-opcode ISA 50-opcode, per-core
Synapse formats 3 4 4 + compression 10
Memory 1 level 1 level 4 levels (L1–L4) HBM3E 48 GB, 2.0 TB/s
Virtualization NeurOS (680+ networks) NeurOS v2
FPGA validated VU47P (16-core) VU47P (28 tests) F2 + K26 F2 + K26
ASIC projection 19–38 mW (28 nm) 100–300 mW (28 nm)

Publications

Fourth generation 29 pages

Catalyst N4: A 512-Core Dual-Chiplet Neuromorphic Processor with 134M Virtual Neurons, Spike Tensor Core, and Hardware Neuroscience Primitives

512 cores, 134M virtual neurons via TDM, 16x16 Spike Tensor Core, 8-head spiking attention, KV cache, hyperdimensional computing, Hopfield memory, hardware backpropagation, federated learning, post-quantum security, RV64GC with 14 custom opcodes.

DOI 10.5281/zenodo.19332513
Third generation 22 pages

Catalyst N3: A 128-Core Hybrid Neuromorphic Processor with Hardware Virtualisation, Per-Tile Learning, and Silicon Metaplasticity

128 cores, 16 tiles, NeurOS virtualisation (680+ concurrent networks), ANN/SNN hybrid mode, 4-level memory hierarchy, per-tile 28-opcode learning ISA, silicon metaplasticity.

DOI 10.5281/zenodo.18881283
Second generation 18 pages

Catalyst N2: Full Loihi 2 Parity in an Open Neuromorphic Architecture

155-feature Loihi 2 parity audit, 5 neuron models, programmable microcode engine, 3-factor learning with homeostatic plasticity, graded spike payloads.

DOI 10.5281/zenodo.18728256
First generation 13 pages

Catalyst N1: A 131K-Neuron Open Neuromorphic Processor with Programmable Synaptic Plasticity

128 cores, 131,072 CUBA LIF neurons, STDP + reward-modulated plasticity, 14-opcode learning ISA, SHD 90.6%.

DOI 10.5281/zenodo.18727094

Software

The SDK includes three execution backends. The CPU backend is the reference implementation. Every commit is tested against all three.

CPU
Reference implementation. Deterministic, exact integer arithmetic. Ground truth for all comparisons.
GPU (CUDA)
Parallel backend for training. Statistically equivalent to CPU. Differences limited to floating-point operation reordering.
FPGA
PCIe MMIO driver targeting AWS F2 and Kria K26. Bit-exact spike output match against CPU simulator.
Papers
Open access
Zenodo with DOI
N1 RTL
Apache 2.0
Open source
N2+ RTL
BSL 1.1
Proprietary
SDK
BSL 1.1
Proprietary
API client
MIT
Open source