New analog chip that is 1k times faster than high-end Nvidia GPUs (www.livescience.com)

Researchers at Peking University published a Nature Electronics paper describing an analog processor built from arrays of resistive RAM (RRAM) cells that they say can deliver up to 1,000× higher throughput and ~100× better energy efficiency than top-end digital GPUs (tested against chips like Nvidia’s H100 and AMD Vega 20) on certain tasks. The device performs computation in-place by steering continuous electrical currents across RRAM crossbars, avoiding energy- and time‑consuming data shuttling between memory and compute. In benchmarks focused on communications problems — notably matrix inversion used in massive MIMO systems — the chip matched the accuracy of digital processors while using far less energy, then scaled to much higher throughput with modest adjustments. The team addressed the classic precision and stability limits of analog computing by using a two‑stage architecture: a fast, approximate analog pass followed by iterative refinement circuits that converge to digital-grade accuracy. The chip was fabricated with a commercial process, suggesting practical manufacturability, and the authors argue this approach could revive analog in domains where bandwidth, latency and power dominate (AI model kernels, real‑time 6G signal processing). Key implications: in‑memory analog compute with RRAM could change hardware trade-offs for matrix-heavy workloads, but results are workload‑dependent and will require larger integrated designs and robustness testing before displacing general-purpose GPUs.