LeoDISK / MEMORY INTERFACE
The implementation
and the tools to understand it.
Independent memory lanes feed LeoDISK. Its controller combines the traffic and presents one ordered port through the host interface you choose.
Memory
Interface
Choose the memory side.
Combine memory building blocks with independent controllers and clocks. Faster channels can receive larger shares of large accesses.
Choose the host side.
Select the interface for the application: PCIe, CXL, SerDes, or an HBM-like port, with qualification tracked for each integration.
Predict, generate, validate.
The memory compiler sizes the subsystem and predicts its bandwidth. Simulation and board checks compare the implementation with that configuration.
MEMORY INTERFACE FAMILY
Compile and Design Flow
Explore how your requirements become operating hardware. Select any stage in the flow.
01 / WHAT IT DOES
Give ordinary memory
a wider path.
LeoDISK joins many standard DRAMs into one fast memory and presents them to a processor as a single, ordered address space. Choose DDR5, LPDDR5, DDR4 or HBM, as many as the system needs. It works with our AI core and with other hosts.
The problem
AI is limited by memory bandwidth.
A language model streams its whole model through the chip for every token it generates. Tokens per second follow the memory’s GB/s, not the compute.
HBM is fast but scarce and costly.
It needs advanced packaging, its supply is limited, and it ties a design to one memory type and one packaging flow.
Each DRAM channel is narrow.
DDR5, LPDDR5 and DDR4 are everywhere and affordable, but one channel alone delivers only a fraction of what an accelerator needs.
What happens to each access
- ReceiveThe host sends a request through one port, in its own protocol.
- DistributeEvery large access is spread across all memories, so they all work on it at once.
- Work in parallelEach memory runs on its own clock. A memory paused for refresh does not stall the others.
- Reassemble in orderLeoDISK puts the pieces back together, in order, before they return to the host.
02 / WHY LEODISK
Hardware built for your system,
not a system built around the hardware.
Memory products on the market are fixed: one memory type, one host interface, one capacity, one bandwidth. System designers pick the closest part and design around its limits. LeoDISK is a compiler with a path to product. You state what your system needs, the memory, the host interface, the bandwidth and the capacity, and the compiler generates the memory controller RTL for that spec. From that RTL the product is built through ASIC design, manufacturing, test and packaging; FPGA builds serve as the proof of concept.
Built to your spec, not to the catalog
State what your system needs. The compiler generates the memory interface hardware to that spec, instead of you designing the system around the closest part on the market.
Any memory, any interface
Choose the memory: DDR5, DDR4, LPDDR5, HBM, or SSDs for even wider bandwidth. Choose the host side: PCIe, CXL, SerDes or an on-chip bus.
As wide and as large as you need
Bandwidth comes from the number of memory channels and capacity from the memories you choose, so both are set at compile time to your target. On the board, adding memories made an AI model 1.85× faster per token.
A path to product
The compiler generates the controller RTL for your spec. From there it goes through ASIC design, including place and route and timing closure, then manufacturing, test and packaging. FPGA builds serve as the proof of concept.
The LeoDISK compiler vs fixed memory products
| LeoDISK compiler | Fixed memory products | |
|---|---|---|
| Starting point | Your system spec. The compiler builds the memory interface to it. | A catalog part. The system is designed around what exists. |
| Memory type | Your choice: DDR5, DDR4, LPDDR5, HBM, or SSDs for wider bandwidth. Many options, chosen per design. | The one memory type the product was built for. |
| Host interface | The interface you need: PCIe, CXL, SerDes or an on-chip bus. | A fixed interface. |
| Bandwidth | As wide as you need: add memory channels until the target is met. | Fixed by the product’s channel count and speed. |
| Capacity | As large as you need: sized by the memories you choose. | A few fixed capacity options. |
| Channel count | A build parameter, from one memory to many. | Fixed in the product. |
| Planning | The sizing worksheet predicts bandwidth for your spec before anything is built. | Datasheet numbers for one part. |
| Path to product | Compile the controller RTL for your spec, then ASIC design (place and route, timing closure), manufacturing, test and packaging. FPGAs are used for the proof of concept. | No path to a part built for your spec: you buy what is offered. |
| Spec changes | Supported: recompile for the new spec, then take it through the same design and manufacturing path. | No path: you are stuck with what is available. |
The fixed-product column describes the usual pattern, not any one product. LeoDISK reaches HBM-class bandwidth by adding commodity channels; it does not claim to beat an HBM stack. SSD-backed configurations are scoped per design.
03 / SIZING WORKSHEET
Choices on both sides
of the controller.
Pick the memory building blocks on one side and the host interface on the other. The worksheet counts the devices LeoDISK joins to match the host, the capacity they give, and where the bottleneck sits. It uses the same math as the worksheet in the LeoDISK demo console, on public peak rates.