Date: December 27, 2025

Topic: Hardware Architecture / Artificial Intelligence

Reading Time: 25 Minutes

The problem is not just transistor density; it is the Von Neumann Bottleneck.

In every digital GPU, CPU, or TPU, data lives in memory (DRAM) and processing happens in logic cores. For every single calculation, data must be fetched, moved across a bus, computed, and written back. In the era of Large Language Models (LLMs) like GPT-5 and Claude 4, this constant shuffling of weights consumes up to 90% of the total energy. We are effectively burning gigawatts of electricity just to move numbers around, not to calculate them.

2. The Physics of Efficiency: Kirchhoff vs. Boolean Logic

To understand why analog is winning, we must look at the fundamental operation of Deep Learning: Matrix-Vector Multiplication (MVM).

A neural network is essentially a massive collection of "weights" (synaptic strengths). When you ask an AI a question, your input vector is multiplied by these weight matrices. In a digital computer, this is a brute-force process. To multiply a 1,000-value input by a 1,000x1,000 matrix, a digital core must perform one million separate multiplication operations and one million additions, sequentially or in parallel chunks, moving bits back and forth for every step.

Analog computing cheats. It utilizes two fundamental laws of electricity:

1. Ohm’s Law ($V = I \times R$): If you pass a voltage (input data) through a resistor (the weight), the resulting current is the product. The multiplication happens instantly, physically.
2. Kirchhoff’s Current Law: If you wire multiple resistors together, the currents naturally sum up.

In an Analog Matrix Processor, the neural network’s weights are encoded directly into the conductance of memory cells (using technologies like ReRAM or PCM). You apply the inputs as voltages along the rows of a memory array. The physics of the circuit instantly multiplies the voltages by the conductances and sums the currents along the columns.

3. The Hardware Trinity: ReRAM, PCM, and Photonics

The concept is elegant, but the execution requires exotic materials. In 2025, three distinct technologies have emerged as the frontrunners in the race to kill the digital GPU.

A. Resistive RAM (ReRAM): The Edge Champion

Resistive Random Access Memory (ReRAM) has become the gold standard for edge inference. ReRAM cells work by forming and breaking conductive filaments inside a solid dielectric. By controlling the size of this filament, engineers can tune the resistance of the cell to represent a specific neural network weight.

• The Breakthrough: Just months ago, researchers at Peking University demonstrated a commercial-process RRAM chip that solved high-precision matrix equations 1,000 times faster than flagship GPUs, with accuracy rivaling digital fixed-point systems. This addressed the historic "Achilles' heel" of analog: precision.
• Key Players: Mythic has been a pioneer here. After navigating financial turbulence in 2023, their late-2025 "M2000" series chips are now seeing deployment in drones and security cameras, delivering 30 TOPS (Trillion Operations Per Second) for just a few watts. Ambient Scientific is also pushing boundaries with their DigAn™ architecture, blending analog MACs with 3D memory stacks for battery-powered smart devices.

B. Phase Change Memory (PCM): The IBM Bet

While ReRAM relies on filaments, PCM relies on heat. A PCM cell contains a chalcogenide glass (like Germanium-Antimony-Tellurium). By heating it, the material switches between a conductive crystalline state and a resistive amorphous state.

• The Tech: IBM Research has spent years perfecting this. Their latest "Analog AI" cores utilize PCM to store weights. The advantage of PCM is its stability and density. IBM's recent "NorthPole" digital-analog hybrid concepts have shown that you can run transformer models (the architecture behind LLMs) with 14x better energy efficiency than 4nm digital nodes.
• The Nuance: PCM is temperature-sensitive. The "drift" of resistance over time was a major hurdle, but 2025 has seen the introduction of active drift-compensation algorithms that recalibrate the analog values transparently, making these chips viable for enterprise servers.

C. Photonics: Computing at the Speed of Light

If ReRAM and PCM are the evolution of electronics, Photonics is a different species entirely. Companies like Lightmatter and Lumai are not using electrons; they are using photons.

4. The "50% Utilization" Crisis and the Analog Fix

Why are data centers desperate for this switch? In early 2025, a startling metric came to light: GPU utilization in inference workloads often sits below 50%.

Digital GPUs are designed for massive batch processing—crunching huge blocks of data at once. But real-world AI is "bursty." A user asks a chatbot a question; a car sees a pedestrian. These are single, immediate events. Digital GPUs struggle to keep their thousands of cores fed with data for these small batches, leading to idle silicon burning power.

Analog chips do not have this overhead. Because the weights are stationary in the memory arrays, they have zero set-up time. You can wake up an analog chip, run a single vector inference in microseconds, and shut it down. This "instant-on" capability makes them the only viable solution for the massive scale-up of AI agents anticipated in 2026, where billions of autonomous software agents will require constant, low-latency inference.

5. Overcoming the Analog Curse: Noise and Precision

If analog is so superior, why didn't we switch years ago? The answer lies in the messy reality of physics.

Digital is perfect. $1 + 1$ is always $2$.

Analog is probabilistic. $1.0 + 1.0$ might be $1.999$ or $2.001$ depending on temperature, wire resistance, or electronic noise.

For years, this noise made analog chips useless for deep learning, which requires high precision (especially for training). However, the "Algorithm-Hardware Co-Design" revolution of 2024-2025 changed the game.

A. Noise-Aware Training

Startups like Rain Neuromorphics realized that we shouldn't try to eliminate noise—we should train the AI to survive it. By injecting noise into the model during the digital training phase (quantization-aware training), the neural network learns to be robust. When deployed on the "noisy" analog hardware, the AI performs flawlessly because it has already "seen" that level of variance.

B. Digital-Analog Hybrids

The most successful architectures of 2025 are not pure analog. They are hybrids. They use analog arrays for the heavy lifting (the matrix multiplication, which is 99% of the work) but use high-precision digital circuits for the accumulation and activation functions. This "sandwich" approach gives the best of both worlds: analog efficiency with digital stability.

C. The 24-bit Breakthrough

The recent work from Peking University using RRAM achieved 24-bit fixed-point precision. This is a massive leap. Previously, analog struggled to get beyond 4-bit or 8-bit precision. 24-bit is sufficient for almost all inference tasks and even some training workloads, effectively removing the precision argument from the table.

6. The Industry Landscape: Who is Winning?

As we survey the field in December 2025, a few distinct tiers of companies have emerged.

• The Hyperscalers (Internal R&D):