3

I'm looking for a good ethereum mining GPU. I heard it's mostly memory bandwidth intensive, so I rounded up the specs of the latest cards from Wikipedia and came up this (prices are in CAD.):

card                     price     watts      GB/s    GB/s/$    GB/s/W  $/(GB/s)/year
GeForce GTX 1050        148.82        75       112      0.75      1.49          1.548
GeForce GTX 1050 Ti     189.78        75       112      0.59      1.49          1.974
GeForce GTX 1060 3GB    271.69       120       224      0.82      1.87          1.413
GeForce GTX 1060 6GB    339.96       120       224      0.66      1.87          1.768
GeForce GTX 1070        517.44       150       256      0.49      1.71          2.355
GeForce GTX 1080        817.81       180       352      0.43      1.96          2.707
GeForce GTX 1080 Ti     954.34       250       484      0.51      1.94          2.297
NVIDIA TITAN X         1638.35       250       480      0.29      1.92          3.977
NVIDIA TITAN Xp        1638.35       250     547.7      0.33      2.19          3.485
Radeon RX 460           189.78        75       112      0.59      1.49          1.974
Radeon RX 470           244.39       120       211      0.86      1.76          1.349
Radeon RX 480 4GB       271.69       150       224      0.82      1.49          1.413
Radeon RX 480 8GB       326.30       150       256      0.78      1.71          1.485
Radeon RX 550           107.86        50       112      1.04      2.24          1.122
Radeon RX 560           135.16        80       112      0.83       1.4          1.406
Radeon RX 570           230.73       150       224      0.97      1.49            1.2
Radeon RX 580 4GB       271.69       185       256      0.94      1.38          1.236
Radeon RX 580 8GB       312.65       185       256      0.82      1.38          1.423

Everything there's pretty self-explanatory except for the far right column. It's an estimated TCO per unit memory speed per year including cost of power in Ontario. So, you can probably just ignore it if you don't live in Ontario. In fact, you can probably ignore it even if you do.

Anyway, the GB/s/W is the dominating factor when it comes to electrical bills. By that measure, the Radeon RX 550 is winning, but you'd need quite a few of those in a single system to really reap the benefits.

In the rest of my build, I'm assuming that mining throughput scales more or less linearly with additional GPUs. If this is true, communication between GPUs isn't necessary (i.e. crossfire support unneeded) and it doesn't matter too much what the PCIE bus speed is. I'm assuming linear scaling because I think this is how it works on the network at large, with eth generation scaling linearly with number of miners. Is that a sound assumption?

Anyway, to those who have actually done this in real life, am I at all close here? Am I missing anything critical?

BTW, this is the script I used to generate that table:

#!/usr/bin/env python

# constants for doing conversions.
hour = 60*60
year = 365.25*24*hour
kilo = 1000
watt = 1
CAD = 1
USD = 1.36529*CAD

# Data is from Wikipedia.
# https://en.wikipedia.org/wiki/GeForce_10_series
# https://en.wikipedia.org/wiki/AMD_Radeon_400_series
# https://en.wikipedia.org/wiki/AMD_Radeon_500_series
card = [
        'GeForce GTX 1050', 'GeForce GTX 1050 Ti', 'GeForce GTX 1060 3GB',
        'GeForce GTX 1060 6GB', 'GeForce GTX 1070', 'GeForce GTX 1080',
        'GeForce GTX 1080 Ti', 'NVIDIA TITAN X', 'NVIDIA TITAN Xp',

        'Radeon RX 460', 'Radeon RX 470', 'Radeon RX 480 4GB',
        'Radeon RX 480 8GB', 'Radeon RX 550', 'Radeon RX 560',
        'Radeon RX 570', 'Radeon RX 580 4GB', 'Radeon RX 580 8GB',
        ]
price = [
        109*USD, 139*USD, 199*USD, 249*USD, 379*USD,
        599*USD, 699*USD, 1200*USD, 1200*USD,

        139*USD, 179*USD, 199*USD, 239*USD, 79*USD,
        99*USD, 169*USD, 199*USD, 229*USD,
        ]
watts = [
        75, 75, 120, 120, 150, 180, 250, 250, 250,

        75, 120, 150, 150, 50, 80, 150, 185, 185,
        ]
gflops = [
        1862, 2138, 3935, 4372, 6463, 8873, 11340, 10974, 12150,

        2150, 4940, 5834, 5834, 1211, 2611, 5095, 6175, 6175,
        ]

memspeed = [
        112, 112, 224, 224, 256, 352, 484, 480, 547.7,

        112, 211, 224, 256, 112, 112, 224, 256, 256,
        ]


# http://www.ontario-hydro.com/current-rates
powerprice = (12*0.087 + 6*0.18 + 6*0.132)/24 * CAD/(kilo*watt*hour)

# Expected lifetime of GPU before resale. PCIE4 is coming out in two years, 
# probably destroying all of this.
lifetime = 2*year

# Expected percent depreciation of GPU's resale value after above lifetime has 
# elapsed. Optimistic?
depreciation = .2

print('{:20}{:>10}{:>10}{:>10}{:>10}{:>10}{:>15}'.format(
    'card', 'price', 'watts', 'Gflops',
    'Gflops/$', 'Gflops/W', '$/Tflops/year'))
for i in range(len(card)):
    gpd = gflops[i]/price[i]
    gpw = gflops[i]/watts[i]
    # Average total cost of ownership per year per teraflops of computing 
    # power, including cost of power if for some reason running at full tilt 
    # all year round.
    dpT = 1000/gpd
    tco = dpT*(depreciation/lifetime + powerprice) * year
    print('{:20}{:10.2f}{:10}{:10}{:10.2f}{:10.3}{:15.4}'.format(
        card[i], price[i], watts[i], gflops[i], gpd, gpw, tco))

print()

print('{:20}{:>10}{:>10}{:>10}{:>10}{:>10}{:>15}'.format(
    'card', 'price', 'watts', 'GB/s',
    'GB/s/$', 'GB/s/W', '$/(GB/s)/year'))
for i in range(len(card)):
    gpd = memspeed[i]/price[i]
    gpw = memspeed[i]/watts[i]
    dpg = 1/gpd
    tco = dpg*(depreciation/lifetime + powerprice) * year
    print('{:20}{:10.2f}{:10}{:10}{:10.2f}{:10.3}{:15.4}'.format(
        card[i], price[i], watts[i], memspeed[i], gpd, gpw, tco))
3

Ethereum, like others have said is memory intensive. Therefore AMD cards such as the 470/480 570/580 shine. What you need to look for is the hash rate for each graphic card to determine what is the most cost effective. But, you also have to consider if stock is even available for those cards.

For Ethereum, I would only consider 1060 6GB, 470/570 480/580. Ethereum requires cards that are over 2gb at the moment and it will soon grow to be over 3GB as its DAG size grows overtime. Therefore cards like the RX 550, despite being a watt/cost winner, will not work. Although 1060 3GB will still work, it's not a sound investment as dag is approaching 3GB.

If I were to rank the cards, it would be.

  1. 470/570
  2. 480/580
  3. 1060 6gb

The 470/570 wins as Ethereum does not depend much on the number of cores or GB/s texture fill, but rather memory speeds. They can reach 28~32 Mh/s, equivalent to the 480/580.

The 480/580 is more expensive, but still beats nvidia in terms of performance for a lower MSRP cost. Stock cards run at 22~24 Mhash/s but with timing tweaks, memory overclock it can achieve 28~32 Mh/s. The card draws 100W when underclock to the core and power limit is lowered from personal experience.

The 1060 6GB can mine at ~18 Mh/s, but with memory overclock it can achieve 22~23 Mh/s at 90W.

| improve this answer | |
0

Instead of relying on theoretical considerations, it is wise to take into account real measured hash rates. Fortunately, since many tech sites have taken upon themselves to do these benchmarks, and publish the results, you do not have to actually purchase the cards to get hard numbers for yourself. These numbers can reveal certain unexpected deviations from the theoretically expected hash rates: For example many miners have found an unexplained decrease in hash rate on the newer Polaris GPUs at larger DAG sizes. (1)(2) This can mean trouble over the long term, as those cards become obsolete for mining before paying for themselves.

Looking at the benchmark results, we can immediately see some interesting figures. The older R9 390/Xs perform superbly, and do not appear to degrade at higher DAG sizes. It's somewhat surprising to see the GTX 1070 above the 1080, something we would never know just looking at the 1080's higher core count and memory bandwidth. It appears that GDDR5 has superior latency characteristics for mining than the GDDR5X used in the 1080 (Another reason spec sheets do not tell the whole story: latency information is missing). The GTX 1060 also performs adequately, though for future proofing, the 6 GB version is a must.

Overall, due to the higher power consumption and lower availability of the older Radeon cards, although the R9 390 is a viable candidate, I do not heavily recommend it. Essentially, the decision is between the 1070 and the 1060. The choice between the two essentially boils down to how much capital you plan on investing into your mining gear. Since for a more expensive GPU, the rest of the system makes up for a smaller proportion of total costs, I would recommend the GTX 1070 as a first choice and the GTX 1060 or R9 390/X GPUs as secondary options.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.