Xilinx C1100 FPGA (ETH & KASPA)

[Andrew_Carr]

I just picked one of these up and wanted to do a write-up since I haven't heard much about them and info is kinda scarce.

Pros:
-Very efficient at mining ETH (on par with an A11 Pro ASIC roughly at around 1.25W / MH)
-Can be reprogrammed for RVN, etc. algorithms
-Open source bitstream from xilinx (60MH/s), closed source bitstream from teamredminer (70-80MH/s) at 90-110W.
-Firmware available from teamredminer as well

Cons:
-Expensive ($1300 or $16/MH/s at MSRP)
-Availability (50 week lead time suddenly, was 4-6 weeks previously)
-More confusing initially than GPUs
-Cooling
-No cables or manuals included in the box, simply the card and some foam
-extra USB cables are annoying to deal with
-4% dev free from TRM

As far as setup guides go, team red miner has a good guide here for all the software setup needed:
https://teamredminer.info/documentation/fpga-guide/
https://github.com/todxx/teamredminer/blob/master/doc/FPGA_GUIDE.txt

Hardware setup is similar to a passive server GPU. It's single slot, half width and takes 1 x 8 pin PCIE power cable (optional but needed for mining) and one micro-USB to USB A for data output. Data over PCIE is possible but not with teamredminer I believe. You need to force airflow in to cool it. I tried the setup here (https://www.reddit.com/r/FPGA/comments/rh5pz6/having_trouble_cooling_a_data_center_fpga_card_i/) but my temps with it started hitting 73C so I'm waiting on a 3d printed fan bracket to arrive and then I'll mount a better fan to it as seen here:

Thanks to teamredminer these can be tuned pretty much just like GPUs in terms of overclocking and undervolting, although you'll want to update to their firmware for better undervolting. They do have some basic limits in place so you don't destroy your card, and it looks like they're pretty difficult to brick. The xilinx bitstream isn't great out of the box but at least you'll always have an open source option and other devs will hopefully work on these as well and bring down the dev fee in the future. I ran mine on hive on an existing rig (no GPUs or other cards, a CPU mining rig). All I had to do was switch the flight sheet to one with teamredminer, wait a few minutes, and it installed the bitstream in the background and started mining. Xilinx has separate instructions on their website for their bitstream and so install videos but I skipped that since performance isn't as good.

Once I'm able to install a better fan I'll do more tinkering and try to report ravencoin hashrates, but overall setup was pretty seamless. Some of their other FPGA models seem to be more finnicky when it comes to voltage and other things but so far these ones seem to be fairly idiot-proof.

 

[Andrew_Carr]

Are you controlling the fan speed or running it at 100%? I was planning on just running it at 100% but I saw a fan speed controller in someone else's video and they mentioned running it at around 68% fanspeed and getting temps around 45C (probably 72F ambient). I'm not sure why xilinx went with a single slot design on these though since a larger heatsink would be beneficial and all HPC server cases I see are designed around dual slot cards. Maybe they underestimated performance or had demand for single slot solutions or something. A waterblock would've been nice in that case since they're already kinda specialized. Kinda tempted to take mine apart and see what old GPU heatsinks might match up.

Long-term, I don't see TON mining being useful and while in the future ETC is still something to mine, you'll be competing against ASICs there and won't have any efficiency advantages most likely (it seems like ETH ASIC design slowed down due to the 2.0 merge news but Jasminer has some 5GB ASICs that are much more efficient). I hadn't seen anything on ravencoin but had heard there was a bitstream for it. I figured no one had bothered posting results yet since it's not the most profitable thing to mine currently and performance is usually roughly half of the ETH hashrate anyway, but I guess if there's no bitstream yet that explains things, lol. Have you tried removing the heatsink and repasting/repadding? I doubt it would make a difference with the tiny heatsink though. At $1300 a year ago this would've been a great product vs $1500 3080s or something. Right now, on a cost/hashrate basis it's not great. It's about $300 more than what you'd want (1.3 x $800 example 3060ti non-LHR. Above $10/MH) and this could get way worse if GPU prices tank. On an efficiency basis it's ok, but with ETH going away soon* (who knows) its future is unclear, since worst case it'll be competing against more efficient ASICs on ETC.

Best case, it looks like hashrate will max out at around 80MH/s. Currently the best I've seen is roughly 80.5MH/s at around 102W + fan. Maybe the power usage will be improved but it doesn't seem like by much.


[Edit: Found a picture of a disassembled one. Looks like a GPU waterblock might need a 30mm x 30mm or 40x40mm so shim to avoid the frame around the die.]

Here's a close-up on the fan shroud design someone else did. The rest of it looks pretty much the same as yours.

 

[Andrew_Carr]

No Im running it at 60% controlled via PWM. These 1.85A fans start to really vibrate around 40%, 100% speed is going to break something eventually just due to the extreme vibration/resonance. I can actually feel the vibration in my full 42U rack, it's really making me nervous. But Ive got some stuff arriving this week to test some better thermal solutions.

I had to immediately re-paste them, as xilinx has assembly problems with these cards and it's very common that the heatsink is torqued improperly at their factory, leading to cases like this where the HBM isn't even making contact with the heatsink.

And yes there is a brace around the die/HBM that is above the die height. You can see the machined recess in the stock heatsink.

That... doesn't look very ideal, guess I'll take mine apart today while I wait on my mount to arrive. I have a random assortment of 0.46A, 1.4A, and 2.6A fans in that size so maybe I'll start with the 0.46A first and see if it's enough.

 

[cdabc123]

That's interesting. Since this is a xilinx fpga im assuming its documented well enough to do dev work for it independently. Unlike many other mining fpgas. Honestly I would love to get my hands on a hbm fpga thats a cool device. I've been meaning to stretch my legs outside of basic digital logic design to some performance intensive stuff.

 

[Andrew_Carr]

Glad I took it apart. Looks like the same problem with the thermal paste. Used a knife to save the warranty sticker but otherwise disassembly was easy.

 

[Andrew_Carr]

Alright, my mount arrived today and I tried out a couple of fans. The 0.46A fan wasn't enough, but the 1.70A fan (Nidec Gamma30 model) seems to be just barely enough at 90F ambient. The heatsink design does seem really terrible though. I think even something cheap like that blue one you're showing would've worked better. Even though this fan puts out a ton of airflow, by the time it gets out of the heatink it has slowed down considerably.

I guess because they really wanted a 1 slot design they had to compromise and direct some of the airflow out of the top, which doesn't make sense to me since it'll just blow hot air around inside whatever case it's put into. I might try it in my supermicro GPU case to test since it has high speed exhaust fans behind the GPU backplate, but even with that much airflow it struggles to cool a passive vega64 so I'm not sure if it's even worth bothering with this.

So far no vibrations at all really. xenium, does your fan mount screw into the three screw holes on the right side of the card? I screwed those in pretty tightly (see my previous pic) and then installed it and I'm not really getting any vibrations through the card. Currently getting 74MH/s at stock settings, will play around with tuning later and hopefully reduce temps and increase hashrate.

Testing an added heatsink + .5mm thermal pad. The "little" 15mm thick heatsink is definitely working and helps shave off a few degrees, but it really needs a proper shroud as well to channel airflow through. I have a 30mm heatsink coming tomorrow to test which I had a feeling will be a much, much better solution.

Also played with fan settings on the 1.85A B97 fans. These have a bad vibration zone around 40-60% fan speed and rattle the hell out of the card. I raised them to 70% and it made a huge difference.

I just wish there was an axial fan/heatsink solution for these like the FK fpga's from SQRL. These cards are great, but the shitty heatsink is their Achille's heel.

 

[Andrew_Carr]

Yeah, the model number on the 1.7A fan is A34710-58. The 0.46A fan is A34123-55. They came out of some old 1U servers. Does yours have slits on the side that go over the PCB? Maybe that helps with the vibrations. Here's his install guide if that helps: https://files.catbox.moe/gxw78t.pdf But your design looks really similar so maybe it's just the fane models being different.

I'm not really sure why they would design it for such a low temperature range since even datacenters are starting to run at higher temperatures to save on cooling. I guess if you thought that 60MH/s was the best the card could do then you wouldn't think it would have problems running at higher temps.

 

[Andrew_Carr]

Yeah, that might work better. The fans should have way more airflow than is really needed for these, it's just with how the heatsink is designed a lot of seems to be wasted. I have a bunch of 0.5mm thermal pads so I might strap a old heatsink as well.

 

[cdabc123]

I know its a barrier with the raised die edge but a fpga like this would be really nice under a waterblock. Maybe just find a small enough copper block that it can fit inside of the edges.

I'm abit worried about the power delivery stage of these. It may be worth looking into the components used to see if it would really be comfortable pushing 125w. Fpgas are weird. They dont have safeties built in and you can push most fpgas far beyond their power or thermal limitations with the right type of workload. This was a problem with Chinese based boards using off the shelf fpgas that underestimated power delivery. It can begin to effect logic latching and precision after a point.

I would have less concerns about a xilinx ultrascale oem board. However xenium is right about beginning to see hbm degradation past a point. Especially with the frightening die contact.

 

[Andrew_Carr]

New mod works great. Even with a crappy 6w/mK thermal pad, it still works incredibly well. Going to test a better thermal pad soon. The biggest advantage here is running a lower fan speed and saving 10w+. The stock heatsink fins are so restrictive, there's a lot of diminishing returns as you turn up the fan speed, at high RPM. It just phsyically can't move enough air through the card, and the difference between 50 and 100% fan speed is minimal.

Will probably be selling these shrouds as a kit if anyone is interested.

View attachment 453333
View attachment 453330View attachment 453331

Do you have the dimensions on the heatsink too? Apparently these can be pushed to around 250W (if cooled) and would do well in dual mining ton + eth so having the extra cooling capacity would be nice.

 

[Andrew_Carr]

Just came across this: https://www.ospreyelectronics.io/product-page/c1100-heat-sink
GPU style heatsink for the c1100. Kinda pricey, and I wish it didn't extend above the board so much, but seems like it might be effective. My 3D printed fan mount is still working fine though, especially with the unit being undervolted (new TRM release improved efficiency slightly).

 

[xenium]

Just came across this: https://www.ospreyelectronics.io/product-page/c1100-heat-sink
GPU style heatsink for the c1100. Kinda pricey, and I wish it didn't extend above the board so much, but seems like it might be effective. My 3D printed fan mount is still working fine though, especially with the unit being undervolted (new TRM release improved efficiency slightly).

They work really well, but with a caveat. Shaved ~7w per card vs the san ace blowers, and they run basically silent now (mining eth). The downside is the Osprey heatsinks have no direct contact with any of the VRM components (unlike the stock heatsink). On a low power bitsream like Eth Im not too worried, especially since they're indoors and rarely see ambient above 80F, but I likely wouldn't have bought those pricey heatsinks if I knew they weren't actually engineered to contact all the VRM components. The heatsink only contacts the FPGA/HBM, and relies on the through-heatsink airflow to cool the rest of the SMD/VRM components.

They went from design to production/shipping in like 4 months so Im not too surprised that the design is just lazy. On mine they also had to hand grind out the USB ports on their I/O brackets, and one of them still didn't fit the USB cable, so I had to just re-use the stock I/O bracket. Not a big deal, but again, shows how rushed the product is. They shipped with some god awful bubblegum thermal pad on them as well.

Mind the fan cable mess, they have individual cables for each fan and are PWM controlled via the mobo/bios (the top card runs a bit hotter/higher fan speed).

 

[Andrew_Carr]

Update to this. Kaspa bitstream is now up on teamredminer as of last week or so. I'm not able to run the same clock speeds as ETH with it. Stock settings gets almost 2.4GH/s, but still fiddling with and trying to get better settings to work.

These settings worked for me:
--fpga_update_fw
--fpga_vcc_int=760
--fpga_vcc_bram=875
--fpga_vcc_mem=1180
--fpga_clk_mem=1215


Mostly similar to the ETH settings, but no matter what I set for the core clock it seems it crashes. Since this is a new bitsream these probably aren't the best settings, but they do seem to have dropped temps a bit.

 

[xenium]

Update to this. Kaspa bitstream is now up on teamredminer as of last week or so. I'm not able to run the same clock speeds as ETH with it. Stock settings gets almost 2.4GH/s, but still fiddling with and trying to get better settings to work.

These settings worked for me:
--fpga_update_fw
--fpga_vcc_int=760
--fpga_vcc_bram=875
--fpga_vcc_mem=1180
--fpga_clk_mem=1215


Mostly similar to the ETH settings, but no matter what I set for the core clock it seems it crashes. Since this is a new bitsream these probably aren't the best settings, but they do seem to have dropped temps a bit.

you can leave out mem settings, HeavyHash doesn't touch the hbm at all. Great to see the C1100 get some love from TRM though. More profitable than ETC right now. Fwiw 760 vcc is high, keep an eye that you aren't hitting the 170A limit for the regulators. Personally I run mine at 550-600 for efficiency. They get power hungry over 600.

Also adding --debug will show you vccint regulator temps. Something to keep an eye on, other than just core temps.

Example from one of the TRM devs, tuned for efficiency:

 

[Andrew_Carr]

Ah, that's what I was doing wrong. I figured the ETH settings would be safe enough, but didn't think I was that much higher than the best settings. Lowering core clock to 550Mhz fixed the hashrate and now it's at 4.4Ghash. Thanks, just doubled my profitability, lol. I'll have to play around and see how much lower I can drop the voltages. 720mv vcc bram and 600vccint seems stable though.

 

[cirial]

Where did you pick these up?

 

[Sycamore]

Hi guys, have you done any crypto mining with the

Xilinx Artix-7 FPGA M.2 Development Board (A100T FPGA/512MB DDR)​

 

[Andrew_Carr]

Where did you pick these up?

I got mine on eBay. There’s also an FPGA discord where people sell stuff but there’s scams and such you have to watch out for.