GOWIN FPGA

[Piotr Wyderski]

Hi,

The GW1N family looks so much better than the Lattice MachXO2 series,
but they are unobtainium, as far as I can see. How do you get them in
small quantity (<=100 pieces) in QFN? Desoldering them from eval boards
doesn't look like a plan...

Best regards, Piotr

[Rick C]

Why do you say they are not available? Have you tried talking to a distributor?

I have two who have quoted me delivery on the -9 part in the QFP100 package. I'm pretty sure other packages are also very available.

--

Rick C.

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

[Rick C]

On Sunday, March 14, 2021 at 5:08:04 AM UTC-4, Piotr Wyderski wrote:

Oh, I should mention that the documentation is still a work in progress. They have some real crap but at least they update it fairly often.

--

Rick C.

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

[Piotr Wyderski]

Rick C wrote:

> Why do you say they are not available? Have you tried talking to a distributor?

No major distributor has them: Mouser, DigiKey, Edge. They either have
never heard of the chips (they only have an eval board) or label them as
"Discontinued". I need 15 pieces of the 3.3V GW1N in QFN48 for a
prototype run. Not even close to the production stage to consider a
reel, too many to request samples.

What distributor should I talk to?

Best regards, Piotr

[Rick C]

The Gowin web site lists Edge and Rutronik. Both were very happy to provide good prices ~3-4$. Edge has a part number search but it is not very effective. Best to contact someone and put in a sample order. They will sell you 15. Here is an email contact at rutronik, Neal.Enzenauer at rutronik.com. Just let him know how many you will buy the rest of the year. Tell him Rick, the guy working on the ventilator sent you.

Gowin wants $80 for their programming cable. It uses the FT2232HL which you can get in generic cables from FTDI (Digikey etc.) at half that price.

Have you test driven the development software yet? It's a bit rough around the edges still, but at least supports VHDL now.

--

Rick C.

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

[Piotr Wyderski]

Rick C wrote:

> The Gowin web site lists Edge and Rutronik.

Checked both, no luck with an on-line search. Maybe contacting someone
there would help, you're right. Thanks!

> Have you test driven the development software yet?

No, I am still at the chip selection phase. The requirement is a
small-non BGA package with about 30 IOs (which excludes TQFP), which
narrows down the choice to QFN48. This basically leaves 3 viable
options: Lattice MachOX2 640HC, ICE40UP5K and GW1N. About the same price
in low volume, the 640 and the GOWIN are capable of running from a 3.3V
directly, which is a big advantage. The FPGA can be SRAM-based; I am not
going to use the built-in flash.

Lattice Diamond is a big PITA, too; I got used to the quality of Quartus.
And it turns out that 13 years is not enough to implement decent support
for VHDL-2008 in LSE, Synplify and Modelsim at the same time. The
documentation sucks even more, so I don't know if the Gowin's can be any
worse. I would rather use a Xilinx or Altera/Intel part, but they seem
not to care about the low-end applications anymore, where the only fast
part required is a differential Manchester multidrop LVDS link. All the
rest is a boring CIC decimator, an I2C master and so on. It is basically
MCU-like stuff, but the only FRAM-based MCUs are the MSP430s, which are
way too slow for a 50Mbps low-latency link. This is the only reason for
an FPGA here.

Best regards, Piotr

[Jim MacArthur]

Mouser signed a disti agreement with Gowin in December:

https://www.gowinsemi.com/en/about/detail/latest_news/65/

Piotr, what is it about the the GW1N series that makes it better than the MachXO2?

-Jim

[Rick C]

On Sunday, March 14, 2021 at 2:42:32 PM UTC-4, Piotr Wyderski wrote:
> Rick C wrote:
>
> > The Gowin web site lists Edge and Rutronik.
> Checked both, no luck with an on-line search. Maybe contacting someone
> there would help, you're right. Thanks!
> > Have you test driven the development software yet?
> No, I am still at the chip selection phase. The requirement is a
> small-non BGA package with about 30 IOs (which excludes TQFP), which
> narrows down the choice to QFN48.

TQFP is excluded because of the physical dimension? Just too big?

> This basically leaves 3 viable
> options: Lattice MachOX2 640HC, ICE40UP5K and GW1N. About the same price
> in low volume, the 640 and the GOWIN are capable of running from a 3.3V
> directly, which is a big advantage. The FPGA can be SRAM-based; I am not
> going to use the built-in flash.

What about the iCE40 Ultra and UltraPlus which come in the same 48 pin QFN?

Lattice has a very useful selection guide.

> Lattice Diamond is a big PITA, too; I got used to the quality of Quartus.

I don't know why people rag on the vendor tool. I don't spend much time with it. I do most of my work in the editor and simulator. Lattice used to provide Active HDL from Aldec, but now it's Modelsim which I understand is a bit more clumsy for some things. I was working with someone who said he couldn't view the state of variables until he was running the code for that unit. AHDL lets you throw the up on the waveform display as soon as you start the simulation.

> And it turns out that 13 years is not enough to implement decent support
> for VHDL-2008 in LSE, Synplify and Modelsim at the same time.

With Gowin you won't care so much about Synplify I think. Gowin has it's own synthesis tool, but I suppose you can buy the Synopsis tool. For simulation you are on your own. I'm working in the AHDL with the Lattice tools.

It's funny sometimes. There's someone who doesn't like every tool. Lots of folks rag on Xilinx and even Altera tools.

> The
> documentation sucks even more, so I don't know if the Gowin's can be any
> worse.

They are, believe me. Trying to use their math blocks I finally had to reverse engineer the simulation code to see how it works. Rather complex too with all the generics for the variations. Eventually I was able to distill it down to something I could understand.

> I would rather use a Xilinx or Altera/Intel part, but they seem
> not to care about the low-end applications anymore, where the only fast
> part required is a differential Manchester multidrop LVDS link. All the
> rest is a boring CIC decimator, an I2C master and so on. It is basically
> MCU-like stuff, but the only FRAM-based MCUs are the MSP430s, which are
> way too slow for a 50Mbps low-latency link. This is the only reason for
> an FPGA here.

Yep, they go after the real money. X and A have no interest in becoming a Microchip type company selling a bazillion different small parts with low margin. I think there is a market for that. FPGAs are fast enough now that many apps can be implemented on the slowest chips without even thinking about it. The high end will always create demand, but that doesn't mean there isn't also significant demand at the low end. I've read a lot about Anlogic as well, but they don't seem to support English as well as Gowin does. Some of the Anlogic data sheets are only in Chinese.

Keep in touch. If you get to using the programming tools before we do I'd like to hear your results. We are skipping the cable thing and are putting the FTDI chip on our prototype boards. Then we don't have to worry with who has programming cables and who doesn't. The people on this project are scattered around the globe.

--

Rick C.

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

[Rick C]

They were listed on Mouser for a bit, but Gowin was put on the US CCMC list (Communist Chinese Military Company) which creates restrictions in funding. They can become further restricted so that you can't ship your product to various locations, but not yet. I expect this was an issue for a global disti.

--

Rick C.

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

[Piotr Wyderski]

Jim MacArthur wrote:

> Piotr, what is it about the the GW1N series that makes it better than the MachXO2?

The GW1N-1-UV is roughly equivalent to the 1200HC, but for some unclear
reason, there is no 1200 in QFN48. It is available in QFN32, but then
there are only 21 IO lines, which are not sufficient in this
application. This way, Lattice forces me to use the 640HC in QFN48,
which is quite small and does not allow me to use the LatticeMicro8 soft
CPU, as there are only 2 EBRs, for which I already have found a better
application.

I cannot really compare the documentation quality, but that for the
Lattice parts is appealing. For example, after two days of scratching my
head thinking about using the I2C hard cores, I decided to roll my own
in VHDL, which took me 6 hours. The PFU documentation also lacks
details, e.g. I still don't know if I can access the intermediate carry
chain or not. The carries in a ripple counter make perfect clock
enables/triggers -- but is the tool wasting some resources to emulate
the "true" carries for me? No idea. It makes a difference if you have
got 256 or even 640 LUTs.

Best regards, Piotr

[Piotr Wyderski]

Rick C wrote:

> TQFP is excluded because of the physical dimension? Just too big?

Yes, precisely. The BGA/CSP would make the PCB too expesive -- and I am
BGA-phobic anyway...

> What about the iCE40 Ultra and UltraPlus which come in the same 48 pin QFN?

These would certainly do, but I have found the QFN48 variant a bit too
late. They require an external regulator for V_core, BTW. :(

> Lattice has a very useful selection guide.

Mouser has even better. ;)

> They are, believe me.

I thought that the Lattice docs mark the bottom. Not sure if I am
willing to pursue the Gowin direction then, time will tell.

> Keep in touch. If you get to using the programming tools before we do I'd like to hear your results.

Sure. For some time I will be playing with the MachXO2 series, as I have
got an eval board.

Best regards, Piotr

[Rick C]

When I have questions like that I take the section of code and synthesize it, then look at the result in the RTL schematic viewer. Even Gowin has one of those.

Once I was getting two carry chains when I expected one only to find out I had a very small discrepancy in my coding that made them distinct. Fixed that and reduced the LUT count by dozens from the many places I was using that method.

--

Rick C.

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

[Rick C]

On Sunday, March 14, 2021 at 5:32:42 PM UTC-4, Piotr Wyderski wrote:
> Rick C wrote:
>
> > TQFP is excluded because of the physical dimension? Just too big?
> Yes, precisely. The BGA/CSP would make the PCB too expesive -- and I am
> BGA-phobic anyway...

I used to be that way. Recently someone on eevblog convinced me to consider them as the cost difference is gone for the 1 mm and even 0.8 mm ball spacing parts. LCSC sells the Spartan 6 at incredible prices and the BGA256 may be too big for you, but it fits in the 100 pin QFP my board uses. I would probably only use half the IOs or less, maybe a third. Xilinx has the Spartan 7 in a 196 pin BGA with 1 mm ball pitch if I remember correctly, so smaller than the 256 a bit, still not 48 QFN size though. I've only found Mouser pricing which is not realistic if you build more than a couple thousand.

If I have to spin my current board because of the Lattice part price increasing, I can't use the Gowin part because of being on the CCMC list. So the Xilinx part looks very attractive. I'd need to find a way to get a pricing quote they will honor for years. One really good thing about Gowin is the Edge web prices are not much different from the quotes.

> > What about the iCE40 Ultra and UltraPlus which come in the same 48 pin QFN?
> These would certainly do, but I have found the QFN48 variant a bit too
> late. They require an external regulator for V_core, BTW. :(

Is the Vcore really that big a deal? It's $0.25 part, no?

> > Lattice has a very useful selection guide.
> Mouser has even better. ;)

I've been down this road and other than the Chinese startups Lattice is the only vendor supporting the low end. So their selection guide has red boxes around the "good" parts to make them easy to find.

> > They are, believe me.
>
> I thought that the Lattice docs mark the bottom. Not sure if I am
> willing to pursue the Gowin direction then, time will tell.
> > Keep in touch. If you get to using the programming tools before we do I'd like to hear your results.
> Sure. For some time I will be playing with the MachXO2 series, as I have
> got an eval board.

All the other small fish and even the big fish in the FPGA pool have been bought up. I wonder how much longer Lattice will swim free?

--

Rick C.

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

[Piotr Wyderski]

Rick C wrote:

>>> What about the iCE40 Ultra and UltraPlus which come in the same 48 pin QFN?
>> These would certainly do, but I have found the QFN48 variant a bit too
>> late. They require an external regulator for V_core, BTW. :(
>
> Is the Vcore really that big a deal? It's $0.25 part, no?

Real estate is. That would be another part with non-trivial cooling
requirements. The QFN48 has a huge thermal pad, which would help me a lot.

> I've been down this road and other than the Chinese startups Lattice is the only vendor supporting the low end. So their selection guide has red boxes around the "good" parts to make them easy to find.

I have absolutely no problem with Chinese parts, no national pride hurt
here. ;-)

Best regards, Piotr

[Rick C]

On Monday, March 15, 2021 at 2:51:35 AM UTC-4, Piotr Wyderski wrote:
> Rick C wrote:
>
> >>> What about the iCE40 Ultra and UltraPlus which come in the same 48 pin QFN?
> >> These would certainly do, but I have found the QFN48 variant a bit too
> >> late. They require an external regulator for V_core, BTW. :(
> >
> > Is the Vcore really that big a deal? It's $0.25 part, no?
> Real estate is. That would be another part with non-trivial cooling
> requirements. The QFN48 has a huge thermal pad, which would help me a lot.

I've never designed an FPGA where I had to give thought to the power dissipation other than to estimate so I could say it was estimated. Are you really going to be cranking on it so it needs real current? Dropping 2 volts even at 100 mA is not much more than an 0603 resistor can handle. What's a reasonable dissipation for an SC-70 regulator? I've never considered how much heat an SC-70 can get rid of. You did say you could consider a 640 LUT device, right? I would point out there is a reason why they call them iCE40 devices. They really are low current.

> > I've been down this road and other than the Chinese startups Lattice is the only vendor supporting the low end. So their selection guide has red boxes around the "good" parts to make them easy to find.
> I have absolutely no problem with Chinese parts, no national pride hurt
> here. ;-)

The key word there was "start up". Gowin seems to be doing ok, at least until they got on the CCMC list. Don't know how that will impact them. The others seem to be focusing on markets where they are not so visible. Their docs are a lot less English too, not to mention support, tools, etc.

There's one that is a bit different that *is* with either Mouser or Digikey, can't remember which and I can't remember their name. I think they might have a larger QFN, but maybe not because every time I consider them they drop off my list. They do the is it logic or is it routing thing. So you don't get to use all of the LUTs. In theory it means lower price since you aren't paying for as much routing that you won't use. I remember years ago the FPGA FAE chant was, "We sell you the routing and throw in the logic for free". I don't see this in this company's prices, but they are not real prices as we all know. You have to get a quote to know what they will really sell their chips for.

--

Rick C.

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

[Rick C]

Here is an example of the poor quality of the Gowin docs. In fact, this one is pretty extreme. UG166 lists the pin outs of different family members in each package. This document shows a consistent match between the designators TBLR as in IOT21A being bank 0, 1, 2 and 3 (not necessarily in that order). But in UG114 the they list IOT I/Os as being part of banks 0, 1 and 3! If you are not running all I/Os at the same voltage this would be a big issue!!!

We just caught this and I've sent an email to the FAE. He usually gets back within 24 hours.

--

Rick C.

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

[Piotr Wyderski]

Rick C wrote:

> I've never designed an FPGA where I had to give thought to the power dissipation other than to estimate so I could say it was estimated. Are you really going to be cranking on it so it needs real current? Dropping 2 volts even at 100 mA is not much more than an 0603 resistor can handle. What's a reasonable dissipation for an SC-70 regulator? I've never considered how much heat an SC-70 can get rid of. You did say you could consider a 640 LUT device, right? I would point out there is a reason why they call them iCE40 devices. They really are low current.

OK, I am abandoning the Gowin direction. iCE40UP5K wins hands down and
the volume is not likely to be big enough to pay for the NRE. If I am
wrong here, Lattice will probably make me a special offer... ;)

Best regards, Piotr