Consensus on most efficient GPU in terms of hashes per kw?

[jamesg]

Hi yochdog,

I see you have two concerns, space and power.

On the power side, it might be beneficial if you run your rigs off of a 240v circuit (think dryer or AC) rather than 120v. This would basically halve your amperage requirements giving you more room per circuit for rigs.

Example: If you are running a 1000w rig....

1000w / 120v = 8.334 amps

1000w / 240v = 4.167 amps

Also, PSUs run more efficient off of higher voltages so this might help with your efficiency also.

This doesn't alleviate any of your space concerns, but would give you more headroom on the power side.

[RandyFolds]

Mousepotato is on the deals for the 5870s: https://bitcointalk.org/index.php?topic=55188.0

[BkkCoins]

I really think anyone looking at new investment now should be looking at FPGA solutions because, even though capital costs are higher, they offer a far more stable ability to profit under various future scenarios for a much longer time. The higher capital costs should simply be factored in as a monthly cost like a loan would be.

Here's some example math. Naturally everyone has slightly different conditions.

Right now I have $610 invested in 3x5830 + system. It produces 930 MH at 530W. Let's say I treat that investment as if borrowed money at 7% and the salvage value after 2 years is $200 (you would be taking a bigger risk to keep the HW than selling). According to my loan calc my HW cost is $18/mo. My electric cost at 0.10/kwH is $38/mo. And according to my btc calc I can mine 24 BTC/mo. at todays difficulty.

At current $4/btc that's 24*4 - 18 - 38 = $40/mo. return on $610 invested.
But when $2/btc that's 24*2 - 18 - 38 = $8/mo loss. So I've turned off my miner.

My cost per BTC is $2.33 each and I need to sell higher than that.

Likewise next year when block prize drops to 25 btc it's the same as difficulty doubling and making half as many btc (though granted I may be helped by others dropping out). It's not very resilient to market vagaries and you only really profit in a bull market.

Calculate this again with current FPGA costs. $365 ZTex board using 9W to give 200 MH/s (x 5 for comparison). 365x5 + 50 PSU +5 USB hub = $1880. Salvage about $500 after 5 years. My loan calc says $26/mo HW cost. My electric drops to $3/mo. and I can mine 26 BTC/mo.

At current $4/btc that's 26*4 - 26 - 3 = $75/mo. return on $1880 invested.
But when $2/btc that's 26*2 - 26 - 3 = $23/mo. return. So I'm still mining.

My cost per BTC is $1.11 and I have to sell higher than that.

Also on a 240/20 Amp circuit I can mine 8,370 MH/s with 27 GPUs.
On the same circuit with FPGAs I can mine 106,600 MH/s with 533 FPGAs (ie. unlimited).

So if you're limited by power feed you can reach so much higher revenues with FPGAs. The real kicker is after you amortize your capital costs I'm willing to bet the FPGA has a longer life than the GPUs (no fan!). So after 2 years you are shutting off GPU miners or replacing HW but with the FPGA your running cost is now super low.

After 5 years, your BTC are costing you 12 cents each and that's the position you want to be in a few years from now while everyone else is dropping out you are mining BTC practically free.

My main point is that lowering your operating costs puts you in a much more resilient long term position and if you are in it for the long term then the capital is better invested for that. Plus, in regions with higher kWH prices you would still be mining under almost any conditions = safer investment.

(Not to mention I could actually live in a room with 50GH/s of FPGAs working!)

[nmat]

....

I agree with what you wrote, but note that the cost to generate will not always be 12 cents. Difficulty will likely increase a lot in the next years so an FPGA bought now might not be profitable in 5 years.

[BkkCoins]

....

I agree with what you wrote, but note that the cost to generate will not always be 12 cents. Difficulty will likely increase a lot in the next years so an FPGA bought now might not be profitable in 5 years.

Very true. Wherever it goes you will be better off with lower operating costs.
It's hard to say but I suspect that the resale market for FPGAs may be more long term than the video cards as the gaming market is fickle and out of date cards don't fair well. I think the crypto market for FPGAs may be more stable but who knows... in the future I'd expect low energy costs to be a real advantage.

[DeathAndTaxes]

Good analysis but I think a 5 year effective lifespan is unrealistic.  Maybe 3 years.  Moore's law applies to FPGA.  28nm FPGA will be hitting eventually and will (once discounted) provide roughly double the performance for half the cost and half the wattage.   The drop to 20nm will happen "eventually" and likely within next 5 years and will once again roughly double performance per watt and per dollar.

Assumming units retain 33% resale value after economically obsolete is also the stuff of unicorns and rainbows.  They will be utterly worthless after 2 process drops have occured.  The FPGA will be roughly as powerful as a $50 20nm part and consumes 4x the power. 

Still your idea of ammortizing hardware to get monthly cost is a good idea I just think your assumptions are somewhat pie in the sky.

[BkkCoins]

Good analysis but I think a 5 year effective lifespan is unrealistic.  Maybe 3 years.  Moore's law applies to FPGA.  28nm FPGA will be hitting eventually and will (once discounted) provide roughly double the performance for half the cost and half the wattage.   The drop to 20nm will happen "eventually" and likely within next 5 years and will once again roughly double performance per watt and per dollar.

Assumming units retain 33% resale value after economically obsolete is also the stuff of unicorns and rainbows.  They will be utterly worthless after 2 process drops have occured.  The FPGA will be roughly as powerful as a $50 20nm part and consumes 4x the power.  

Still your idea of ammortizing hardware to get monthly cost is a good idea I just think your assumptions are somewhat pie in the sky.

I thought about 3 years but reasoned that they should still mine just as well even if newer models do hit the market so the mining life is probably still feasible. In that case the problem is what are other miners using then and can you compete with old HW any more.

My reasoning on the resale was that older FPGA kit floating around eBay still seems to get some interest even if just because hobbyists love playing with these things to see what they learn. But ya, if too many are sold and then dumped later the market will be trash.

Looking back at previous Spartan gens the chip density always goes way up but the chip pricing is held quite high since unlike CPUs they don't have much competition. Even now you really can't find another part that competes for mining with the Spartan 6 - Altera has one that kind of comes close but not really.

In the past when Xilinx has come out with new series the old ones drop in price so I can see valid reasoning that you would want to wait for Spartan 6 to reach that drop and then buy in. I haven't checked where Spartan 6 is in it's life cycle now. The FPGA mining cards are mostly FPGA cost and as times goes by that margin for PCB and assembly should drop as dev costs get let go.

[yochdog]

Thanks everyone for the responses.....always amazing how much knowledge is out there.  Even cooler that people take time to share it with complete strangers.  

I have decided to mazimize the use of 5970's for my farm.  If those will not fit or are unsuitable for the location I will be using 6970's or 5850's.  Definitly phasing out all 5830's, 5770's, and 6870's.

One of the main reasons I am avoiding FPGA's at the moment is the high upfront capital cost.  To replicate my 14,000 MH/s, I imagine it would cost somewhere on the order of $20,000+.  Given that my electricity is insanely cheap, I don't really get an edge from a operating cost standpoint.  

It will be an interesting few months as I try to push right up against the capacity constraints and see how much hashing density the infrastructure will support.  

[Mousepotato]

To replicate my 14,000 GH/s, I imagine it would cost somewhere on the order of $20,000+

 

[yochdog]

To replicate my 14,000 GH/s, I imagine it would cost somewhere on the order of $20,000+

 

LOL!  Nice catch.....corrected! 

[Syke]

I have decided to mazimize the use of 5970's for my farm...  To replicate my 14,000 MH/s

What's your power setup look like? I'd guess 6x 20amp circuits. That should hold at least 36 5970s for a total of 25 GH/s if you keep the circuits below spec at 80% (16amps continous). You could probaly fit a few more in if you overclock/underclock to optimize your power draw per circuit.

[jjiimm_64]

Im with everyone else about the 5970's, I love them...  2900 Mhash pulls 1190watts at the wall. 

[legolouman]

Before I accidentally reiterate anything already said (as I didn't read many posts), I'd like to apologize for reiterating what's already been said.

By simply looking at the mining hardware comparison, it looks like the 5970 and 6990 have the highest hashing power per card. The 6990 is slightly higher(average) but the 5970 is probably easier to find in large numbers, thinking in terms of used hardware.

Of course, Nvidia's Tesla S2070 seems to hash at like 750mh/s (despite costing like $2.5k) You could go that way, if you can conjour up cards from thin air.

With the 7000 series looming over us, I'd be careful making any large decisions, as it could change everything. 5xxx/6xxx prices could plummet or skyrocket, and the 7xxx series could be entirely power horse and efficient or not. We only know so much, the rest is speculation. Oh and from what I've seen, the 7870 (I think that was the model) was like $600 for a pre-order.


Good luck in making some sense out of my late night forum posts.

[Transisto]

Good luck in making some sense out of my late night forum posts.

Please go to bed.  You're not helping.

[phelix]

underclock to become more energy efficient.... I think 3MH/W for a whole rig might be possible. use a low power cpu and as many underclocked (double) GPUs as possible

my cards run at 3MH/W, whole rig at 2.7MH/W  https://bitcointalk.org/index.php?topic=40056


it would be very interesting to see how far you can underclock a gpu... below bios limit I mean

Just for fun I tried and got 5970 down to 120 Mhz then again the 5970 is rather adaptable I run the memory down to 160Mhz for optimal efficiency.  I didn't have kill-a-watt connected but I will check again and give you an idea of peak efficiency.

also lowering voltage decreases power use much more

[DeathAndTaxes]

underclock to become more energy efficient.... I think 3MH/W for a whole rig might be possible. use a low power cpu and as many underclocked (double) GPUs as possible

my cards run at 3MH/W, whole rig at 2.7MH/W  https://bitcointalk.org/index.php?topic=40056


it would be very interesting to see how far you can underclock a gpu... below bios limit I mean

Just for fun I tried and got 5970 down to 120 Mhz then again the 5970 is rather adaptable I run the memory down to 160Mhz for optimal efficiency.  I didn't have kill-a-watt connected but I will check again and give you an idea of peak efficiency.

also lowering voltage decreases power use much more

Yup generally speaking change in voltage results in change in power squared.  So if you can undervolt 10% you can cut power by 19%.  If you can undervolt 15% you can cut power ~27%.