Topic: BFL Single and BFL mini-rig seems to have inferior performance (Read 6682 times)

bitfury might have a much better handle on how to simulate and get the correct numbers since he's taken his designs and actually put them onto silicon. While there could be something specific about the EP3SL150F780C4 that causes it to be so far out from simulations, it's probably pretty likely that if bitfury's simulations match his power draw for his own chips, they won't be that far off for modeling the same high level design on a different chip.

Still, why are we necroposting this?

You do realise those photos are from last June and those are the FPGA rigs, right?

good news for you !!!





you can see here how many users they are on eclipse
https://eclipsemc.com/

Active Miners    1497
Current Speed:    1.73 TH/s

of course they are asic miner who mining !!!!

here the production line
http://www.butterflylabs.com/mini-rig-production-line/

^^^ This.

I suspect you're bumping into the same reason their initial specs didn't match the delivered product: the power (and perhaps clock?) estimates out of these tools are wildly wrong because they don't account for the very high toggle rates.

Your numbers would have been more in line with what BFL originally claimed for the product but couldn't deliver.

1GH/s @ 20W.

What was the original hashrate BFL claimed when they were making a prototype of bitforc esingle? wasn't it like 1.5ghash or something at 40 watts?

This is what I am doing. Just not boards, but producing 1-U servers with 48 spartan on boards. Soon first limited try-out will be sold, and I will inform everybody here as well.

About taking any board - if it would be piece of cake - like spending one day and done - I would do this. But actually only compilation takes 24 hours... Then it would take even more for adoption and debugging, multiply it per number of boards. Not knowing exactly volumes - means you would spent effort and it won't produce any fruit.

Also with pricing... such prices as $250 per spartan seems ridiculous. Like some vendors offer. I understand that as "dev board" for geeks to play it is nice offer, but as for mining - that is bad offer. Maybe this played a bit as well, as I disclosed how much it costs to build boards, etc. 1U will be more expensive though than initial setup of bitfury with rack ... more density ==> higher costs (say without active cooling it is unlikely that you would put more than 24 spartans in 1U).

Buy the board, program it, demonstrate your work, then people will believe and have buying interest, right now it's just promises without any backing

Licensing terms were written on this forum - basically $25 per spartan for low volumes and $20 per spartan for higher volumes. + Additional offer to scalp even more performance for $5 per spartan, if spartan volumes will be higher than 5000 total subscribing. Enterpoint shown no interest in this, as well as other board vendors except starting project - shalab.si

Maybe this happened because I wrote that these works will be open-source when spartans will be phased-out from generic mining rush. It seems to starting happening by the way. Depends on what will happen with ASICs.

Depends. Potentially yes. 12 W power per Spartan Chip. That power makes it difficult to use on many boards, for example with ZTEX boards.

You may look on our web site how many there capacitors below the chip (www.bitfury.org). On Icarus it maybe just 260-280 Mh/s, because it maybe wise to _lower_ voltage there to 1.15 V and decrease frequency accordingly. So it would run about 210 Mhz - 220 Mhz and consume about 9 - 9.5 W taking about 8 Amps power.

But exact combination of voltage and frequency should be determined based on power supply components and bypass capacitor calculations.

so your bitstream it´s  300Mh/s? per Spartan Chip?  it would be 600Mh/s at Icarus?

About board voltage - yes - 1.3 V, and 1.26 V on chip. We have board in production with separate power supplies and LVDS high precision clock supplied based on SI5338A external PLL. It may give superior performance to Spartan internal clock generators.

2 mrb: thanks for your input. 66 W per single means about 33 W per chip and including COP of system it is probably 26-28 W per chip (I expect 12 V -> 1.1 system should have COP about 80-85%).

If I recall, don't you run your Spartan6s at higher voltages to achieve the MH/s that you are with your design? I imagine you'd hit the same problem that the tricone design is hitting with sagging voltages on boards that don't feed a slightly higher core voltage.

Jothan's numbers are incorrect (perhaps he is trying to subtract a baseline idle power consumption that is incorrectly measured, or his watt meter is defective, or his singles are throttling)... I measured power consumption with a clamp meter around the 12V input, and I get 66W per Single (rev 3 without the big 80mm fan an the bottom, the ones with the fan consume 68W), with an average 81W at the wall measured with a kill-a-watt, meaning the power adapters are 66/81 = 81% efficient.

And BFL seems to be using core voltage of 1.1V per the test points in the top right corner of their PCB: https://i.imgur.com/vrzol.jpg
I measured 1.13V with a multimeter on this 1.1V test point.

thank you for the insight.

With tricone model unrolled round design is necessary, because otherwise it is near impossible to protect design. for me this means start another design from scratch for unrolled round. I already had ideas how to fit actually 2 full unrolled rounds there at clock 150-160 Mhz approx. (300 - 320 Mh/s). and little amount of small hashers (4-5) each giving 4 Mh/s approx. So total about 316 Mh/s to 340 Mh/s which is more than sea of hashers but also power envelope not 12 W but only 8W. This design would beat also Stratix approach with BFL mini-rig, as basically spartans would consume for 25 Gh/s about 820 W from plug compared to 1.25 kW of mini-rig. And will work on all boards as well. But this is due found solution with round expander which needs to wide buses with dual-clock design.

Plus another big part of work is testing for compatibility with every board around - I need these boards.

About question for hardware - about 655 spartans installed and mining using current bitstream (one full and one 5/6 filled rack).
So yes - verified in hardware of course, not in software. About BFL estimation - only simulations though, so it may fail due to power problems... but actually not likely, as these conditions are not putting chips to their edges at all, compared with 12 W on Spartan device.

Do you own the hardware? If you don't have at least 2 of each for testing go shit in your hat.
Whats your products performance? NO SIMULATIONS either.

If you can do this, why not use the tricone model? Publish a bitstream for cairnsmore that extracts a commission and go straight to the end user. Every single person would use it. You could direct 200mh/s? to yourself and make your goal of $25/chip in no time. There are hundreds of these ready to go.

Depends on architecture. For example for spartan sea-of-hashers - it is not. Mh/s is based on how many small cores inside placed. Fractional could be there.

Basically for example in Spartan - 61 clock cycles used for calculation 4 cycles used for loading of new job. 65 clock total. 82 cores. You get * 82 / 65 Mh/s hash rate.

For Altera solution I've put into chip that is 2 unrolled sha256, so output 2 * Mhz.