Product Overview | What is CUBE?
Winbond 's "CUBE (Customized Ultra-Bandwidth Elements)" is a custom memory that achieves high bandwidth, ultra-low power consumption, compact size, and flexible design.
It is especially optimized for edge AI applications with high computational loads, maximizing the performance of SoCs in areas such as generative AI, image recognition, voice processing, and inference engines.
Furthermore, in response to the challenges posed by HBM, such as high cost, high power consumption, thermal design, and short life cycle, CUBE's low-power design reduces heat dissipation load, while its flexible design and stable supply system provide a sustainable alternative solution that can be used stably over the long term.
Comparison with HBM | Why CUBE now?
Currently, HBM (High Bandwidth Memory) is widely adopted as a high-performance memory, and the latest generation, HBM3E, is a representative technology that supports high-speed processing in the cloud and HPC fields.
Let's compare the latest HBM standard, HBM3E, with CUBE and see what the differences are!
|
characteristics |
HBM3E |
CUBE |
|
Stack Layer |
Max 12Hi |
Maximum 8Hi (flexible design) |
|
Transfer speed |
Up to 9.2Gbps |
Up to 2Gbps (low power design) |
|
bandwidth |
Up to 1.2TB/s |
Up to 4TB/s (depending on configuration) |
|
Number of I/Os |
1024 |
max 4096 |
|
power efficiency |
<4pJ/bit |
<1pJ/bit |
|
supplyability |
Short product cycle, EOL risk |
Long-term supply available |
|
Appropriate for use |
High-performance cloud-focused |
Ideal for edge AI and L4 cache |
HBM3E vs CUBE comparison chart
HBM is widely adopted for cloud AI and HPC applications as a next-generation DRAM that achieves high bandwidth, high density, and low power consumption. However, it faces the following challenges:
- Product life cycle is short, and EOL (End of Life) risk is high
・Thermal design is difficult, and in some cases it is difficult to demonstrate actual performance.
・High cost and complex design are required, and advanced technology is required for integration with SoC
On the other hand, in edge AI, there is a tendency to prioritize low power consumption, compact size, and long-term supply over high bandwidth, and HBM can sometimes be considered excessive.
To address these challenges for HBM, Winbond has designed the CUBE as a long-term product.
Product Features | Why choose CUBE?
High bandwidth exceeding HBM, ultra-low power consumption, anda 3D stack structure with excellent thermal design
It achieves power efficiency of less than 1 pJ/bit, consuming approximately one-quarter the power of HBM, reducing heat generation and simplifying the thermal design of the entire system.
・Flexible configuration for optimal application
Die capacity, I/O configuration, bus width, etc. can be flexibly customized according to the application, and the package can be made smaller and more dense.
・Supporting stable product development through long-term supply
This avoids the "EOL issue" that is often an issue with HBM, and is ideal for products that require long-term supply, such as industrial and medical equipment.
CUBE Product Specifications | Technical Specifications
|
item |
content |
|
memory capacity |
0.5〜16GB |
|
power consumption |
<1pJ/bit |
|
Number of I/Os and transfer speed |
Maximum 4096・2Gbps |
|
Total Bandwidth |
128GB/s to 4TB/s (depending on configuration) |
|
Stacking Configuration |
Supports both with and without TSV, up to 8Hi |
|
packaging technology |
TSV, Chip on Wafer, Wafer on Wafer, 2.5D/3D compatible |
CUBE technical specifications table
Summary | Beyond HBM, the future of AI memory created by CUBE
CUBE is more than just an HBM replacement.
This custom memory is flexible and scalable for a wide range of applications, from edge AI to HPC and even L4 cache.
It combines performance and power efficiency, ensuring a stable supply over the long term.
Inquiry
Please feel free to contact us with any questions about product details or technical information.