Yizhu Technology is reshaping artificial intelligence computing hardware by combining 3D-DRAM memory with a fully digital in-memory computing architecture. According to eu.36kr.com, the approach aims to cut data transport costs and slash total cost of ownership for large model inference by up to 88 percent compared to traditional cloud GPUs.
Addressing Data Transport Bottlenecks Through In-Memory Architecture
Traditional computing hardware separates processing units from memory. This separation forces systems to shuttle data back and forth continuously, creating severe latency and driving up power consumption. Yizhu Technology integrates logic directly into or near the memory medium. Calculations occur as data flows out of storage, reducing multi-level cache dependencies and eliminating redundant transport steps. The design does not replace high-bandwidth memory or 3D-packaging techniques. Instead, it complements them by minimizing the physical distance data must travel.
The architecture pairs specialized in-memory computing units with universal SIMT units. In-memory units handle fixed-parameter, data-intensive tasks like matrix multiplication and addition. Universal SIMT units tackle complex control logic, scalar operations, and irregular calculations. This heterogeneous fusion preserves the energy efficiency of in-memory computing while retaining the programming flexibility of standard GPUs, maintaining hardware compatibility with mainstream software ecosystems like CUDA.
Overcoming Capacity and Precision Limits With 3D-DRAM and Digital Circuits
Memory capacity has historically restricted how many model parameters an in-memory chip can handle. Yizhu Technology adopts 3D-DRAM to expand storage capacity while simultaneously boosting memory bandwidth and data throughput. This physical stacking allows chips to scale far beyond the megabyte limitations of older designs, ensuring efficient token generation during large language model inference.
To preserve calculation accuracy, the company relies on fully digital circuits rather than analog pathways. Digital in-memory computing avoids the precision losses typical of analog components. Third-party validations cited by eu.36kr.com show that Yizhu’s universal in-memory design achieves multiple times the performance of mainstream cloud GPUs while consuming about one-fifth of the power and costing roughly one-sixth as much.
Expanding Software Ecosystems and Agentic AI Support
Hardware improvements are supported by a full-stack software system developed in-house. This stack includes a proprietary instruction set architecture, hardware-software co-designed architectures, heterogeneous computing IPs, a CUDA-compatible compiler, and a cluster-wide KV-cache pooling system. The CUDA-compatible instruction set and PyTorch-compatible development environment lower migration barriers for developers accustomed to mainstream hardware.
The architecture is engineered to handle emerging demands from agentic AI, supporting one-million-token context windows, mixed-precision calculations, and Mixture-of-Experts sparse computations. Yizhu Technology currently distributes its products across cloud data centers, edge workstations, all-in-one devices, and distributed clusters to meet varying enterprise scale requirements.
Frequently Asked Questions
What core memory technology does Yizhu Technology use?
Yizhu Technology builds its architecture around 3D-DRAM as the primary storage medium, enabling higher memory capacity and bandwidth for large AI models.
How does digital in-memory computing differ from analog approaches?
The company uses fully digital circuits to maintain the determinism of digital systems, avoiding accuracy degradation caused by analog component variations.
What efficiency gains have third-party tests recorded?
Third-party validation indicates the architecture reduces inference total cost of ownership by up to 88 percent while consuming approximately one-fifth of the power of mainstream cloud GPUs.
Worth a look