Introducing Hub Chiplets

Primemas hub chiplets deliver rich pluggable SoC infrastructure that can help accelerators and CXL solutions reach the market 5x-10x faster with over 10x cost reduction.

What are Hub Chiplets?


Hub chiplets are System-on-Chip (SoC) modules implemented in a pluggable chiplet format. Similar to conventional SoCs, they house a rich IP infrastructure (high-performance CPU cluster, Network-on-Chip bus, memory controllers, resource scheduler, and others) but are also equipped with high-bandwidth, low-latency die-to-die interfaces that enable their seamless resource sharing with neighboring dies.

Hub chiplets are extremely versatile and can be employed in a wide range of systems, including CXL memory solutions, AI accelerators, crypto engines, and data analytics accelerators. On top of that, they are also designed with modularity in mind, which allows them to be combined in large hub chiplet clusters or chained together with other chiplet dies, enabling effortless system scale-out.

Image Description

SoC Production
Conventional Flow

Conventional SoC production flow encompasses seven distinct phases and can take anywhere from 24 to 30 months (sometimes longer) to complete.

If the vendor needs to modify or scale out his existing system (e.g., to expand the target market), he must fall back to Phase 1 of the flow and repeat the entire process, effectively doubling the production time.

SoC Production
Chiplet-Based Flow

With Primemas, our partners can effortlessly produce new or modify their existing SoCs by simply pairing our hub chiplets with new accelerator dies or scaling out their configuration by clustering hub chiplets together.

These modifications only require reiterating through Phases 6-7 of the production flow, which leads to the TTM for the new SoCs being reduced by 5x (up to 10x if we account for bugs that commonly occur in Phases 1-6) compared to the conventional flow.

Hub Chiplet Scale-Out
Memory Solution

Increasing memory capacity in traditional SoCs requires vendors to either supply them with additional memory controllers or alter the existing controllers to support different memory types. In either case, this effectively means producing a new SoC, which, in turn, has to pass through the entire 24-30 month production flow, bringing along the associated manufacturing costs.

Solutions built with Primemas chiplet hubs, on the other hand, are much more flexible and can be scaled out by simply clustering multiple hub chiplets together, with each chiplet bringing its nominal capacity to the system. As a result, SoC vendors can quickly adapt to customer needs and produce new solutions in under 6 months by merely packaging hub chiplets in different configurations.


Memory Solution Scale-Out Using Hub Chiplets.

Hub Chiplet Scale-Out
Computational Solutions

With modular architecture, Primemas hub chiplets offer an excellent low-cost solution for scaling computational systems, e.g., Neural Processing Units (NPUs), from the edge to server infrastructure. By packaging multiple hub chipsets and NPU dies together, our customers can effortlessly compose high-performance computational systems, which can then be further interconnected to create large server-size infrastructure.

As such, Primemas hub chiplets offer a truly flexible SoC infrastructure that can be easily applied in computational solutions targeting both cost- and power-sensitive markets.


Computational Solution Scale-Out Using Hub Chiplets

© Copyright 2023 Primemas. All Rights Reserved.