Hi Asmi
(1)
I would stick to the manufacturer's dev boards for the moment, no doubt Digilient will come out with something.
I have ordered a :
https://www.altera.com/products/devkit/po-3196/agilex-5-fpga-e-series-013b-development-kitthere is also a similar for A3C in same form :
https://www.altera.com/products/devkit/po-3000/agilex-3-fpga-and-soc-c-series-development-kitmake sure anythign you buy if used isnt the engineering silicon version. And for those boards, there is a PCI card adaptor so you can plug it into a PC and exercise the PCIexpress caps.
These are the lowest end vendor's dev boards , but appear to be fully supported for examples.
They've got LPDDR4 on them
(2) IP - yeah Altera in the catalog, the Base IP is 6k ( FIR, FFT etc) . If you make a case that you are jumping from Xilinx to become an Altera shop, they might waive this for you as a single developer. Your mileage may vary. There is ALOT of pain moving from Xilinx to Altera and that I think can justify getting this free since I will be the one with pain.
Now I am not so fussed now i think about IP, given how many high level tools there are now of different ilks, as I can write the cores. There are many high level tools to make the HDL for you with agents doing the code work fro your spinal HDL, pythonhdl etc,
However ! is that when you use the IP cores,
vendor's IP cores are highly optimized for their architecture.example : If implementing a FIR at say 300 MHz, with 300 Msps sample rate, you dont have ANY room to waste a clock cycle, so you really have to know what you're doing... Wasteing a clock cycle could mean a 2:1 increase in DSP block usage... and all the extra timing and route issues that might go with that ,,,,
I think the embedded tools in ALtera QPP Agilex are sooo much nicer than the nightmare that VITIS is..... They got a separate company to do it in a closed package. It is easy and uncomplicated. Vitis sure you can do anythign on the planet and the complexity is clear but most of us dont need 90% of that.
(3) Xilinx now with the 5.5k perpetual license cost (but you get everything) is the new world it seems.
5.5k for perpetual enterprise seems like a pretty good deal I think- But Xilinx dont have the devices I want anymore. They're obsessed with the Military RADAR and AI datacentre accelerator revenue.
Even if Xilinx did me a deal, I am not sure that I would stay with Xilinx. Xilinx have done nothign new IMO in the mid range, and ALtera (now) have the best mid range options ever in the industry, i think, from small parts (A3C025) to large devices in Agilex 5E and D. and their devices have modern features (DSPs with FPU, tensor).
I was goign to try and stick it out with Xilinx but the licensing change has given me 2nd thoughts. And as other's have written, where are the new Zynqs

the cheapest you can get a MPSoC ZU1 part asking nicely is still MORE THAN DOUBLE the street price of a Z7010 !! I am hooked on SoC , the hard , fast processor buys you 10x the number of DMIPS compared to a soft processor, and that frankly, buys alot of software time that you dont need to optimize ! My backup was to try and get a rock bottom price on a ZU1 to replace my Zynq designs.
(4) EFINIX.
yeah. LUT4, interesting fabric etc, interconnect seems to work OK. (BTW not all interconnect is through XLRs, there is some localized private routing in groups - see the patent) -- you need to place an extra pipelining stage in there if you want to cross the chip , however I think , excellent stuff. the tools seem to work
and now they have recently introduced partitioned designs - this is new and exciting. and they fixed the interface designer BS so now you instantiate IO ports in your HDL not separate GUIs.
The thing I have found is, Efinix is not all that cost effective unless you are buying lots of trays, they have a fairly steep and seemingly inflexible pricing strategy. The fabric etc is all fast but it needs to be with their interconnect arch. The DSPs will run at 1000 MHz, but they dont have much in extra features (which I guess makes them fast)
. They can do BLOAT16 MUL ACC = FP32 though.....
I like the simplicity of the tools. Their hard and soft SoC is excellent . The SOFT RV64 IP is faster in the real world than their RV32 soft soc with all the branch predictors etc and what I would recommend.
of course, support , well if you need it you need the FAE who knows someone at the factory. They do seem to take notice of bugs and fix them. even if they seem a bit non responsive in the forums
Yes, you need to buy a simulator..... the Quartus Prime Pro built in sim is pretty well equiv to the xilinx I think . so far.....
SU+ is well priced, the packages are a bit annoying for me - 400 ball 0.5mm , or 600 ball 0.8 no like 300 ball 0.8mm package like 7 series., but ultrascale+ is still fast, and the tools and the routing work well, US+ in all variants is an absolute pleasure to work with...... I consider running a 64 bit fully optioned microblaze instead of going the SoC way. But, turns out, Xilinx in SU+ is only supporting LPDDR4X and LPDDR5 (lower VIO) but with all these memory issues, I really just want to stick to using what I have trays of - ECC LPDDR4..... (I use built in ECC LPDDR4 for all my work) . And, since I have been using hard controllers, I cannot be arsed anymore with the hassles of trying to run soft fabric memory controllers fast..
I've also been dabbling in on FPGA fixed burned in LLMs / transformers- where you want LOTS of hard muls and LUT-ROMs( hard weights synthesised into the fabric) . Efinix might be a good solution for the simply, mass fabric requirement there without giant wide data lanes... Efinix with their new 12nm devices, they've been announced for a year but nothign has happened so far.... but they go to 127 Mbits of block ram.....
The efinix advantage is- no money out of tools and IP, (although I would argue that USD 5k on tools once for perpetual isnt a big impost if I play the long game)
but, with Efinix i think that you very much are on your own with just a solid set of build tools. Price wise... Yes you can get more transceivers and useful high speed IOs per $ than Agilex and Xilinx. but in most cases what Agilex and Xilinx give you is sufficient.
Xilinx have trashed their low end student market with the license change, the hobbling of the BASIC tool is just a F___ing joke. Useless debugger, sim, tools.... I am prepared to say "Good riddance" with that. But, Altera is a bit unknown now, just what their longevity is, and the continuously moving understanding of exactly how the tools are licensed- even the local FAE doesnt really understand it . I still dont fully understand their renew etc model, no one seems to, make sure you get that in writing from your FAE so you can wave it around if you have trouble.....
hence the Efinix play....
-glen