Hybrid Analog-Digital Architectures
Analog & In-Memory Computing · Frontiers track
9 min read
In this module
The Best of Both Worlds
Analogue excels at matrix multiplication — performing it in O(1) time through physics. It struggles with precision, nonlinear operations, control logic, and programmability. Digital is the reverse. The logical move: don't pick one. Use both.
Hybrid analogue-digital architectures use analogue crossbar arrays for the compute-heavy matrix multiplications and digital circuits for everything else. Analogue handles the work it's built for; digital handles the rest with full precision and programmability. Neither compromises the other.
Anatomy of a Hybrid Analogue-Digital AI Accelerator
A typical hybrid architecture consists of the following components:
Analogue Compute Tiles
The core of the chip: arrays of crossbar tiles, each containing a memristive crossbar (e.g., 256x256 or 512x512 devices) that stores neural network weights and performs matrix-vector multiplication.
Each tile includes:
- DACs (digital-to-analogue converters) on the input side - converting digital input values to analogue voltages
- The crossbar array - performing the matrix multiplication
- ADCs (analogue-to-digital converters) on the output side - converting analogue output currents back to digital values
- Sample-and-hold circuits - capturing the analogue output before digitisation
Digital Processing Units
Surrounding the analogue tiles are digital circuits that handle:
- Activation functions: ReLU, sigmoid, GELU, and other nonlinear functions applied element-wise to the output of each layer
- Batch normalisation: Scaling and shifting activations based on learned statistics
- Pooling: Max or average pooling for convolutional networks
- Residual connections: Adding the input of a layer to its output (skip connections)
- Softmax: Computing probability distributions for classification outputs
- Quantisation/dequantisation: Converting between different numerical representations
Data Flow Controller
A digital controller orchestrates the flow of data through the chip:
- Load input activations into DACs of the appropriate analogue tile
- Trigger the analogue computation (apply voltages, read currents)
- Digitize the analogue outputs via ADCs
- Route the digital outputs to the digital processing unit for nonlinear operations
- Feed the processed outputs to the DACs of the next analogue tile
- Repeat for each layer of the neural network
Memory Hierarchy
Like digital AI accelerators, hybrid chips include:
- SRAM buffers: Fast on-chip memory for storing intermediate activations
- Input/output interfaces: For loading input data and reading results
- Weight programming interface: For writing weights to the crossbar arrays (typically done once per model, not during inference)
Trace one layer round the loop: the controller feeds a tile, the tile's ADCs hand digital numbers to the processing unit, and the processed activations go back to the controller to be fed into the next tile. Every trip round that loop crosses the analogue-digital boundary twice, which is why converter cost matters so much.
A pure analogue system would need to implement activation functions, normalisation, and control logic using analogue circuits - which are imprecise, inflexible, and hard to design. A pure digital system wastes enormous energy moving weight data for matrix multiplications. The hybrid approach gives each technology its strongest task: analogue handles the O(N²) matrix multiplications (which dominate energy consumption), and digital handles everything else with full precision and programmability. The result is better than either approach alone.
The DAC-ADC Challenge
The conversion between analogue and digital domains is the critical bottleneck — full stop. Every joule the chip saves on matrix multiplication can be squandered right here if the ADCs are wasteful.
DAC Design
Input DACs convert digital activation values to analogue voltages. Requirements:
- Resolution: 4-8 bits (matching the analogue precision)
- Speed: Must convert at the rate of inference operations (MHz range)
- Power: Must be low (many DACs needed - one per row of the crossbar)
- Area: Must be small (DACs sit at the edge of each crossbar tile)
Modern designs use current-steering DACs that can operate at hundreds of MHz with 6-8 bit resolution in compact areas.
ADC Design
Output ADCs are more challenging. They must:
- Resolve the accumulated current from an entire column (which is the sum of many partial currents), requiring higher dynamic range
- Operate at the output rate of the crossbar (often higher than the input rate if time-multiplexed)
- Consume minimal power (typically the dominant power consumer in the chip)
ADC power consumption scales exponentially with resolution. An 8-bit ADC consumes roughly 4x more power than a 6-bit ADC. This creates a tension: higher ADC resolution means better accuracy but higher power consumption.
Research solutions include:
- Successive approximation register (SAR) ADCs: Good balance of speed, power, and area for 6-10 bit resolution
- Sigma-delta ADCs: Very high resolution but lower speed; suitable for low-throughput applications
- Time-domain ADCs: Convert analogue values to time durations, which can be measured digitally. Promising for low-power designs.
- In-memory ADC approaches: Use the crossbar array itself to perform multi-bit quantisation, reducing the need for separate ADC circuits
Handling Negative Weights
A fundamental issue with analogue crossbar computing is that memristive conductances are always positive (current flows in one direction through a resistor), but neural network weights can be negative.
Differential Pair Approach
The most common solution uses two memristors per weight:
- G+ stores the positive component
- G- stores the negative component
- The effective weight = G+ - G-
Two column lines are used: one collects the positive currents (through G+ devices) and one collects the negative currents (through G- devices). The difference is computed either by an analogue differential amplifier or digitally after ADC conversion.
This doubles the number of devices needed but provides full signed weight representation.
Offset Approach
An alternative: bias all conductances by a fixed offset. If the weight range is [-1, +1], program conductances in the range [0, 2], then subtract the offset (1) from the digital output. This uses fewer devices but reduces the effective precision because part of the conductance range is used for the offset.
Mapping Neural Networks to Hybrid Hardware
Layer Partitioning
When a neural network layer has a weight matrix larger than a single crossbar tile (e.g., a 4096x4096 matrix on 256x256 tiles), the matrix must be partitioned:
- Row partitioning: Split the input vector across multiple tiles, each handling a subset of rows. Partial column currents from different tiles are summed digitally.
- Column partitioning: Split the output across multiple tiles, each computing a subset of output elements.
- Both: For very large matrices, partition in both dimensions.
The digital controller manages the data routing and partial sum accumulation for partitioned matrices.
Pipeline Execution
For multi-layer networks, the hybrid chip can pipeline execution:
- Tile A computes layer 1 on input batch 1
- While tile A processes batch 2, the digital unit applies activations to batch 1's output
- Tile B computes layer 2 on batch 1 (with activations applied)
This pipelining keeps the analogue tiles utilised while digital processing happens in between.
Batch Processing
Multiple input vectors can be processed in sequence through the same crossbar (which stores the same weights). This amortizes the cost of weight programming and ADC/DAC static power across many inferences, improving overall efficiency.
Energy Analysis: Where Does the Energy Go?
A detailed energy breakdown of a hybrid analogue-digital inference:
| Component | Fraction of Total Energy |
|---|---|
| ADCs | 30-50% |
| DACs | 5-15% |
| Analogue crossbar | 5-15% |
| Digital processing (activations, etc.) | 10-20% |
| Data movement (SRAM, routing) | 10-20% |
| Leakage/overhead | 5-10% |
The crossbar computation itself - the matrix multiplication - is remarkably cheap (5-15% of total energy). The ADCs dominate because they must convert many analogue values to digital at sufficient precision. This is the fundamental bottleneck: the analogue computation is efficient, but the analogue-to-digital interface is expensive.
Despite performing matrix multiplication essentially for free through physics, analogue AI accelerators spend 30-50% of their energy on ADCs - the circuits that convert analogue results back to digital form. This means the ultimate energy efficiency of analogue in-memory computing depends as much on ADC technology as on the crossbar array itself. Every improvement in ADC efficiency translates directly to improved system efficiency. This is why ADC design is one of the most active research areas in analogue AI chip development.
Comparison with Digital AI Accelerators
How do hybrid analogue-digital chips compare with state-of-the-art digital accelerators?
| Metric | GPU (NVIDIA A100) | TPU (Google v4) | Hybrid Analogue | Advantage |
|---|---|---|---|---|
| Peak TOPS | 312 (INT8) | 275 (INT8) | 10-100 (varies) | Digital |
| Energy/MAC | ~0.5 pJ | ~0.3 pJ | ~0.01-0.1 pJ | Analogue 3-50x |
| Weight memory | External HBM | External HBM | In-crossbar | Analogue |
| Precision | INT8/FP16/FP32 | INT8/BF16 | 4-8 bit | Digital |
| Programmability | Full software | Full software | Model-specific | Digital |
| Maturity | Production | Production | Early commercial | Digital |
The key insight: analogue chips win on energy per operation but lose on absolute throughput and precision. They are best suited for edge inference where energy budgets are tight, while data centre workloads (requiring high throughput and precision) still favour digital accelerators.
The Future of Hybrid Computing
Several trends are driving hybrid analogue-digital architectures towards wider adoption:
Improving device technology: Memristive devices are becoming more reliable, more uniform, and easier to manufacture. New materials and device structures are reducing variability and improving endurance.
Better ADCs: Research in time-domain, SAR, and in-memory ADC designs is steadily improving the efficiency of analogue-to-digital conversion.
Hardware-software co-design: Training algorithms that account for hardware constraints (noise, limited precision, drift) are making analogue-friendly models more accurate. The gap between digital-quality and analogue-quality inference is narrowing.
CXL and chiplet architectures: The CXL interconnect and chiplet packaging technologies enable heterogeneous systems where analogue compute tiles are integrated alongside digital processors on the same interposer, communicating through a standard interface.
Emerging applications: Edge AI, IoT sensing, autonomous vehicles, and satellite computing all require extreme energy efficiency. These are natural markets for hybrid analogue-digital chips, where the energy savings justify the reduced precision and programmability.
Analogue in-memory computing is a genuine rethink of where AI computation happens. Exploit the physics of resistive devices to perform multiplication and summation in place — no weight movement, near-zero energy for the core operation. Surround that with digital circuits for everything analogue can't handle cleanly. The result is a credible path towards AI hardware that's 10–100x more energy-efficient than today's digital accelerators. With AI workloads compounding at their current pace, that gap matters.