ETH FPGA's coming soon. Want to get the inside scoop? Read on in this thread!

1. Three contestants! Xilinx, SQRL, & Osprey/Dracaena. Who will be first to market? No one knows! SQRL has announced their bitstream is in BETA.

2. A far cry from the originally advertised, championed and theorized 300 MH/s, SQRL will be launching at 40 MH/s with a target of 70W. Possibly up to 75 MH/s can be achieved, at a later date.

The target hardware is a Forest Kitten 33 - Xilinx's VU33P - with a $ of 1950 USD.

3. Next up is Osprey Mining, now having split off their bitstream development team until the brand "Dracaena Mining". Currently, they sit at 20 MH/s for 50W, with a potential to get up to 30 MH/s @ 50W. Osprey's target price? $1350-$1500.

4. Last, but not least, Xilinx. Not much is public about Xilinx's plans, except for what's been leaked by some folks on an FPGA discord. No target hashrate and performance, but given it's modelled on the same hardware, we can expect in a similar range, with better price point.

5. All in all, you're ending up at about 1.75W per MH/s, with a potential floor of 1.25W per MH/s. For contrast, a GPU can pull 1.46W per MH/s to 2.5W per MH/s. More energy efficient? Sure!

6. Unfortunately where FPGAs fall flat is in the $/W/MHs. At 37.5 USD per MH/s (1500 / 40) a Vega64 still beats it flat on it's back at 6.25 USD per MH/s (350 / 56).

7. Many influencers claim they demolish GPUs, and that's only true for Blackminer-style devices - where tons of cores can be packed into one tight little box.

8. Unfortunately, all of the above FPGAs are standard style cards - much like GPUs. Even worse, Osprey's device requires a custom riser or backplane. Unnecessary infrastructure investment for miners.

With ASICs from Linzhi and new Innosilicon miners revisions looming...

9. ...It's feeling like FPGAs will be an enthusiast class piece of equipment, at best.

Is there room for improvement? Tons. We'll have to wait and see what Xilinx has been cooking up, and perhaps the next generation FPGAs will bring better value.

10. CC: @cryptochrisw, @WhaleBearMan, @H4SHR8, @RedPandaMining, @BitsBeTrippin and @greerso (since ya'll seem to be the only ones interested in ETH mining on my feed).

11. But wait! What about the alternate use cases I keep hearing about? Surely my FPGA will be good for rendering, AI, and all those other fancy use cases?!

12. I've spoken about this at length before, but NO SERIOUS CONSUMER OF COMPUTE WANTS YOUR CRAPPY MINING HARDWARE. NONE.

Look, you might get a small scrappy provider, and you might get some tests, but chances are your mining farm doesn't have the right infrastructure.

13. In the 'compute' and 'render' business, it's not just the infrastructure setup or the hardware - it's the trustworthiness of the company you're renting the hardware from. The CFO of DreamWorks isn't going to sign off on rendering his multi-million dollar movie on your farm.

14. But then, what are FPGAs good for outside of cryptocurrency? Well, traditionally they sit mostly in R&D labs to assist simulations, model development for AI (specifically NLP is a big use case right now, some NN's, etc) and HFT for the BIG funds.

15. But they're not in the density that mining farms have - a set of 16 FPGAs might be enough to cover the needs of an R&D lab and their researchers. Aside from aiding product development, we see a lot of them embedded in other datacenter products for network acceleration...

16. ...and storage accleration. Essentially helping to speed up and improve the performance of other hardware (this, honestly, is where they shine).

I could rant about this specific piece a lot (maybe save it for a podcast, eh?), but TL;DR: no one is going to rent your FPGA.

17. So where do I see FPGAs being a good investment for miners?

2 years down the road, I think we'll see ZKP and VDF's and most of Eth's L2 infrastructure accelerated by FPGAs. Specifially, 381-bit primes and 126-bit primes will shine on newer-generation FPGAs. CC @StarkWareLtd

18. FPGAs will be a critical piece for blockchain infrastructure where PoW isn't utilized, and where low latency needs with tight compact formfactors are desired. Why run thousands of GPU rigs when you can condense it to 4 FPGAs?

19. See https://github.com/bsdevlin/fpga_snark_prover for an example of accelerating SNARKs through FPGAs.

20. While all three FPGA models are 'advertised' to be capable of dual mining, the pickings for the cards are slim and no ETH bitstream has yet been realeasd with the capability, with only Tellor making sense.

21. And in an amusing but very predictable twist, FPGA miners are not happy I’ve shared this information.

The Discord flaming continues.