Latest research from arXiv, written for makers and embedded engineers.
Researchers show that applying an FFT before training a feature extraction network can cut prediction error by up to 50% in low-data, resource-constrained settings.
2026-07-05
Memristors promise denser, lower-power memory than SRAM, but reliability threats stand in the way. Here is what embedded engineers need to understand.
2026-07-05
Researchers found that after a GPU command finishes on Apple M4 Pro, leftover cache displacement measurably slows the CPU. Here's what that means for builders.
2026-07-02
A detailed reverse-engineered breakdown of Apple's Neural Engine reveals the hardware, firmware, and dispatch paths that Core ML hides from developers.
2026-07-02
Researchers built a software-only method to verify which GPU you're actually running on in the cloud, using latency measurements and network probes.
2026-06-25
A new survey traces how hardware synchronization primitives evolved from the NYU Ultracomputer to modern GPU clusters, with lessons relevant to any multi-core embedded design.
2026-06-20
DataGuard embeds differential privacy enforcement directly into AI accelerator hardware, removing the need to trust third-party software with sensitive training data.
2026-06-19
HAMON encodes time-series data onto light and uses diffractive optics to compute forecasts passively, outperforming digital models on several benchmarks.
2026-06-17
A new binary tree approach to Gaussian random number generation cuts energy by 3.7x and boosts throughput density by 5.8x, making Bayesian Neural Networks practical on ultra-low-power hardware.
2026-06-16
A new compression technique stores neural network weights as small deltas instead of full values, cutting memory use by nearly 50% on resource-constrained FPGAs.
2026-06-16