Ethereum GPU mining is back and — currently, at least — highly profitable. But there's more to it than just firing up the software and letting it run in the background, especially if you've managed to procure one of the best graphics cards. Most of the graphics cards in our GPU benchmarks hierarchy can theoretically earn money right now, depending on how much you pay for power. However, you'll want to tune your graphics card with the optimal settings, and the brand and card model can have a big impact on overall performance and efficiency.
First, let's note that we're not trying to actively encourage anyone to start a mining farm with GPUs. In fact, based on past personal experience that some of us have running consumer graphics cards 24/7, it is absolutely possible to burn out the fans, VRMs, or other elements on your card. At the same time, we know there's a lot of interest in the topic, and we wanted to shed some light on the actual power consumption — measured using our Powenetics equipment — that the various GPUs use, as well as the real-world hashing rates we achieved. If you've pulled up data using a mining profitability calculator, our figures indicate there's a lot of variation between power and hash rates. Don't be surprised if your particular card doesn't reach the level of performance others are showing.
There's also the old saying: If you can't beat 'em, join 'em. Part of what drives up the price on cryptocurrency is if there's demand for it that can't be met. Right now, you could potentially make a few hundred dollars a month, per GPU. Sure, the price could go up and you might miss out. It could also drop. And if more people mine, the rewards for mining become less, which means maybe the large mining farms will stop snapping up all the graphics cards (and gaming laptops!) in pursuit of Ethereum. Do as you will, but don't think that us discussing this will really make the situation any worse. The only way this gets better is if profitability of mining plummets, and the more people that mine in their spare GPU cycles, the less there is for the mining farms.
We'll start with the latest generation of AMD and Nvidia GPUs, but we also have results for most previous generation GPUs below. There are a few things you should know before getting started. First, Ethereum GPU mining requires more than 4GB of VRAM, so if you're still hanging on to an RX 570 4GB, it won't work. Second, there are a lot of different software packages for mining, but we're taking the easy route and using NiceHash Miner 3.0.5.6. It includes support for the most popular mining solutions, and it will even benchmark your card to determine which one works best. Except, it's not always accurate, so we've standardized on NBMiner for the Nvidia GPUs and PhoenixMiner on the AMD GPUs. (Our test results are generally within a few percent of each other on either miner, but Phoenix did a bit better on AMD. YMMV.)
Nvidia Ampere and AMD Big Navi Mining Performance
We've run these initial benchmarks using NiceHash Miner, including its built-in benchmark that's required to get started. We tested each graphics card in stock mode, and then we also attempted to tune performance to improve overall efficiency — and ideally keep temperatures and fan speeds at reasonable levels. We let the mining run for several minutes before checking performance, power, etc., as often things will slow down once the graphics card starts to heat up.
It's also important to note that we're reporting raw graphics card power for the entire card, but we don't account for the power consumption of the rest of the PC or power supply inefficiencies. Using an 80 Plus Platinum PSU, we should be running at around 92% efficiency, and wall outlet power consumption is typically about 50-80W higher than what we show in the charts. About 40W of power goes to the CPU, motherboard, and other components, while the remainder depends on how much power the GPU uses, including PSU inefficiencies.
There's a lot to discuss with these charts, specifically, what do we mean by "tuned" performance. The answer: it varies, often massively, by GPU.
Let's talk about the big picture quickly before we get into the details. The fastest GPUs for Ethereum mining right now are the RTX 3080 and RTX 3090, by quite a large margin. Our baseline RTX 3080 measurement got 85MH/s, and the baseline 3090 got 105MH/s. Additional tuning improved the 3080 performance to 93MH/s, while the 3090 FE limited us (on memory temps) to around 106MH/s.
Meanwhile, the RTX 3060 Ti and 3070 cards all started at close to 52MH/s — even though the 3070 is theoretically faster. That's because Ethereum hashing depends quite heavily on memory bandwidth and perhaps somewhat on capacity. Overclocking the VRAM on those GPUs got performance up to around 60MH/s. AMD's RX 6000 cards started at close to 60MH/s, and with tuning, we achieved 65MH/s — there wasn't much difference between the three AMD GPUs, mostly because they're all using the same 16GB of 16Gbps GDDR6 memory.
You can see the power chart (which is reversed because it would take extra effort to get my charting tool to do low to high), but the overall efficiency chart is more important than raw power use. Here, the lower power RTX 3060 Ti and 3060 shoot to the top, followed by the custom Colorful RTX 3080 Vulcan. The RTX 3080 FE and RTX 3090 FE are roughly near the same level as AMD's RX 6800, 6800 XT, and 6900 XT after tuning.
Finally, we have temperatures. These are GPU core temperatures, but they're actually not the critical factor on many of the cards. AMD's cards ran hot at stock settings, but all of the cards benefit greatly from tuning. More importantly, while we couldn't get GDDR6 temperatures on the 3060 Ti and 3070, we did get VRAM temps on the 3080 and 3090 as well as the AMD cards. At stock, the 3080 and 3090 Founders Edition cards both hit 108-110C on the GDDR6X, at which point the GPU fans would kick up to 100% (or nearly so). The cards settled in at 106 degrees Celsius, with GPU clocks fluctuating a bit. AMD's RX 6000 cards also peaked at around 96C on their GDDR6 at stock, but tuning dropped VRAM temps all the way down to around 68-70C. This brings us to the main area of interest.
How to Tune Your Graphics Card's Ethereum Mining Performance
Let's start by noting that every card model is different — and even cards of the same model may vary in performance characteristics. For the 3080 and 3090 cards with GDDR6X memory, keeping that cool is critical. We've seen examples of cards (specifically, the EVGA RTX 3090 FTW3) that can run at up to 125MH/s, while the memory sits at around 85C. That's because EVGA appears to have put a lot of effort into cooling the memory. Without altering the cards, the Nvidia 3080/3090 Founders Editions let the memory get very hot while mining, which can dramatically hinder performance and/or reduce the card's lifespan. Let's take each card in turn.
GeForce RTX 3090 Founders Edition: While technically the fastest card for mining that we tested, we really don't like the idea of running this one 24/7 without hardware modifications or serious tuning. At stock, the fans end up running at 100% to try to keep the GDDR6X below 110C, and that's not good.
For our purposes, we tuned the card by dropping the GPU core to the maximum allowed -502MHz, set the memory clock to +250MHz, and put the power limit at 77%. That gave us a GDDR6X temperature of 104C, which is still higher than we'd like, and performance remained at around 106MH/s. Power use also dropped to 279W, is quite good considering the hash rate.
Alternatively, you can go for broke and max out fan speed, set the power to 80%, drop the GPU clocks by 250-500MHz, and boost the VRAM clocks by 750-1000MHz. If you don't mod the card to improve GDDR6X cooling, you'll typically end up at 106-110C (depending on your card, case, cooling, and other factors), and the maxed out fan speed is not going to be good for fan longevity. Consider yourself warned.
GeForce RTX 3080 Founders Edition: Tuning this card was very similar to the 3090 FE. It doesn't like stock settings, as the GDDR6X gets very toasty. We again dropped the GPU core the maximum allowed (-502MHz), set the memory clock to +750MHz, and put the power limit at 70%. That resulted in the same GDDR6X temperature of 104C as the 3090 FE, and performance was only slightly slower at 93-94MH/s.
Again, maxing out fan speeds and memory clocks while dropping the GPU core clocks and power limit are key to improving overall hash rates. Modding the card and replacing the VRAM thermal pads with thicker/better pads is possible and will help cooling and performance. We'd prefer using an RTX 3080 with better GDDR6X cooling, however. Which brings us to...
Colorful RTX 3080 Vulcan: This card has better VRAM cooling than Nvidia's reference card, so the memory didn't get quite as hot. However, we still found we achieved the best result by dropping the power limit to 80-90% and then setting the GPU core clocks to the minimum possible value in MSI Afterburner (-502MHz). Then we overclocked the memory by 750MHz base clock, which gave a final speed of 20Gbps (the Ampere cards seem to run at 0.5Gbps below their rated memory speed when mining by default). That yielded average GPU clocks of just 1425MHz (down from 1920MHz), but overall performance increased to 93MH/s, while fan speed, GPU temperature, and power consumption all dropped. Most importantly, at the same (relatively) performance as the 3080 FE, GDDR6X temperatures stabilized at 100C. It's not ideal, but at these temperatures a 4C difference can be significant.
The RTX 3070 and RTX 3060 Ti have the same 8GB of 14Gbps GDDR6, and as we'll see with the AMD GPUs, that appears to be the limiting factor. Our initial results were poor, as these were the first cards we tested, but we've revisited the settings after looking at the RX 6000 series.
GeForce RTX 3070 Founders Edition: The main thing for improving performance on the 3070 was to boost the GDDR6 clock. We were able to add 1200MHz, giving a 16.4Gbps effective speed in theory, but the memory actually ran at 16Gbps (vs. 13.6Gbps at stock settings). Along with the boost to memory clocks, we dropped the GPU clock down to the maximum -502MHz in Afterburner and set the power limit at 55%. That resulted in actual GPU clocks of 960MHz on average. You'd think that wouldn't be sufficient, but boosting the GPU clocks up to 1.9GHz resulted in the same performance while substantially increasing the amount of power used. At the 55%, meanwhile, the 3070 is second only to its little sibling in overall efficiency.
GeForce RTX 3060 Ti Founders Edition: As with the 3070, we bumped the memory speed up as the main change to improve performance. We went from stock to +1200MHz (compared to +750, which is about as far as we could go for gaming overclocks), giving a maximum speed of 16.4Gbps, with a 400MHz negative offset for GPU compute workloads. Other settings were similar: -502MHz GPU clock, 55% power limit, and 50% fan speed. Performance was very close to the 3070 while using less power, making this the overall winner in efficiency.
Asus RTX 3060 Ti TUF Gaming OC: This is mostly here to show that, unlike the 3080 and 3090, third party cards aren't markedly different in hashing performance with the 3060 Ti GPU. Our tuned settings ended up with higher clocks (due to the factory overclock) and more power use than the 3060 Ti Founders Edition, but basically the same hashing performance. Optimal efficiency can vary a bit, but for the GA104-based cards, GDDR6 speed is the limiting factor on performance.
Radeon RX 6900 XT (Reference): Tuning all three of AMD's reference RX 6000 cards ended up very similar. The GPU clocks can go very high at stock, but the memory bandwidth appears to be the main bottleneck. Running with GPU clocks of 2.2-2.5GHz just wastes power and generates heat without improving performance. We cranked the power limit to the maximum 115% just to ensure the VRAM wasn't being held back, then set the memory at +150MHz (the maximum allowed in Radeon Settings), enabled fast RAM timings, and dropped the maximum GPU clock down to 70%. That gave us final clocks of 1747MHz compared to 2289MHz at stock and about 8% higher hash rates overall. More importantly, power consumption took a massive dive, and efficiency improved to one of the better results in our testing. But this actually isn't AMD's best overall showing.
Radeon RX 6800 XT (Reference): Same approach as above, but due to differences in the core configuration and … honestly, we're not sure what the deal is, but we ended up with an optimal maximum GPU frequency setting of only 50% this time, which gave us clocks of 1206MHz instead of 2434MHz — and performance still went up, matching the RX 6900 XT and RX 6800. At the same time, power requirements dropped substantially, from 281W to 186W. Whatever is going on behind the scenes, it appears different AMD Navi 21 GPUs run optimally at different "maximum frequency" settings. Like Nvidia, AMD's GPUs are largely limited in performance by their memory speed, and without tools to overclock beyond 17.2Gbps, there's not much to be done.
Radeon RX 6800 (Reference): With only 60 CUs (compared to 72 on the 6800 XT and 80 on the 6900 XT), you might expect the 6800 vanilla card to end up slower. However, the memory proves the deciding factor once again. We set the GPU power limit at the same 115%, which does make a difference, oddly — average power dropped about 15W if we set it to 100%, even though the card was running well below the official 250W TGP. We also maxed out the memory slider at +150MHz (17.2Gbps effective), and this time achieved optimal performance with the GPU set to 75% on maximum clocks. That resulted in a 1747MHz clock compared to 2289MHz at stock, but fan speed was higher this time. That's because we set the fan to run at 40% at 50C, 60% at 60C, 80% at 70C, and 100% at 80C — and it ended up at 50% speed, which is perhaps more than is required, but we feel it's better safe than sorry if you're looking at 24/7 mining.
Comparing Profitability and Performance
We're just getting started, but after all of the testing we've completed so far, one thing we want to point out is that the NiceHash Profitability Calculator is … well, let's just call it generous. It's not like NiceHash has a vested interest — like a few percent of all commissions — to get people mining with its software, right? Actually, it's not all bad, but our results definitely don't match up on some GPUs. Here's the quick summary of discrepancies, and again, this is without accounting for PSU inefficiency. We're using the optimal tuned settings for our own results, and obviously some cards with the same GPU perform better than others. But we're mostly looking at reference models, and our results don't quite match up with NiceHash's figures.
| TH Hashrate | NH Hashrate | % Difference | TH Power | NH Power | % Difference | |
|---|---|---|---|---|---|---|
| RTX 3090 FE | 106.5 | 120 | -11.3% | 279 | 300 | -7.0% |
| RTX 3080 FE | 93.9 | 98 | -4.2% | 234 | 240 | -2.5% |
| RTX 3070 FE | 61.3 | 60.6 | 1.2% | 123 | 160 | -23.1% |
| RTX 3060 Ti FE | 60.6 | 60.5 | 0.2% | 116 | 180 | -35.6% |
| Colorful RTX 3080 | 93.4 | 98 | -4.7% | 221 | 240 | -7.9% |
| Asus RTX 3060 Ti | 60.7 | 60.5 | 0.3% | 127 | 180 | -29.4% |
| RX 6900 XT | 64.6 | 64 | 0.90% | 183.3 | 220 | -16.70% |
| RX 6800 XT | 64.5 | 64.4 | 0.20% | 186.3 | 190 | -1.90% |
| RX 6800 | 64.6 | 63.4 | 1.90% | 166.3 | 175 | -5.00% |
There are definitely some interesting results. The RTX 3090 Founders Edition is definitely not the best sample of mining performance, and NiceHash's number (perhaps with maxed out fans) is 12% higher than what we got, but also used 7% more power — or if you prefer, our numbers were 11% slower while using 7% less power. On the 3080, the FE ended up just 4% slower while using 3% less power, but the Colorful card was 5% slower while using 8% less power. That's one of the clear cases where the model of the card used makes a big difference.
NiceHash specifically recommends running your fans at 90-100% on the 3080 and 3090, which can definitely boost performance. Our stance is that this is a Very Bad Idea (tm). Not only will the fans make a lot of noise, but they're also destined to fail sooner rather than later. If you're okay replacing the card's fans in the future, or if you want to mod the card with better cooling pads in the first place, you can definitely achieve the NiceHash performance figures. Power use (as measured using Powenetics) is higher than quoted, but not drastically so.
We were relatively close on the 3060 Ti performance, but here's where things get interesting. Our 3060 Ti and 3070 power use were much lower than NiceHash's 160W and 180W figures, with the 3060 Ti using less power instead of more. It's unfortunate there's no way to check memory temperatures on the 3070 and 3060 Ti. Presumably they'll be much lower than GDDR6X, and we're not totally sold about the 16Gbps being fully stable on these cards. Still, assuming you can find the 3060 Ti at a reasonable price, it's potentially one of the best options.
Shifting over to AMD, things are very close on performance — we're within 2% across all three cards and actually came out slightly ahead. As with the 3060 Ti and 3070, we also achieved our results while using less power than what NiceHash shows. Of the three cards, the vanilla RX 6800 comes out at the top, thanks to it's lower power use and potentially lower price.
The current thinking for a lot of miners is that Nvidia's RTX 30-series cards are superior to AMD, but that's really only true if you look at pure hashrates on the 3080 and 3090. Factor in power efficiency, and things are much closer. Not like you can buy any of these GPUs right now — unless you're willing to fork out a lot of money or have some good industry contacts for building your mining farm.
Mining with Previous Generation Hardware
Nvidia Ampere and AMD RDNA2, aka Big Navi, GPUs are getting most of the headlines today, but what about previous generation hardware? The principles we've outlined above generally apply to the older GPUs as well. We've included Nvidia's RTX 20-series and GTX 10-series cards, along with AMD's RX 5000, Vega, and 500 series parts. We'll see about adding the last few remaining cards (GTX 1660 series) tomorrow.
We're going to skip all the baseline performance metrics this time, and just jump straight to optimized performance. Note that outside of the RX 580 and 590, all of our tests were done using the reference models from AMD and Nvidia, which tend to be more of a baseline or worst-case scenario for most GPUs. We'll list our optimized settings below, but here are the results (so far).
For Nvidia's Turing GPUs, performance again correlates pretty much directly with memory bandwidth, though with a few interesting exceptions. Notice that the 2060 Super, 2070 Super, and 2080 Super all end up with nearly identical performance? That's not an error. The odd bit is that the 2080 Super requires substantially higher memory clocks to get there.
The 2060 Super and 2070 Super both had to run with a +1100 MHz offset in MSI Afterburner, giving an effective speed of 15.8Gbps (with a 400Mbps negative offset for compute). The 2080 Super meanwhile has a base GDDR6 speed of 15.5Gbps and we were able to max out the memory overclock at +1500 MHz, yielding a final speed of 18Gbps for compute. Except we still ended up at the same 44MH/s hash rate. Most likely the memory timings on the GDDR6 in the 2080 Super are more relaxed (looser), so even though bandwidth is higher, bandwidth plus latency ends up balancing out. 2080 Ti meanwhile hits the same ~60MH/s as the 3060 Ti and 3070, thanks to its 352-bit memory interface.
Stepping back one generation further to Pascal (GTX 10-series), the approach changes a bit. Maximum memory clocks are still critical, but core clocks start to matter more—the architecture isn't tuned for compute as much as Turing and Ampere. We got our best results by overclocking the GPU core and memory speed, but then setting a power limit. Except, being four years old, two of our GPUs (the GTX 1070 and GTX 1080) really weren't happy with memory overclocking. Anything more than a 200MHz bump on the 1080 caused a hard PC crash, and while the 1070 managed +500MHz, our hashing results were still a bit lower than expected.
Tweaking AMD's previous generation GPUs is a bit different, in that rather than limiting the power, the best approach is to tune the voltage and clock speed. Not surprisingly, the older the GPUs get, the lower the hash rates and efficiency become. Let's start with the previous generation and move back from there.
The RX 5000 series GPUs (aka RDNA, aka Navi 1x) all continue to perform quite well in Ethereum mining. RX 5700 XT and RX 5700 are nearly as fast as the latest RDNA2 chips, the main difference being in memory bandwidth. Most cards won't go much beyond around 1800MHz on GDDR6 clocks, which is equivalent to 14.4Gbps, though some might be able to do 14.8Gbps. The RX 5600 XT is the same GPU but with a 192-bit memory interface, which cuts performance down by about 25% again, right in step with memory bandwidth. Overall, the 5700 and 5700 XT end up as the most efficient previous generation GPUs for mining.
Move back to the Vega architecture and the large memory bandwidth that comes from HBM2 comes into play. The Radeon VII with its monster 4096-bit memory interface and a modest overclock for good measure ends up with 1178 GBps of bandwidth—only the RTX 3090 comes close. But Vega was also a power-hungry architecture, and it benefits from turning down the GPU clocks. We ended up at a 1650MHz setting (stock is 1800MHz) on the Radeon VII, and we set the core voltage to 880mV. That gave mining clocks of 1620MHz.
Vega 64 and Vega 56 used similar settings, but half the memory capacity and bus width limits performance quite a bit relative to the Radeon VII. Also, our results on the reference blower cards are probably far less than ideal—just about any custom Vega card would be a better choice than these blowers.
And then there's Polaris. Much like the Pascal GPUs, our tuning efforts took more time and effort. Many places recommend VBIOS flashing, but that's not something we want to do (as we're just testing mining performance, not building a 24/7 mining farm). As such, our maximum hash rates ended up about 10% lower than what some places show. Besides setting a low voltage of 800-900mV, depending on the card, we set the memory timings to level2 in Radeon Settings, and that gave the best results with reasonable power use. YMMV.
Here's a gallery of all the 'tuned' settings we used for the legacy cards. Use at your own risk, and know that some cards prefer different miner software (or simply fail to work with certain miners).
Should You Start Mining?
This brings us to the final point we want to make. Suppose you already have a graphics card and want to mine using your GPU's spare cycles. In that case, it's certainly worth considering, particularly if you live in an area where power isn't super expensive. You could theoretically net over $10 per day from the right 3080 or 3090 card, and $7.50 or so per day with one of the Big Navi or lower-tier Ampere GPUs. If you're looking for the best option for mining, from the perspective of paying off the GPU as quickly as possible, the RTX 3060 Ti would be the card to get — or at least, it would be if you could find it anywhere for less than $900. No, that's not a typo, and no, you shouldn't pay $900 for a 3060 Ti!
At the same time, we strongly recommend against 'redlining' your card for maximum hash rate at all costs. The fans on consumer cards aren't designed to spin at 100%, 24/7 without failing. They will burn out if you run them that way. We also have serious concerns with any component temperature that's consistently at or above 100C (or really, even 90C). But if you do some tuning to get fan speeds down to the 40-50% range, with temperatures below 70C, you can probably mine with a card for quite a while without having it go belly up. Will it be long enough to recover the cost of the card? That's the big unknown.
Here's the thing to remember: Cryptocurrencies are extremely volatile. Right now, Ethereum's price is up 67% just in the past month, and often swings by 5-10% (or more) in one day. It's also up 300% during the past quarter, and 750% since one year ago. This means that, as fast as the price has shot up, it could plummet just as quickly. If you're looking at $10 per day per GPU in mining profits and thinking it's only three months to break even on some GPUs, you're assuming the price and profitability will remain steady. It won't. There's a chance it could go up, but the opposite is more likely given the recent jump in pricing. Just ask the GameStop 'investors' how that worked out if you think the sky's the limit.
Again, if you already have a GPU, putting it into service isn't a terrible idea — it's your hardware, do with it what you please. Paying extreme prices for mid-range hardware to try and build your own personal mining mecca, on the other hand, is a big risk. You might do fine, you might do great, or you might end up with a lot of extra PC hardware and debt. Plus, what about all the gamers that would love to buy a new GPU right now and they can't? Somebody, please think of the gamers!
Anyway, there's more to come. We'll check out some previous-generation GPUs as well, and RX 5700 series in particular seems promising. If you've got one of the previous generation Turing or Navi cards already, you might not be able to make quite as much from mining, but even GTX 1070 and RX 570 8GB cards can currently do over $3 per day. Just remember to account for power costs and cash out enough coins to cover that, and then hopefully you won't get caught holding the bag.
https://www.tomshardware.com/how-to/optimize-your-gpu-for-ethereum-mining
Comentários
Enviar um comentário