ice40 Hx8K project not configuring from external flash

[woofy]

The hx4k is available in a TQ144.  I've heard it's an hx8k inside, but never used it and no idea if it's true or not.

[julian1]

The hx4k is available in a TQ144.  I've heard it's an hx8k inside, but never used it and no idea if it's true or not.

That's interesting.
I noticed in documentation, that the bitstream sizes (generated by icestorm) are the same size.
A quick grep through the nextpnr code on the arch flags for ArchArgs::HX8K and ArchArgs::HX4K doesn't show obvious handling differences.

For hx8k packages, the 256 pin bga is 0.8mm pitch versus 0.5mm pitch for cb132 and would be more tolerant of a home reflow process.

[SiliconWizard]

If it's the same, then yes the only way of using it as such will be with yosys/nextpnr. I wouldn't be surprised - this is relatively common.
Another example of this with Lattice is with the ECP5. The -12 and -25 are the exact same die. The only difference is the internal chip ID - so if you use Lattice tools, they won't let you use a-12 as a -25, but with yosys/nextpnr, no problem!
(And it's about half the price.)

[woofy]

I think it was David Shah (nxtpnr) who first discovered this but I can't find a reference now. Lattice tools  will not generate hx8k code for the hx4k but nxtpnr could, though not recommended.

[SiliconWizard]

Can't speak for the hx4k, but for the ECP5, it *is* the exact same and there's no difference. So, if it "not recommended", that would be by Lattice for obvious reasons.

[julian1]

With the tq144 package - a combined eda symbol can support both hx1k or hx4 parts (some 1.2V supply pins get managed with jumpers).
This may be useful if the firmware complexity is not known upfront, or else for product scope changes, without having to change the pcb.
If the hx4k actually has the same dff+logic resources as the hx8k, this brings more flexibility.
Although the dependency on the open-source toolchain, will be a limitation for some.

[SiliconWizard]

Yes, yosys/nextpnr isn't for everyone and still has limitations for sure. (Just for the record, for those who use VHDL, you can use it with yosys and a ghdl plugin. Works fine. So it's not limited to Verilog either.)

Now of course, while (at least for the ECP5, maybe same for the hx4k) it *appears* to be identical dies to the more expensive models, there's of course no guarantee that it will work reliably. Even if it's indeed the very same dies, for all we know, the manufacturer may sort the dies during production and sell the ones that pass tests for the lower number of logic blocks but don't pass tests for the nominal number. Something that *some* manufacturers do with *some* chips. OTOH, they may not do any sorting at production and just distribute the dies according to marketing figures. We just don't know, unless someone works at Lattice and can tell us, but I doubt they'd be allowed to.