Bitcoin Forum

I have seen people saying that the GCN architecture is not good at mining comparing to VLIW architecture.
The 7970 seems to do quite well in mining.
Made me wonder, from architecture point of view, is it really a bad architecture for mining?  Can it be better than the old architecture?

It's not bad, it's just not as "optimal" as VLIW for bitcoin mining.

If you take all the scheduling hardware in a design like GCN and removed it and replaced it with more ALUs, you would get even higher performance.

Well, the performance is actually roughly in ratio to the increase in shader count and frequency. It's more of some disappointment that it didn't improve on bitcoin mining over the 5800 series. Combine that with the relatively high prices and the main reason to go GCN would be for MH per PCI-E slot density.

It's significantly more power-efficient than 58xx just from the process shrink...

True to an extent, more MH/W but not by as much as I would have thought going from 40-28nm.

https://bitcointalk.org/index.php?topic=57410.0

You can't compare the MH/J at 550 MH/s. The 58xx if you could clock it that high would require MASSIVE power consumption.

If you were to take both cards, set a target MH/s and then compare on MH/J, the 7970 would completely dominate the 58xx.

Going to take quite some time to earn back that $300 price difference through power savings...

No one smart mines a graphics card until it is worth nothing... what you really need to do is compare ROI.

In 6 months, let's assume a 7970 will lose $50 in value, while a 5830 wont lose anything:

As such, we should consider an analysis like this:

A 4x5830 does 1100 MH/s at ~550W and costs ~$800 to buy the system
A 4x7970 does 2880 MH/s at ~1100W and costs ~$2100 to buy the system

Net revenue over 6 months (at current diff):
5830: 127.44 BTC
7970: 333.72 BTC

At current exchange rate ($5.6/BTC):
5830 revenue: $713.27
7970 revenue: $1868.83

Power costs (6 months or 182 days):
5830: 2402.4 kWh
7970: 4804.8 kWh

At 15c/kWh:
5830 cost: $360.36
7970 cost: $720.72

So net profit is:
5830: $713.27 mined - $360.36 power cost - $0 depreciation = $352.91
7970: $1868.83 mined - $720.72 power cost - $200 depreciation = $948.11

ROI:

5830: $352.91 / $800 = 44.1% ROI
7970: $948.11 / $2100 = 45.1% ROI

As shown above, GCN is definitely not inferior to VLIW5 as far as mining goes. Note that I have not included cooling costs, as that would skew the ROI a couple more % towards 7970 nor did I take into account the reduced maintenance nightmare from running less rigs. On the other hand, if your power is significantly cheaper, the ROI% moves towards 5830.

Also, one could make the observation that 58xx pricing is mostly driven by mining demand. Thus the ROI % being so close together is no coincidence.

If the 5830s were being undervalued, demand and price will increase, until the ROI% dropped down to 7970 levels.

If the 5830s were being overvalued, demand and price will decrease, until the ROI% dropped down to 7970 levels.

This is simply market economics at play here.

So we are in agreement that the main advantage is slot density. Secondary advantage for newer cards is less dependance on bitcoin mining value for resale purposes.

No, I'm just saying that at over a certain $/kWh, 7970 is the solution that will give you a higher ROI% on your investment due to its higher MH/J, even if it costs more initially.

Over a certain $/kWh you probably shouldn't be mining with GPU to begin with.

Yes, but that point is much higher than 58xx/7970 point. I'm not sure wtf the point of this post was.

The real issue up to this point is that there has been more market drive for solutions with more functional blocks than ALUs.  I expect this will change heavily within a year if we levy some of the major silicon suppliers, or some of you that have been in this from the beginning set up a business model that could rival the giants.  FPGAs and ASICs designed with mining in mind would offer substantially better margins than any equipment out there now.

As far as GCN goes, I have a tahiti, and have not found it to be substantially more efficient with current software options,  Of course until you an get a transistor by transistor map of what is in a chip, there's little to be said publicly about how to improve an option.  Though bitcoin mining of course favors clean manipulatable logic cells over all the fancy bits in modern silicon solutions.