Collaborative
Request of Collaboration

High Capacity Memory Subsystem Development

Introduction

This project aims to design and implement a high capacity memory subsystem for A series CPU based SoCs. 

Project Milestones

  1. Architectural Design

    Design Flow

    Architectural Design

    Architectural Design

    Target Date

     

    Determining the information needed to begin architectural design (Specification)

    The team are reviewing the JEDEC DDR4 specification. The team are also reviewing available resources and projects with the DDR3 protocol. One resource extends understanding of the JEDEC timing specification with explanation on the timing information of signals.

    A final Specification will need to include a list of proposed use cases, specific algorithms to be implemented, data movement requirements, time critical operations, etc.

    Separation of Concerns:

    Separate aspects of the design task to specific parts of the system by defining a unique sub-system. Srimanth Tenneti produced an initial subsystem breakdown. 

     

    Identify reusable design patterns:

    To be determined. The AXI and APB interfaces to the System buses are well known and architected. 

    Identify existing third party blocks:

    To be determined. 

    Identify that need to be developed uniquely:

    To be determined. 

    System Throughput  Considerations:

    Review JEDED specifications.

    Design Decomposition / Partitioning:

    Once the set of IP parts has been identified the work tasks of considering the interfaces between them will begin. Starting with the connections internally between subsystems, and those to the external system environment.

  2. APB interface

    Navya Deepika is working on the definition of the APB interface and the interactions with the external SoC and internal interactions with blocks such as the Timing Parameters Module being worked on by Dhanaji Pawar.
     

  3. AXI Interface

    Sandeepan Roy is working on the main data flow to/from the rest of the SoC via the AXI interface. 

Team

Name
Sandeepan Roy
Research Area
Computer architecture, Machine Learning, Accelerator Design, Architecture Verification
Role
Independent Researcher
Name
Soumya
Research Area
SOC WITH AI
Role
Student
Profile picture for user Dhanaji_Pawar
Name
Dhanaji Pawar
Research Area
Accelerator Design , Architecture Verification,Computer architecture
Role
Design and Verification engineer

Comments

Hi All, 
Could you please use this link to join the introductory meeting this friday at 3pm UK time (BST)
https://teams.microsoft.com/l/meetup-join/19%3ameeting_NzJmOWMzZjQtMTYxNi00NTA3LTlhMTQtMjMwNDY4N2NhMDQw%40thread.v2/0?context=%7b%22Tid%22%3a%224a5378f9-29f4-4d3e-be89-669d03ada9d8%22%2c%22Oid%22%3a%22da03259c-2f3e-4038-96bb-de5e01994a6c%22%7d

Hi,

We have just finished our 2nd video call for the project. We did not record it. Not everyone has signed up on the slack channel which will be a location for frequent communication. We will make a regular update to the project so that everyone can keep aware of the weekly drum beat. Srimanth said we would share the meeting summary so hopefully we will draft that up and share it later.

Srimanth discussed his rough block diagram of the DDR Command Generation module. 

An initial design decomposition focusing on the signal interchange was felt to be a good start to allow people to engage. 

Various blocks need fleshing out and we need to assign people have a first look and then we can see why the blocks are misaligned. 

We will continue to meet weekly and try and move things forward. Look out for the more details meeting summary. 

John. 

 

 

 

Hi,

To try and reduce the amount of email notifications we have added a moderation step. The owner of the project will get all comments and they can decide if the comment should be forwarded to all of the project team via email. We hope this way that important comments will flow via email and you can catch up on other minor comments when you visit.

Let us know if this works as we can change things if needed.

Hi,

We have just finished our 3rd video call for the project. Last weeks meeting we stated that various blocks need fleshing out and people assigned to do the initial work so we can see why the blocks are misaligned. 

@Sandeepan Roy had put his hand up for the Host Interface Logic - AXI. Work has progressed on the initial external interface but more needed on the internal interfacing but that need other blocks to form. There was discussion on potential separation of responsibility between the AXI interface with a primary responsibility on data movement and the potential use of APB interface for configuration. This led on the a discussion on address translation and the various registers that will need to be accessed.

@Srimanth had taken the Command Module. There was some discussion on the hand off from the AXI interface and the Command Module and the form of address to be expected by the Command Module. This led to a need to define the responsibilities for each of the blocks. 

People where directed to following on line book on System on Chip Design with Arm Cortex M Processors.

@Dhanaji Pawar was to look at the Timing Parameters Module. The meeting did not get around to discussing this block. 

Some new team members joined the call and it was good for them to get involved. @Srimanth said he will produce some more initial statements of block functionality to help people decide how best to engage on investigation of the various blocks.

Thanks all,


 

We have just finished another virtual meeting.

@Sandeepan Roy talked people through his memory system architecture document allowing the meeting to review it and use the question and answers to help everyone with a shared understanding of the design. The architecture extends the initial subsystem breakdown adding two new blocks 

Specification:

The controller architecture is based on DDR4 as specified in the JESD79-4.

Separation of Concerns:

@Sandeepan Roy in the period worked on defining the AXI interface block and interactions with the other blocks. A Memory Arbiter which samples the address and sends them in the address queue. A Memory Scheduler that passes address requests  whenever requested by the command module. A good discussion was had on how data read/write requests flow through the AXI interface. 

@Sandeepan Roy and @Srimanth need to finalise the hand off from the AXI interface and the Command Module.

Not much definition has happed on the APB interface. The start up sequence and how the information held in the Params module is to be established in combination with the system boot sequence was discussed to help determine some of this blocks interactions. 

Verification:

A discussion on the verification strategy occurred and we agreed that the second meeting in December would be dedicated the verification.

Angelo has made the point that if the controller is for AI/ML applications, then the controller needs to be fast. He believes UberDDR3 is fast compared to commercial DDR3 controllers like the Memory Interface Generator (MIG) in Vivado, but uses much less area than the MIG. 

His recent blogpost discusses a Dhrystone Test comparing the performance of MicroBlaze with UberDDR3 against MicroBlaze with MIG. 

Dhrystone test performance results:

  • MicroBlaze with UberDDR3 = 0.3154 DMIPS/MHz

  • MicroBlaze with MIG = 0.3061 DMIPS/MHz

UberDDR3 uses 32.63% fewer Slice LUTs and 32.33% fewer Slice Registers than MIG.

  • UberDDR3: Slice LUTs = 3083, Slice Registers = 2743
  • MIG: Slice LUTs = 4576, Slice Registers = 4047

The PHY used by UberDDR3 is specific to Xilinx 7-series FPGA. 

Add new comment

To post a comment on this article, please log in to your account. New users can create an account.

Project Creator
Srimanth Tenneti

Researcher at University of Cincinnati
Research area: Machine Learning | SoC Design

Submitted on

Actions

Log-in to Join the Team