What is the maximum number of cameras and MMICs that can be input into a holoscan sensor device?

[13eta]

Hello. I would like to ask a few questions before implementing HSB in our product.

First, we use TI’s MMICs. These are radar-specific ICs, and recently, application notes like the one shown in the photo below can be found on TI’s website. We intend to apply these MMICs to HSB, but each radar device is equipped with four MMICs. Each MMIC operates at 450 Mbps per lane and is configured with 4 lanes.

We intend to configure approximately 12 of these radar devices, and the configuration is as follows:

Radar Device (MMIC4) * 12 = MMIC48

Considering the data bandwidth, since each lane is 0.45Gbps, the total bandwidth is 86.4Gbps. If we consider the overhead of the HSB and assume it is 15%, the bandwidth becomes 99.36Gbps.

When configuring the block diagram as shown above, four MMICs are converted to LVDS using CrossLink-NX, and SFP+ outputs are sent via CertusPro-NX (which has Holoscan IP) to a switching hub, which is then connected to Thor.

While I would need to receive a response from Lattice to confirm whether this configuration is actually feasible, I am curious whether NVIDIA’s Thor can handle these 48 MMICs, regardless of the FPGA implementation status at present.

In the existing GMSL2 SERDES environment, data pipelines had to be configured by assigning CSI ports, setting Vi-ids, and allocating PHY controllers for each MMIC. I have attempted to configure 12 MMICs on a single Thor using this configuration, but I am wondering if data processing close to 100Gbps is possible software-wise when configuring 48 MMICs on a single Thor using HSB.

Additionally, I would like to know if individual software settings for the 48 MMICs are required on HSB, and if this is even possible. (I am unsure if it is possible to assign IPs to each FPGA board and IDs to the internal MMICs.)

Consequently, there are 12 HSB boards. Each HSB board is equipped with 4 MMICs, and these HSBs utilize SFP+ to occupy a data bandwidth of 8.6 Gbps.

Approximately 12 of these HSBs are connected to a switching hub and then connected to Thor via a QSFP28 port.

I would like to know if Thor can process all the data from each HSB and 48 MMICs, if this can be implemented in software and if there are any references worth consulting, or if technical support is available.

[ShaneCCC]

I think the bottleneck is the ethernet bandwidth due to HSB transfer the data via ethernet.

Thanks

[13eta]

Hello. The core of my question concerns the maximum number of MMICs that can be operated on an HSB.

I understand that the number of cameras that can be connected is currently limited in the case of CoE or SIPL. The key point is whether it is possible to operate 48 MMICs on an HSB by configuring, for example, a LinuxReceiverOperator or RoceReceiverOp.

Please review whether this is possible.

Of course, if data bandwidth itself is the issue, the bandwidth per MMIC lane can be limited to 300 Mbps, allowing for operation at 66.24 Gbps while taking into account the HSB overhead.

[ShaneCCC]

Maybe consult with Lattice for the detail.

Thanks

[13eta]

Please read the text again.

What I mean is, regardless of the Lattice FPGA, I am curious about the maximum number of configurable MMICs in the HSB.

Just like the sentence ‘I understand that the number of cameras that can be connected is currently limited in the case of CoE or SIPL,’ I need to verify if the number of MMICs that can be specified is actually limited within the HSB driver.

Please review this and reply.

[ShaneCCC]

I think your question is the FPGA(HSB) able to support 48 MMICs right?

[13eta]

So far, I have been asking questions regarding Thor, its HSB drivers, and the software environment.

My questions are as follows:

1. Thor will receive IP and device IDs via Ethernet; is there sufficient computing resource for Thor to handle 48 MMICs? Thor has QSFP28 ports; is it actually capable of handling speeds up to 100Gbps?

2. For example, I understand that when connecting cameras in Orin, settings such as Vc-ID and CSI Port must be configured, and consequently, the maximum number of cameras that can be connected is fixed.

Similarly, I would like to know the maximum number of MMICs that can be operated when MMICs are connected to Thor via HSB and the data pipeline is configured software-wise. To be more specific, based on what I have already seen in other forums, I understand that camera drivers such as CoE or SIPL have a limit on the maximum number of cameras that can be connected, typically around 2 or 16. However, regarding MMICs, I am curious about the maximum number of MMICs that can be operated when configuring a ReceiverOperator in the Holoscan SDK, rather than using those drivers.

3. I would like to know if the following configuration is possible in Thor:

12 radar devices → 12 transport configurations (per IP) + 4 internal MMICs for each device, distinguished by packet header / channel ID.

[ShaneCCC]

For the Thor side I believe it depend on the software how you handle it.

[whitesscott]

This might be worth a look at some point. Nvidia PVA sdk Radar Pipeline Architecture Overview

[ShaneCCC]

I think you need to check if the HSB(Lattice FPGA) able connect MMICs by it MIPI interface.