CryptoCurrency Mining with GPUs *spawn*

Admittedly I don't know the algorithms well enough to name a specific handicap. But perhaps there's a specific instruction or two that's really only used by cryptocurrency mining; a bit rotate or such?

I think its the pure theoretical flops the cards can do and driver optimizations, the gtx 1070 hits above 31 mh/s in eth whiling using 140 watts of power (cuda based with linux, based on different diver versions newer drivers can slow it down some), the rx 580 just under that at 30 mh/s in eth, at around 130 watts (open cl based with windows or linux). Both those cards have similar tflops
 
Ethermining really seems to take a hit with GDDR5X and HBM. 1070 is faster than 1080 and 480 faster than Fiji (ex. a Titan X-P, the older one, is pretty much saturated with 1.4 GHz core clock, but scales almost linearly with memory. A model running here right now with 1.36/"5.4" GHz nets me ~35 MH/s. With default memclk which goes down to 4,5 GT/s, it's 30 MH/s).

My guess: VERY unordered and hard-to-coalesce memory access patterns, wasting many many of the potentially parallel bits transferred.
 
Last edited:
Etheriming really seems to take a hit with GDDR5X and HBM. 1070 is faster than 1080 and 480 faster than Fiji (ex. a Titan X-P, the older one, is pretty much saturated with 1.4 GHz core clock, but scales almost linearly with memory. A model running here right now with 1.36/"5.4" GHz nets me ~35 MH/s. With default memclk which goes down to 4,5 GT/s, it's 30 MH/s).

My guess: VERY unordered and hard-to-coalesce memory access patterns, wasting many many of the potentially parallel bits transferred.
Ethereum though has performance issues anyway with Nvidia unless one uses Linux, that is the only way to get competitive performance with this algo.
Cheers
 
Ethereum though has performance issues anyway with Nvidia unless one uses Linux, that is the only way to get competitive performance with this algo.
Cheers
That's rather old info from around early summer of last year, when Linux was four or more times faster than Windows. For months now, GTX 1070 and RX 480 perform quite similar around the 30-ish MH/s-mark. edit: Individual configuration or hardware issues notwithstanding

edit 2
Just now running a GTX 1070 FE with a bit of clock-tuning (less core, more mem) at 29,5 MHashes/s under windows 10. With an average power consumption of 95 watts according to hwinfo64. So it's no longer "broken by design", but "hampered by memory (controller)".
 
Last edited:
That's rather old info from around early summer of last year, when Linux was four or more times faster than Windows. For months now, GTX 1070 and RX 480 perform quite similar around the 30-ish MH/s-mark. edit: Individual configuration or hardware issues notwithstanding
Agreed but was it fully fixed; Claymore algo (for Ethereum) on good Win10 setups with 1070 can hit 30MHs, on Linux back when the issue raised you could get around 35- 40 MHs with similar setups (to those hitting 30MHs with Claymore).
That Ethereum figure in the link is still low when looking at it from crypto mining sites, and the 1070 is notably competitive with many other algos on Windows albeit with a higher starting cost relative to AMD.

But the point is whether the issue with GDDR5X and Ethereum also came back to that original performance issue with Ethereum as well under Windows *shrug*.
Cheers
 
Last edited:
Sorry for OT, but since this concerns comparison to Polaris, which works in a similar ballpark, I'm interested here as well in the purported Nvidia numbers.
The best I could find is this:
http://bitcoinist.com/gtx-1070-optimized-ethereum/
„After several avenues had been pursued to squeeze every last bit of Ethereum hashpower out of our EVGA test GPU, we came out with an optimized configuration that hashes at 32.4 MH/s, pulling 146 watts under load, and getting up to a maximum of 63 degrees in out test bed case.“
I'd be curious where you got that 40 MH/s number from? IIRC, Claymore has been one of the fastest ethermining programs.
 
Sorry for OT, but since this concerns comparison to Polaris, which works in a similar ballpark, I'm interested here as well in the purported Nvidia numbers.
The best I could find is this:
http://bitcoinist.com/gtx-1070-optimized-ethereum/
„After several avenues had been pursued to squeeze every last bit of Ethereum hashpower out of our EVGA test GPU, we came out with an optimized configuration that hashes at 32.4 MH/s, pulling 146 watts under load, and getting up to a maximum of 63 degrees in out test bed case.“
I'd be curious where you got that 40 MH/s number from? IIRC, Claymore has been one of the fastest ethermining programs.
Was 35-40MH/s (edited before you responded) and I will need to dig it out as you appreciate not many crypto miners with great experience of Linux and Nvidia (including CUDA).

Quite ironically I just noticed that even in the Ethereum blog you linked regarding the Ethereum issue kruisdraad additionally said in comments afterwards.
kruisdraad July 2016 said:
30MH/s with stock 1080 is not bad actually. please note: the 1080 has less good memory then a 1070, but i got the 1070 to a stable 36,1MH/s ... but windows might be a little better and with stable i mean unattended mining. If you dont care about a reboot (you need a SSD for fast boots) you can get a lot better.

The 35 MH/s to 40 MH/s I am thinking of was a different source, there is also the CUDA version to consider; 7.5 vs 8.0.
And this was back in 2016 for what I am thinking and also the kruisdraad quote.
Cheers
 
Last edited:
Admittedly I don't know the algorithms well enough to name a specific handicap. But perhaps there's a specific instruction or two that's really only used by cryptocurrency mining; a bit rotate or such?
What little I've skimmed about GPU mining is that it relies heavily on generating cryptographic hashes of a data set purposefully sized to blow out on-die storage while being modest enough to fit in on-board memory. The overall scheme is apparently designed to exceed what an FPGA or ASIC can keep on-die, while not being scalable to large amounts of parallel compute resources not tied to a common memory and its bandwidth, to help limit what a large cluster might do.
Ethereum seems to use a hash in the family of SHA3, which has a lot of iterative XORs and integer math applied repeatedly to a region of data.

Individual instructions may not really stand out, other than the workload not using FP or any graphics functionality--although simple checks on those items could be readily gamed or have false positives on non-mining workloads or portions of them.

The context of the workflow, which I think is a lot of iterative integer and logic math pulling in streams of memory and iteratively overwriting sections of storage might stand out. It might take something like static analysis of the program or identification of specific programs, the latter check being trivial to defeat and the former probably unlikely to withstand the monetary motivation to defeat it unless the whole GPU and software system were locked down.
I am not sure if there aren't some functions (noise functions?) with iterative math that might yield false positives as well.

The GPU front ends would presently not have much visibility to monitor the behavior of the wavefronts being launched, so their ability to throttle wavefront launches based on the presence of mining-type wavefronts would be low. Their general programmability and microcode store might let them host some more advanced heuristics, if access to such data became available.
The CUs would have more dynamic information, perhaps a flood of XOR operations and shifts in the instruction stream and a distinctive pattern of memory accesses. Perhaps instruction issue could throttle if a count of the various operations or events was accumulated over a time window, but that might be a level of context and tracking the streamlined hardware does not maintain.

Perhaps a more sure way to encourage a mining SKU would be to purposefully double down, and implement something akin to the SHA extensions in x86 for the GPU, and have that subject to SKU-specific throttling like DP. If used in a security context, cryptographic functions often purposefully stall to keep execution times flat and avoid side-channel leaks--which a mining SKU could dispense with.
Getting the level of specificity to the workload given its propensity to not use standards, or to change what it uses if the overall situation is "unfair" (why cryptocurrencies spawned from Bitcoin after ASICs took over, tweaked currencies having low cost to launch) makes implementing higher-level primitives in hardware such as this an uncertain affair.


That's a really good point. I didn't realize NV cards were that close on this generation.

Though the irony is that NVIDIA is probably better suited to handling this anyhow; besides a much larger volume to start with, they have a lot more experience enforcing market segmentation. Which I suppose puts AMD in a no-win scenario...
It seems to me that an issue AMD is facing is that GPU mining specifically targets a rather specific and yet non-differentiating element of GPU architecture that doesn't provide much of a knob or lever that makes it compelling in the other markets that more reliably set GPU prices or create long-term buy-in (professional applications, virtualization, games, machine learning, HPC, rich media, driver support, APIs, blue-sky research, devrel etc.).

Having higher than average local bandwidth and generic operation bandwidth in the design space of a consumer GPU is something that is overly fungible, and emphasizing that in excess of its general utilization of those other elements that make up the general utility/value-add is actually reminiscent of GCN.

I don't have much faith in mining yielding dividends for AMD. Block chains and distributed ledgers used by cryptocurrencies are finding use or will soon for monetary or business transactions, but I don't know enough about the space to know what GPUs offer if the mining component is dispensed with.
 
I'm reasonably sure that NVidia gamers are mining too. You just game when you're not mining. The AMD gamers will just be paying off their card slightly faster when doing the same.

Also if you're a gamer with two cards, you can mine on both cards and game on one card for those few hours per week you do actual gaming. See, there is a reason to run Crossfire/SLI :devilish:

This chart shows how recent the craze is (but it doesn't show the craze from last year just before Polaris arrived, and which did affect Polaris when it arrived):

https://www.tradingview.com/chart/ETHBTC/

If you'd bought a good card dedicated to mining at the beginning of March then it should have paid for itself in the last few days, if you pay typical American rates for electricity. If you mine but don't sell straight away, then you would have paid off the card much earlier. This is the miner's dilemma: sell as I mine or hold in the hope that the coin appreciates?

Still, I don't think most gamers really understand this: if you bought the card for gaming, then this is free money. Non trivial money, too. You do need to be prepared to convert mined coins to another currency if you can't spend those coins. Or you can get the pool you're mining at to sell your coins for BTC, say. And you might have fan failure earlier than expected. GPU chips/VRMs are almost never the reason for card failure.

In general mining doesn't harm cards: they will run just as well at stock settings after years of mining as they did when new. Electromigration is a factor, but cards' stock settings take that into account.

Before Polaris arrived, Fury Nano was the most efficient Ethereum miner, so HBM isn't necessarily a wipeout. But, if HBM is enough of a demerit for mining amongst the current set of cards, then Vega will see no miner uptake. Miners aren't stupid.
 
Free money, except the electricity... But yeah even with that you're "wining" most of the time. I'm testing ether with my fury x (in a custom loop). If i believe the gpu sensors, it's running at approx 185w maximum, which is ok. Classic power usage when only the "compute parts" of the chip is used, like with Folding@home&co. The client (i'm running the classic ethminer) seems pretty stable so far.
 
There are more intricacies, it seems. Just finished running a very short test before leaving for the weekend, and apparently, the memory manufacturer amongst RX580 cards plays a role as well. Saw a difference close to 10 percent with three 580's running at 1,3 GHz/2,0(4,0) GHz, with SK-Hynix memory being handed the short end of the stick.

Regarding Fiji: Yes, it's true that it was the fastest Ethermining card, but given it's TFLOPS rating, it should have been leaps and bounds ahead of Polaris, yet it was only marginally faster. IIRC without many optimizations, you could get around 30-ish MH/s as well.

That said, running a GTX 1080 with the same level of "mojo" (i.e. my limited Claymore-GPU-Miner setup skills and some memory OC), I arrive at 24 MH/s instead of close to 30 with the 1070.
 
Actually just built my first 6 card rig, its getting 1k a month, rx 580's, all of them getting 29mhs per sec dual mining eth and pascal, so its pretty profitably right now. For how long not sure, cause Eth is moving over to proof of stake instead of proof of work early next year or so. At that point eth prices are probably going to sky rocket. Then I have to find another coin of course to mine. My second rig which I'm building now is going to be a 1070 rig, just ordered all the parts and one 1070 to test it out and see how far I can push it in reasonable amount of power 150 watts each card is my goal. I'm expecting around 35mhs per card. Which yeah rate of return is higher but initial cost is more than the return rate increase. But if it works out well, I'm going to building 10 rigs by 3 months, then get a wearhouse and really build up fast, mid next year 100 rigs maybe? Its not hard to do, the software is stable, systems are stable, have them on autorestart every 24 hours. Have network monitoring over my phone and computer also with alerts with bright flashing lights lol. Its all broken down to a per rig per miner alerts, so its pretty cool and easy to see what is going on when something goes tits up.

Even by early next year when I have to switch over to another coin, my goal is to make sure I pay off my investment within the first 6 months, and then the rest of the coins just save up for the market to pick up and make a solid amount of cash. No skin off my back other then keeping things up and going on very stable systems.

Ddin't have time to finish up what I wanted to state,

Ok, dividends wise, it does pay AMD up front by sucking up all the stock of their cards, but the problem is, gamers when they can't get those AMD cards, they choose nV cards, then those customers will they go back to AMD later on? Probably not till the next generation, and at that point, what is the retention rate of people with nV cards? Its pretty high if I remember correctly.
 
Last edited:
From what I read, Zcash will be the next thing when/if eth move to PoS ... ? If true, It could be cool, since you can gpu mine it too, and is asic "resistant".
 
VCZ has a "rumor" that AMD & Nvidia will both release cheaper crypto currency-optimized versions of their GPUs.

https://videocardz.com/70162/amd-and-nvidia-preparing-graphics-cards-for-cryptocurrency-mining

Considering how this topic was just being discussed on B3D (particularly a poignant post from Mr Ryan Smith), I'm wondering if this is a real rumor or just some feedback loop shenanigans.

I just don't think it's in AMD's (or Nvidia's) best interest to sell a cheaper crypto card. But if they can't find a way to gimp gaming cards, then maybe this is for the best.

EDIT It just dawned on me that I probably should've put this is that crypto thread...
 
VCZ has a "rumor" that AMD & Nvidia will both release cheaper crypto currency-optimized versions of their GPUs.

https://videocardz.com/70162/amd-and-nvidia-preparing-graphics-cards-for-cryptocurrency-mining

Considering how this topic was just being discussed on B3D (particularly a poignant post from Mr Ryan Smith), I'm wondering if this is a real rumor or just some feedback loop shenanigans.

I just don't think it's in AMD's (or Nvidia's) best interest to sell a cheaper crypto card. But if they can't find a way to gimp gaming cards, then maybe this is for the best.

EDIT It just dawned on me that I probably should've put this is that crypto thread...

I think it's complete BS.

You also have to realize this dedicated mining GPUs with 'no video outputs' would have 0 resale value. One of great benefits of GPU mining (other than dual use for gaming) compared to ASICs/other dedicated miners is resale value.
 
I think it's complete BS.

You also have to realize this dedicated mining GPUs with 'no video outputs' would have 0 resale value. One of great benefits of GPU mining (other than dual use for gaming) compared to ASICs/other dedicated miners is resale value.

Conversely, one of the risks with buying certain used cards in the possibility that it ran at full load 24-7 for most of its life up until that point.
 
From what I read, Zcash will be the next thing when/if eth move to PoS ... ? If true, It could be cool, since you can gpu mine it too, and is asic "resistant".


Yeah that is true but the problem with Zcash right now is the transaction fees and volatility, there is no way to directly convert Zcash to cash, so volatility in the marketplace can drop your ROI considerably. This is the big plus for Eth mining as there are wallets you can directly convert to cash.

So there are three coins of eth switches to PoS, Monero, ETC, and Zcash. With ETC you can dual mine with one of the less profitable and newer alt coins, like pascal or sia coin which increases profitability or just holding on to the second coin till they go up in price.
 
What do you mean "there is no way to directly convert Zcash to cash" ?

With ETH, I have to do myetherwallet (or whatever wallet you use) => kraken => sold them => withdraw => kraken to my bank account.

What is different with Zcash ?
 
I use coinbase, I can sell eth at market value and get that, Zcash, can't do that, I have to convert it to Bit coin or Eth then send it to my wallet and then cash out :). So there has to be an exchange in the middle for Zcash, and need to trade it for another coin where possibility is to loss money based on what the others want to buy it for.

If there is a wallet that can send directly to my bank with Zcash, please let me know lol, I'll be all over that hehe
 
Back
Top