PKU Unveils World’s First Memristor Neurodynamics Chip and Open-Sources DSpark AI Inference Framework
Researchers at Peking University (PKU), working alongside the Chinese Academy of Sciences (CAS), have announced two significant advances spanning both hardware and AI software. The team has developed what it describes as the world’s first memristor-based neurodynamics chip, while PKU and DeepSeek have also released DSpark, an open-source inference acceleration framework designed to improve large language model (LLM) serving efficiency.
Published in Science, the neurodynamics chip demonstrates how in-memory computing can dramatically accelerate neuroscience workloads. Meanwhile, DSpark introduces a new speculative decoding architecture that significantly boosts LLM inference throughput under real-world, high-concurrency deployments.
Memristor-Based Neurodynamics Chip Eliminates the Memory Bottleneck #
Traditional computer architectures separate computation from memory. As data continuously moves between processors and storage, performance suffers from latency and high energy consumption—commonly referred to as the von Neumann bottleneck.
The PKU research team addresses this limitation by exploiting the analog conductance characteristics of phase-change memristors, allowing memory cells to both store information and perform computation simultaneously.
This compute-in-memory approach dramatically reduces data movement while improving execution speed and energy efficiency.
Chip Specifications #
The prototype was fabricated using a 40 nm manufacturing process and features:
| Specification | Value |
|---|---|
| Process Technology | 40 nm |
| Computing Array Area | 0.28 mm² |
| Single Computation Time | 2.12 ms |
| Computing Architecture | Phase-change memristor in-memory computing |
By completing calculations in just over two milliseconds, the chip enables neurodynamics simulations that were previously impractical for real-time applications.
Performance Compared with Conventional Accelerators #
Experimental results demonstrate substantial improvements over existing hardware platforms.
According to the published benchmarks, the chip delivers:
- Up to 36× higher performance than dedicated neurodynamics accelerators
- Up to 24× lower power consumption
- Between 50× and 478× faster execution than an NVIDIA A100 GPU during cerebral cortex reconstruction workloads
These gains stem largely from eliminating repeated memory transfers rather than relying solely on higher clock frequencies or additional processing cores.
Designed for Brain Science and Medical Computing #
The primary application demonstrated by the researchers is three-dimensional brain reconstruction.
The resulting models preserve complex cortical structures while avoiding artifacts such as unnecessary folds or geometric distortions, making the technology suitable for high-precision neurological analysis.
Potential future applications include:
- Brain-computer interfaces (BCIs)
- Real-time neural signal decoding
- Surgical neuronavigation
- Neurological disease screening
- Alzheimer’s disease research
- Parkinson’s disease diagnostics
- Brain-inspired computing systems
The project received support from multiple national research programs and contributes to ongoing efforts to develop energy-efficient computing architectures beyond conventional transistor scaling.
PKU and DeepSeek Release DSpark for Faster LLM Inference #
Alongside advances in neuromorphic hardware, PKU has also partnered with DeepSeek to release DSpark, an open-source inference acceleration framework focused on reducing latency and increasing throughput for large language models.
The framework specifically targets one of today’s biggest deployment challenges: maintaining fast response times under heavy concurrent workloads.
Why Large Language Models Need Faster Decoding #
Most modern LLMs generate text using autoregressive decoding, producing one token at a time.
Because each token requires a complete forward pass through the model, inference latency increases rapidly as generated responses become longer.
Speculative decoding has emerged as a popular optimization strategy, but existing methods often encounter two limitations:
- Sequential draft models become increasingly expensive as outputs grow longer.
- Parallel draft models experience declining candidate acceptance rates during long generations, wasting compute resources.
DSpark is designed to reduce both inefficiencies simultaneously.
Semi-Autoregressive Candidate Generation #
The first innovation introduced by DSpark is a semi-autoregressive generation architecture.
Instead of relying entirely on sequential decoding, the framework combines:
- A modified parallel backbone network that generates candidate representations in batches
- A lightweight sequential refinement module that restores token dependencies
According to the research team, a design using only two Transformer layers outperforms conventional five-layer parallel draft models while maintaining significantly lower computational cost.
Confidence-Aware Verification Scheduling #
The second major optimization focuses on verification.
DSpark introduces a confidence-scheduled verification mechanism that dynamically allocates computational resources according to:
- Candidate confidence
- Current hardware utilization
- Real-time system workload
Rather than verifying every candidate equally, the scheduler prioritizes higher-probability continuations, reducing unnecessary GPU computation during periods of heavy concurrency.
Benchmark Results #
The framework was evaluated using several mainstream open models, including Qwen3 and Gemma4, across three representative workloads:
- Mathematical reasoning
- Code generation
- General conversational dialogue
Compared with existing speculative decoding systems such as Eagle3 and DFlash, DSpark consistently achieved longer effective generation lengths per verification round.
For Qwen3-4B, the reported improvements include:
- 30.9% higher performance than Eagle3
- 16.3% higher performance than DFlash
These improvements allow DSpark to preserve the low-latency characteristics of parallel decoding while reducing efficiency degradation during long responses.
Engineering Optimizations #
Beyond model architecture, the development team implemented several system-level improvements designed for production deployment.
Training Optimizations #
Training efficiency was improved through:
- Sequence packing techniques
- Optimized data transfer pipelines
- Reduced memory consumption
- Lower overall compute requirements
Deployment Optimizations #
On the inference side, DSpark incorporates:
- Asynchronous scheduling
- Pipeline stall avoidance
- Separation of logical and physical verification
- Dynamic support for variable-length verification workloads
The framework remains compatible with existing CUDA-based GPU infrastructure.
Production Deployment Results #
DSpark has already been integrated into preview deployments of DeepSeek-V4-Flash and DeepSeek-V4-Pro, where testing under live production traffic demonstrated substantial throughput gains.
DeepSeek-V4-Flash #
Reported improvements include:
- 51% higher throughput while maintaining 80 tokens per second
- 661% higher throughput under a 120-token-per-second service target
DeepSeek-V4-Pro #
Under different service-level agreements (SLAs), throughput increased by:
- 52% at 35 tokens per second
- 406% at 50 tokens per second
The framework also dynamically adjusts verification lengths according to current system load, maximizing hardware utilization during light traffic while minimizing contention during peak demand.
Open-Source Availability #
PKU and DeepSeek have open-sourced the complete DSpark ecosystem, including:
- Training code
- Evaluation tools
- Model weights
- Reference implementations for DSpark
- DFlash
- Eagle3
The release aims to provide developers with a practical toolkit for building lower-cost, high-performance LLM inference services while reducing infrastructure expenses and improving user experience.
Outlook #
Together, PKU’s memristor-based neurodynamics chip and the DSpark inference framework highlight two complementary directions in next-generation computing.
On the hardware side, in-memory computing demonstrates the potential to overcome long-standing architectural bottlenecks for neuroscience and brain-inspired applications. On the software side, DSpark addresses one of today’s most pressing AI infrastructure challenges by improving speculative decoding efficiency for large-scale language model deployments.
As AI workloads continue to expand, innovations across both specialized hardware and optimized inference software are likely to play an increasingly important role in improving performance, reducing energy consumption, and lowering deployment costs.