High sampling Quad channel ADCs + FPGA solution suggestion

[Megumini]

I am a newbie with exp 2/1000 to next level (which only took courses and work with System Generator with Xilinx blocks in Simulink once long time ago). But at some point I have to make a decision to purchase new equipment and develop a data acquisition+signal processing system on top of it (hell challenging). I tried to read posts in this forum, tried to contact with vendors. But because it is really big budget (20k including IP license, debugging etc...) plus I am such a beginner, I decided to do as much homework as I can before I make the decision.

The requirements are:

Quad Channel ADCs with sampling frequency within range: 400-600MSPS, resolution 14bits or more. Prefer no decimation before signal processing. Phase information between those channels are really important, which means they should be synced very well. (theoretically Amplitude imbalance between channels less than 1dB, phase imbalance less than 10 degrees..) The sampling frequency band of interest is 100-150MHz, not very wide. Other requirements like CDR more than 70dB, SFDR more than 60dB (most ADCs I checked meet those requirements)

And FPGA...Just need pair with the ADC, virtex 7 or Kintex UltraScale+ or Virtex UltraScale+ I guess? Estimate DSP slices more than 4000. (if do parallel fir), but I saw some ADC has NCO then it can even go down.

Questions:

1. If I found a nice ADC, should I choose Evaluation board from Analog Devices or third party vendor platform solution?

2. Anything else I have not considered yet..Like Vpp and adc solution etc.

3. Should I try to build the system in vivado to get more precise resources report..

4. What usually should I expect to come with products, like IP core? VHDL code? Configuration binaries? Examples?

5. JESD204B can guarantee the synchronization between separate FMC, is that a good option? like 2 * 2channelADCs instead of 1*4channelADCs

6. Oh yeah...suggestions of products.

(did you see how panic I am ? (╯°□°)╯︵ ┻━┻)(no need for answering everything, if I mess up, I gonna just run away from my work, its ok...)

(Thank you so much for even finish reading till here)

 

[thinkthinkthink]

I'd go for an UltraScale+ device with Speed Grade at least -2 to ensure your project will pass timing analysis at those high sampling rates. On UltraScale/UltraScale+ architectures you might benefit from Xilinx's Machine Learning based Intelligent Design Runs that will help with Implementation. It might just be me but I honestly don't think 20k dollars/euros is enough of a budget for this.

Edited December 16, 2021 by thinkthinkthink

[zygot]

I'd suggest that what's important for you to do is plan your approach to assembling off the shelf boards, if they exist. None of your requirements are out of the realm of possibility. I'd certainly not start with the FPGA platform. There are certainly 4 channel ADC devices out there, but can you get a board with them that fits your needs? Communications oriented ADC are not generally good for general sampling and visa versa. Moreover the FMC and HSMC mezzanine boards sold by ADC vendors are generally designed to evaluate the device, not be useful in a real application. Be careful.

I'd start with the analog to digital conversion end of your project. This involves the ADC front end through the presentation of the digital data and controls. Just because a board is in an FMC form factor doesn't mean that you can pair it up with any FPGA board having FMC type connectors on it. There's a lot of due diligence to  be done to avoid tragedy in this mine field.

There are a number of vendors that specialize in FPGA based communications and defense products. You might want to at least see what is available as these will be more integrated solutions to specific application requirements. I've spent many years doing the FPGA development board add-on card search... generally, the results aren't good.

Be careful of UltraScale boards as the FMC connector Vadj is generally 1.5V or lower making them incompatible with older FMC mezzanine cards. Be very careful about pin placements for clocks as FPGA devices have limitations with respect to banks and clocking for LVDS. A lot can go wrong. JESD204B or JESD204C converters are a good choice if you can get them on a board with the right front end.

I wouldn't pin your future on limiting your search to Xilinx devices.

You've left out a lot of important information but perhaps I've got you thinking along the right track..

[Megumini]

2 hours ago, thinkthinkthink said:

I'd go for an UltraScale+ device with Speed Grade at least -2 to ensure your project will pass timing analysis at those high sampling rates. On UltraScale/UltraScale+ architectures you might benefit from Xilinx's Machine Learning based Intelligent Design Runs that will help with Implementation. It might just be me but I honestly don't think 20k dollars/euros is enough of a budget for this.

Thank you for your reply! I will continue searching.

[Megumini]

53 minutes ago, zygot said:

I'd suggest that what's important for you to do is plan your approach to assembling off the shelf boards, if they exist. None of your requirements are out of the realm of possibility. I'd certainly not start with the FPGA platform. There are certainly 4 channel ADC devices out there, but can you get a board with them that fits your needs? Communications oriented ADC are not generally good for general sampling and visa versa. Moreover the FMC and HSMC mezzanine boards sold by ADC vendors are generally designed to evaluate the device, not be useful in a real application. Be careful.

I'd start with the analog to digital conversion end of your project. This involves the ADC front end through the presentation of the digital data and controls. Just because a board is in an FMC form factor doesn't mean that you can pair it up with any FPGA board having FMC type connectors on it. There's a lot of due diligence to  be done to avoid tragedy in this mine field.

There are a number of vendors that specialize in FPGA based communications and defense products. You might want to at least see what is available as these will be more integrated solutions to specific application requirements. I've spent many years doing the FPGA development board add-on card search... generally, the results aren't good.

Be careful of UltraScale boards as the FMC connector Vadj is generally 1.5V or lower making them incompatible with older FMC mezzanine cards. Be very careful about pin placements for clocks as FPGA devices have limitations with respect to banks and clocking for LVDS. A lot can go wrong. JESD204B or JESD204C converters are a good choice if you can get them on a board with the right front end.

I wouldn't pin your future on limiting your search to Xilinx devices.

You've left out a lot of important information but perhaps I've got you thinking along the right track..

I am so glad that you replied. (I was planning to send you a dm after reading your replies on other posts)

Thanks for your suggestions, I gonna check integrated solutions. (they are very expensive). Also had a sad experience based on that. Because the previous all-in-one board stopped supporting, and had very limited flexibility. 

It's just sad that add-on card is not good...

Thanks!

 

[zygot]

 I can't tell you what to do, but might have some useful tips and experience that could be helpful to avoid FUD and remorse.

Yes, the specialty integrators are expensive. When you get down into the details of a project it's best to assume that you will need to know the platform's capabilities and plan on developing your own IP.

As for Vivado ML, it's no good for me because I don't have the money to buy a complete tool license. Whatever it's capabilities are isn't going to help with Artix development projects. For the Kintex and UltraScale+ ZCU106 board that I own my device licenses don't reach that far. Is it possible to buy a device license from Xilinx at a reasonable price so that I can use a new version of the tools for an older FPGA board? If so, I haven't been able to find out. Regular Kintex might be fine for your application, though timing will be hard. For the ZCU106 UltraScale+ I've executed PCIe and PL DDR4 designs but putting both together in one design is very challenging with respect to timing closure. How many Vivado ML demos have your found on Digilent's board support pages? There was a time when it made sense to pay for tools licenses from Altera and Xilinx... but not for a very long while. The quality of the tools from Intel and Xilinx has degraded to the point that I'd refuse to pay for the tools even if I could afford to.

As far as support. I bought the ZCU106 from Xilinx before Vitis was a thing and Vitis has yet to support it directly.... so the fact that a third party vendor has poor support isn't too surprising.

Be aware that UltraScale+ isn't Series7. There are a lot of differences, particularly with IO banks and advanced IO features. The UltraScale has broken my RGMII interface that I've been using for years with Series7 and earlier devices as well as Cyclone and Max10 devices.

With every new FPGA developer tool version comes nasty surprises, especially if you intend on leveraging old design modules done in earlier versions.

I could go on and on but hopefully you get the picture.  

Edited December 16, 2021 by zygot

[Megumini]

On 12/16/2021 at 4:13 PM, zygot said:

 I can't tell you what to do, but might have some useful tips and experience that could be helpful to avoid FUD and remorse.

Yes, the specialty integrators are expensive. When you get down into the details of a project it's best to assume that you will need to know the platform's capabilities and plan on developing your own IP.

As for Vivado ML, it's no good for me because I don't have the money to buy a complete tool license. Whatever it's capabilities are isn't going to help with Artix development projects. For the Kintex and UltraScale+ ZCU106 board that I own my device licenses don't reach that far. Is it possible to buy a device license from Xilinx at a reasonable price so that I can use a new version of the tools for an older FPGA board? If so, I haven't been able to find out. Regular Kintex might be fine for your application, though timing will be hard. For the ZCU106 UltraScale+ I've executed PCIe and PL DDR4 designs but putting both together in one design is very challenging with respect to timing closure. How many Vivado ML demos have your found on Digilent's board support pages? There was a time when it made sense to pay for tools licenses from Altera and Xilinx... but not for a very long while. The quality of the tools from Intel and Xilinx has degraded to the point that I'd refuse to pay for the tools even if I could afford to.

As far as support. I bought the ZCU106 from Xilinx before Vitis was a thing and Vitis has yet to support it directly.... so the fact that a third party vendor has poor support isn't too surprising.

Be aware that UltraScale+ isn't Series7. There are a lot of differences, particularly with IO banks and advanced IO features. The UltraScale has broken my RGMII interface that I've been using for years with Series7 and earlier devices as well as Cyclone and Max10 devices.

With every new FPGA developer tool version comes nasty surprises, especially if you intend on leveraging old design modules done in earlier versions.

I could go on and on but hopefully you get the picture.  

Thanks for your experience sharing~ I appreciate that a lot~ I might share it later after I get my system running. (but they gonna arrive after 5 months....)