MCDMA or Multiple DMAs? Single HP port or Multiple HP ports?

[atmosfir]

Hello,

I am trying to build a axi-stream based co-processor. I want to accelerate a parallizeable algorithm, and I have succesfully implemented and tested my custom verilog based AXI stream IP with a typical DMA-interrupt and through a single HP port. Now, I want to increase parallelism in the co-processor by adding more AXI-Streams. I have found that there are multiple ways to do this.

1. Use multiple DMAs

2. Use Mulitple Channel DMA IP (MCDMA)

My first question, is there a diffrence in performance in both implementations? Say 2 DMAs and 2 channel MCDMA.

I have also read through this presentation about Co-accelaretor architecture

This presentations infers that it recommends to use one HP port and 2 DMAs to feed into 2 accelerators. My question is

1 HP port for multiple DMAs or an HP port for each DMA?

Why and whats the diffrence?

[demarco · Accepted solution]

Hi @atmosfir,

If you are trying to maximize bandwidth, you are better off using AXI DMA over AXI MCDMA. The reason is that while MCDMA has a nicer register interface to support multiple channels, those multiple channels all share a single stream/memory mapped AXI interface. If you use two AXI DMAs, both channels have a dedicated stream/memory mapped interface. 

Similarly, if you need a lot of bandwidth, you should dedicate AXI_DMA0 to AXI_HP0 and AXI_DMA1 to AXI_HP1. This way you have dedicated parallel channels for each AXI DMA.

Regards,

Deanna

[atmosfir]

Hello thank you for your answer. I have a follow up question. Can you please elaborate on what you mean by "dedicating" each HP port to each DMA? Thank you again.

[demarco]

Hi @atmosfir ,

Let's say you have two AXI DMA IPs in your system and they both need to get to PS DRAM. You can choose to have those two AXI DMAs go through a SmartConnect and share S_AXI_HP0_FPD.

But let's say you have analyzed the bandwidth needs of your AXI DMAs and don't think a single S_AXI port to PS DDR can meet your requirements. You could instead choose to have AXI DMA 0 hook up to S_AXI_HP0_FPD and AXI DMA 1 hook up to S_AXI_HP1_FPD. 

Regards,

Deanna

[atmosfir]

Hi @demarco 

Thank you for the timely answer. Is this possible with a zynq 7000 SoCs, that do not have HP_FPD ports?

and can you confirm that this is the architecture you are pointing to?

Thank you so much for your time.

[demarco]

Hi @atmosfir,

Yes, the same reasoning would apply to Zynq-7000 and its S_AXI_HP ports. Your block diagram implements the "I need more bandwidth" scenario.

Regards,

Deanna

[sanjayamv]

Thankyou for your wonderful explanation !!

Is it possible to connect multipltiple AXI_HP ports to a single axi-stream-dma. I have a stream dma of 512 bits wide going into my HLS IP, but each HP port is only 128 bits wide. How do I handle that situation ?

Regards

Sanjaya MV