Some preliminary help for an FPGA selection

I would love to recommend an FPGA for you, but I'm no expert on PCIe or LVD S. They both seem pretty important to your needs.

For low power you might look at a Lattice product. The iCE40 parts are int ended for very low power, but might not be enough for your needs. They hav e the LVDS but I don't recall any PCIe or memory interfaces as hard IP. I' m sure they have an SDRAM interface as a soft core.

They have a variant of the iCE40 line with various hard IP, but I seem to r ecall it is oriented to some particular interface for embedded stuff, video maybe. I think the name is Cross... something. The iCE40 line has RAM co nfiguration memory with non-volatile backup, one time programmable.

They also have a line of RAM configurable parts not unlike the Xilinx line (that's what it is derived from, the old Orca products licensed from Xilinx by Lucent). I seem to recall they have a new product line in that technol ogy. Lattice is focused on low power, so that might be a good place to sta rt looking.

The tools are free, but you need to obtain an annual license. I guess they want to keep control... which is what they all do, even if you buy a perpe tual license. I learned that back when I bought my first Mentor system, ci rca '90. Even though we paid big bucks for it and the software was usable forever (only the maintenance required annual payments) we had to ask for a license each year.

I only use the tools when I have a current job at hand. It seems I sit dow n on a Friday night to use the tools and find the license is expired but ca n't get a new one until Monday! lol They may have automated that so it pr ocesses on the weekend, but I'm not sure.

You probably have more interesting things to do on the weekends.

--

  Rick C. 

  - Get 1,000 miles of free Supercharging 
  - Tesla referral code - https://ts.la/richard11209

fredag den 7. august 2020 kl. 23.35.32 UTC+2 skrev Dimiter:

The free ISE webpack works with spartan6

My advice would be to download the free FPGA tools (Quartus, Vivado, Libero, Diamond) first and see which ones offers the high-level (free) IP you need (PCIe, DDR controller, HDMI, Display controller, FIFO etc). Then run through a small design to see if you like the IDE, when you are happy build/simulate/P&R your design and if it is all working (probably next year) then start looking for a board. There are many to choose from and some offer a custom route as well if you need it.

Like many FPGA engineers I have a box full of FPGA boards from all vendors which I bought for "projects" I never completed, the fun part is always buying the board. I now have the Arduino bug and a second box is slowly filling up with all sorts of breakout boards....

Good luck,

Hans

That sounds like quite an ambitious project for a first FPGA design. How to go at it depends a lot on your other experience - if you are OK with designing PCIE and DDRAM hardware, 400+ 0.8mm pitch pin BGA then working at chip level is OK. If not then use a ready made module with at least the FPGA, PSU, DDRAM and maybe HDMI on board.

For this I would probably go with Xilinx, Spartan 7 or Artix. The free tools do a lot, and there is IP for dev boards from Digilent or Trenz Electronic gmbh.

updates for a year but it goes on working forever.

I actually prefer Altera or Lattice software to Vivado (its just mind bendingly complicated - get training - its worth it). Unfortunately the general support for Xilinx parts (IP, dev boards, general community help etc) in this area is much better.

There are some nice cheap dev boards for Artix and Spartan 7 (Arty and Artys7) - we used them (and Digilent Nexus) when we did a an Artix based design as a sanity checker.

MK

Oh look, someone has done it all for you already !

(Might be worth a look for ideas - and unless you need quite a few will be cheaper than rolling your own !)

There are other references on the web, search 'AX7103 fpga'

MK

MK

Thanks everyone for the pointers and suggestions. It will take a while of looking and head scratching until I can say anything with certainty but at the moment I am looking at spartan 6 and spartan 7 (yet to look at artix).

Looks like spartan 7 does not have a hardwired DDRAM thingie; how much of a problem will that be? I have yet to evaluate the tools but I suppose I will be able to write low level logic, sort of this.d = that.q & whatever, I have no idea what the syntax will be but I am sure everyone gets what I mean. Once I am there I'll find my way (and I fear getting there will not be easy, especially talking to hardwired stuff like DDR and PCIe). But "not easy"is not a problem, I fear the "not possible", especially finding out the latter too late (a friend once quoted a friend of his - both engineers - about the "oh shit" factor being applicable to every project; you either say it at the beginning or at the end... :-).

As you see I am wondering about the basics at this stage, I'll be grateful for any hints from you people with experience before I start downloading toolchains and wrestling to evaluate which and how...

Thanks,

Dimiter

Try a few simple projects in the tools to see which you prefer. Maybe that won't matter since you seem to be zeroed in on the Xilinx parts.

If you are going to do a serious project, maybe start with a few simple pie ces. You can design the RAM interface along with a test bench to test it w ith. In fact, design the test bench first since that will help you learn a bout the RAM without worrying about synthesis. Then you can write the RAM interface code and test it to the test bench knowing what you need it to do . This code will need to be synthesizable. That's not a huge issue, but i gnoring that in test benches makes them easier to design.

None of this requires hardware or downloading or any of the BS you should n ot even think about until you have the design fully coded and verified in s imulation. I don't know your background, but many software people give sho rt shrift to simulation. In FPGAs it pay huge dividends making debugging 1

00 times easier than on the bench. I've completed projects that needed har dly any debug time on the bench because the design was 99.99% working after simulation.

Have you picked a language yet? I like VHDL mostly because that is what I' m used to. It is strongly typed, so requires dealing with conversions. It requires you to be explicit so nothing is assumed. Verilog is much less w ordy, but has defaults you need to know and understand or it may not give y ou the results you expect.

I keep saying I'm going to learn Verilog and Linux but never do.

--

  Rick C. 

  + Get 1,000 miles of free Supercharging 
  + Tesla referral code - https://ts.la/richard11209

One choice is to use a small FPGA module with DRAM and other interfaces alr eady provided (along with RTL examples of their usage). The module should use a FPGA suitable to your needs and should be from a vendor that provides long term (several years or a decade or mode) of availability. Then you h ave a simpler PCB design situation: provide the connectors to plug in the m odule.

Xilinx support DD3RAM up to 800MHz on Spartan 7 - it doesn't much matter to you how much of the support is hard wired and how much is FPGA gates since you just use their IP (free I think for DDRAM on S7). And you will use their IP for DDR3RAM because it represents multiple man years of work.

I agree with Rick about the importance of simulation - but I think you should buy a cheap dev board (like the Arty if you go for Xilinx.)

For two reasons, one because there is quite a lot between a working VHDL or Verilog simulation and hardware (and its good to learn that from day

1) and secondly because it keeps the interest up when stuff actually works. (But don't ever bother even trying stuff on the chip until the best simulation you can do is good.)

MK

I'd certainly explore such an opportunity if I come across something like that - which comes with the sources for the fpga etc.

Dimiter

Thanks Mike, knowing it does not matter hard or soft DDR was important. I really don't want to go into implementing DDRAM myself (I would if I had to but this must have been done by many people already and I don't need the exercise....).

I am not going to use a board, I never have. It will have to run on my board anyway, what is key is that I can upload multiple iterations without wearing out some flash or something; for the coolrunners I have managed to never damage one, i.e. within

20-30 iterations, that utilizing at times 100% of the cells, but this is significantly more logic (apart from the PCIe and DDR interfaces not that much more but well, I can't go without these two, at least not without PCIe (I could use some static RAM). At the moment I even don't know how the spartans are uploaded, is it via jtag or something else (I.e. I have not yet read much). I will have a few pins of the processor (a t1042) to jtag the fpga at least for boundary scans, if I cannot upload it using these well, I'll do what it takes. Key is the processor being able to upload the fpga with the contents of a file from its file system.

Dimiter

r

l

rk.

L

rks.

You can use JTAG to configure a Xilinx part I'm pretty sure, but they have a configuration interface that can be as simple as four signals. The proto col is very simple but there is little feedback on what is wrong if it does n't work. The upside is you only have maybe four things to check to be sur e it will work.

I went through this once at a company I worked for where four engineers (tw o were FPGA designers and one was a manager) could not bring up a new FPGA board I designed. I pointed out the handful of things they needed to do an d they assured me they had done them all. Then I asked if they did them al l at the same time. The guy programming dinked the code to make sure they

I said, "let me know if you need anything else" and went back to my desk. lol It was actually pretty funny, but clearly too many cooks were spoiling that broth.

By any chance are you a software guy? A software guy showed up in one of t he FPGA groups some years ago and we all tried to tell him it would be hard to adapt to the differences between sequential languages and HDLs. He sai d he had tons of room in a spacious FPGA for what he wanted to do and perfo rmance didn't matter. So we advised him, the whole time telling him how po orly it would work. In the end he produced very workable code and complete d his demo task for a customer.

So I don't tell software guys how hard or different HDL is any more. In re ality it's just another programming language. Well, it is a bit different, but not anything a decent teacher can't explain. The guy I learned VHDL f rom was terrible. On top of that the Orcad VHDL tool was so bad they ended up abandoning it a few months later. lol I switched to the Xilinx provid ed tool and later had to switch again because Xilinx switched to another br and. Three tools on one project, but I still managed to finish on time.

BTW, that project was my first, but it taught me the lesson of how importan t simulation is. There were two modes of using the board, recording to the hard drive and playing back from the hard drive, separate configurations, separate designers. I was ready to test on the hardware first and only fou nd a couple of bugs and was done. The other guy took a couple of weeks lon ger in design and spent a couple of weeks debugging on the hardware. Just one anecdotal example, but I see this over and over. An extra week in simu lation saves a number of weeks on the bench.

Simulation is also so much easier than trying to capture signals from insid e a chip.

Thanks for listening. I just came in from doing lawn work and this gave me a chance to unwind.

--

  Rick C. 

  -- Get 1,000 miles of free Supercharging 
  -- Tesla referral code - https://ts.la/richard11209

snipped

If you don't need the RAM to be that fast or that big then you could use SDRAM, it'll save you a lot of trouble, both in board layout and FPGA work.

I wrote my own SDRAM controller a long time ago (when quite new to FPGAs) - not too big a deal at all.

DDRAM is a different thing altogether (reading and writing data in a

300ps window - its a miracle that it ever works !).

Using a pair of 16 bit wide SDRAMs at 100MHz gives you 400 Mbytes/s - chicken feed by DDRAM standards but maybe enough.

The signals are slow enough for scopes and logic analysers that normal people can afford.

MK

Just about - 1080p/60Hz/24bpp is 373MB/s. If you cut the colour depth down you can reduce that bandwidth.

However you also need bandwidth to write into that RAM - you'll need to measure how much you need for that.

On PCIe, that means you're in mid-level FPGA territory as it needs high-speed transceivers. So not the bottom-end Cyclone V E but the fancier Cyclone V GX. And similar across all the vendors lines - that means you're heading towards the top end of Lattice's offerings.

One thing to be wary of is that Intel/Altera at least offer time-limited demos of IP cores that you don't have licences for. The device will work while plugged into the programmer, but not work standalone. There's warnings about this, but you may be able to successfully synthesise despite not having a licence (which might actually be fine for proof-of-concept/eval purposes).

On the system-on-module front there are lots of vendors but I've personally worked with

who have a decent range, although possibly higher-end than you want.

For LVDS I'd think about buffering - you might want a proper line driver to drive a long cable, rather than FPGA I/Os which aren't designed to drive into a long capacitance. We used a third-party HDMI module once and had a lot of problems because the registers in the driver chip (made by ITE) weren't documented - in the end we recorded a log of what the demo code set registers to and replayed that in the same order in our code, a hack but it just about worked. Instead of proper HDMI (audio etc) you might think about DVI-D: HDMI monitors can accept it but it's a simpler protocol. You can also get chips which take in parallel RGB (6-8 bits per channel), Hsync, Vsync, pixel clock and output DVI or HDMI. That means you just generate something that looks like pre-DAC VGA and the chip handles the rest. You might also consider EDID monitor detection, depending on the application.

What I'd do is look for a PCIe-card-shaped dev board, which should come with examples for driving that. For the video side, worst case you can make a resistive-DAC VGA port off some GPIO pins, but you could look at other options - for example wiring a parallelRGB-DVI chip off GPIOs. Maybe the board will come with HDMI or Displayport already (and some demo come), but that's rare on PCIe-card boards. I think at that point you have to live with whatever RAM type the dev board gives you - if SDRAM works for you it'll save some troubles, but I've seen a few where there's only one 16 bit SDRAM, which is low on the bandwidth stakes.

Dev boards sometimes come with a one-seat licence for the tools too. You might also be able to get access by speaking to a sales rep.

Personally I have a dislike of Xilinx tools and I'd prefer Intel or Lattice, but I have more experience of Intel in particular, so I may be biased. On the Intel front, 'Quartus Lite' is the free tools and supports Cyclone, so I'd be looking at a Cyclone IV or V in GX version. For example:

Meanwhile, in Lattice land (tools licence free only for this board):

Theo

More thanks to everyone for the hints.

Looks like I am going to use some of the Artix devices (Spartan 7 have no PCIe as far as I get it, unless they can do it "soft"). I would not go old style SDRAM, the bandwidth would be limiting (I need at least twice the bandwidth needed for refresh if the update via PCIe will work without visible artifacts; much more if I go to make alpha etc.). DDRAM3 is fast enough and as Mike suggests at 800 MHz clock it is a miracle things work but here we are, I have to use that or fall too far behind. Then there will be DDR-4 or 3l on that board for the CPU so it has to work anyway. Last time I used SDRAM was .... 20 years ago, can't really believe it was that long ago.

- another 5 (just 4 mounted) SDRAMS on the bottom side, 64 bit wide in total - this is the second board put together, the first one worked for may be 3 months and fell apart, I had reflown it at least 30 times, was my first board with BGAs on it).

Doing straight RGB as Theo suggests might come to rescue during the design of course, but since HDMI is doable it will have to be done. No MIPI on the agenda yet, they are too secretive so I'll live without a camera at least initially. Or use a USB one.

I downloaded the VivaLdo (I really spelt it Vivaldo having encountered the name first just here in Mike's message, got corrected by google... did not ask for any seasons, luckily :), got a free license for their webpack. At first glance it looks like it is sufficient and free but I will not route in an fpga before I have it at least compile without asking for fees etc., as Hans suggested (or something in that line). Was an adventure of its own,took about 2 hours to download, but it seems to work (not that I know yet what the difference is between vitis HLS, Xilinx Vitis, VivaLdo... DocNav - well, I think I understand what this is :).

Thanks again for the hints, please keep them coming. And please keep your thumbs pressed, I have done crazier projects (at their time) than that (which is by far not just its fpga part), but I don't know if I still have the drive it takes - hopefully I can find it.

Dimiter

====================================================== Dimiter Popoff, TGI

======================================================

PCIe is not something you'd want a soft core for. Spartan7 not only has no PCIe, it has no high-speed transceivers. Artix-7 would be the better choice from Xilinx. I believe the tools are better and the cost lower for a device that has the resources you're looking for. It is supported by the free tools as I recall. If you go with Xilinx, you'd want Vivado for this project. Other tools are more oriented toward special applications like AI (neural networks, deep learning...). Also the latest tools are really made to support the really expensive newer chips in the UltraScale++ family or whatever they're up to now.

Good Luck!

--
Gabor

I remembered the part from Lattice that had the fancy-dancy interfaces, LFD

2NX-40. They call it Certus-NX. There are only two sizes and the larger o f the two at 39 kLUTs has the hard IP PCIe interface. It's only 1 lane and one controller. It also has SGMII (not familiar with that) with 2x2 lanes maybe. DDR Memory is supported, LPDDR2, DDR3/3L.

--

  Rick C. 

  -+ Get 1,000 miles of free Supercharging 
  -+ Tesla referral code - https://ts.la/richard11209

Thanks Rick, they do look interesting indeed but I saw no mention of a "free toolchain" so there probably is none (I did not dig though). I have made up my mind on the Artix device, the webpack tool says it will do all I need for free (the only downside being my code will be defacto public but I don't mind that for this project). So it will be this, I just need a few more weeks to clear up stuff and dig into that VivaLdo (with or without the four seasons... :) and do it.

Dimiter

LFD2NX-40. They call it Certus-NX. There are only two sizes and the larg er of the two at 39 kLUTs has the hard IP PCIe interface. It's only 1 lane and one controller. It also has SGMII (not familiar with that) with 2x2 l anes maybe. DDR Memory is supported, LPDDR2, DDR3/3L.

They are often happy to give out tools. It is possible you might need to t alk to a salesperson about the tools for this device. Don't know, I've nev er bothered with any of the fancy parts.

I'm not clear on the "public" thing. Are you saying Xilinx insists your co de be "public"??? Or is this the thing where they want to have access to y our code so they can analyze the quality of the tools?

I guess if the "public" thing is a problem the tool can be used when off li ne and the files can be sequestered on a removable drive when you want to c onnect to the Internet.

--

  Rick C. 

  +- Get 1,000 miles of free Supercharging 
  +- Tesla referral code - https://ts.la/richard11209