Topic: 7970 from newegg $405 - page 2. (Read 5462 times)

The last couple of posts in this thread have been talking about mining with efficiency (ie. You pay ~10c/kwh).

Where do you see 50W less per card at the same hashrate?


The 5970 & 7970 is roughly the same performance at high clock.  The 7970 has slightly higher efficiency at a modest undervolt but not this game changing nonsense you keep talking about.  Only at significantly lower clock/voltage the 7970 pulls ahead.  Of course at that significant of an undervolt the 7970 doesn't make much economical sense given FPGA still beat it in MH/W and underclocked it gives up a lot of MH/$.

It's not a dog. In fact SIMD is just VLIW with a scheduler slapped on top.

Talking about brute force, 3200 cores is not considered brute force? That's a >50% increase in the number of cores. Whereas the 7970 is clocked only 28% faster.


They're 7970s? Pretty much all of them are reference? What information were you looking for?

You guys have convinced me to drain my loop and do no-card/single-card tests tonight.

I think the biggest problem with his post is that I have not been able to replicate his numbers fully on any of the 7970s I have bought. Without model information the post by itself, while helpful, does not show the whole picture.

You must have missed this then. Unless you consider 50W per card not a significant power savings.

Of course. 

7000 series changed architectures and it is less favorable to mining.  Granted 7970 brute forces its way thought with higher clock & higher # of shaders but measured at the wall I haven't seen anyone "blow away" 5970s.

...

At least not yet.

4x5970
3.2 GH/s
1050W @ the wall.
~3.05 MH/J @ the wall.

The 7990 might change that.

From TSMC's website:

The 28nm high performance (HP) process is the first option to use high-k metal gate process technology. Featuring superior speed and performance, the 28HP process targets CPU, GPU, FPGA, PC, networking, and consumer electronics applications. The 28HP process supports a 45 percent speed improvement over the 40G process at the same leakage/gate.

7970: 4.5B transistors
5970: 4.3B transistors

So 7970 should be 45% faster (minus whatever overhead the scheduling hardware imparts on the architecture) at the same power levels.

I got what I would call a "bad" batch of Black Edition XFX 7970s. They were only able to do 1100/950 stable as well. The normal black editions, for me, were much better. For ~$99 less I think you got a decent deal.

Not to say that your analysis is worth less than mine, but I have taken several graduate-level courses in semiconductor device physics and very-large-scale-integrated circuit design.


Just so that we are clear with our definitions:
Static power is the power consumed by a CMOS transistor when it is "static" (ie not going from 1 to 0 and 0 to 1). Dynamic power is the power used by a CMOS transistor when switching from 0 to 1 or 1 to 0.

There is no "sweet spot" in dynamic power usage when you look at it from a pure computational efficiency stand-point. Lower voltage (so less current leaks when you switch states) and lower frequency (less capacitive effect) is always more efficient (at least, until you reach the threshold voltage, then the device just stops working).

While it is true that static power consumption maybe lower on the 40 nm node, the lower threshold voltage on the 28nm node means that the transistor spends less time in the dynamic power usage. Engineers optimize static power usage and dynamic power usage using a variety of techniques, which is beyond the scope of this discussion. However, for a properly optimized process (ie. something a company would put out), the amount of energy (ie the sum of the static energy and dynamic energy) to do one unit work of computation always decreases as you shrink in process node (otherwise, there is little reason to shrink process nodes given modern ASIC's thermal-limitation).

well, i decided to install one in my gaming rig and it does work - so that is good.  the highest stable clocks (with stock voltage) i can run on it are 1100/950 and it is pulling 656Mh/s with cgminer.  

i don't know enough about these cards right now to judge if that is within expectations or not, but i am suspicious that these could have been returns they just stuck back in inventory.

I ordered 12 of them. I will check mine out to see if there are any problems.

i did grab 2 before the discount expired and they arrived today.  my first 7970's and any HIS card.  i am a bit leery of the state of the packaging (outer lame box tamper seal was not secure and the plastic 'coating' around the outer plastic shroud was not fully adhered) but will likely fire them up and see what happens........

It looks like at higher frequencies the 7970 will win out in efficiency, but a 5970 putting out 550 MH/s is only running at about 500 MHz and 0.9 volts, so dynamic power usage must be at a sweet spot.  Static power use of a semiconductor due to current leakage rises exponentially as the thickness of the insulators decrease, so as dies shrink, static power use can account for over 50% of a chips overall power usage.

I guess you can't write off the 5970s as long as you don't mind underclocking and undervolting.

Anyways, that's my analysis but take it with a grain of salt - I am an electrical engineer, but not a good one

I looked at making an account.  I don't have a business.  I don't see how you could set one up without lying through your teeth.  It requires things like business day and federal tax id# (EIN).  Making that up is asking for trouble IMHO.

Yup I dont trust the GPU-Z numbers anymore. I guess I'll just have to rip out all my cards the next time I redo my loop and measure the no-graphics card wattage. Anyways my total system consumption is 1100 MH/s at 460W which is kind of impressive considering the list of hardware I'm powering in my sig.

Yeah they are WC.

^This is the settings I'm using right now on my gaming/mining PC - just the one 7970 in there right now but my aim for the summer is to get three more and go quad-crossfire and watercooled - the thought of spending £300 just on the extra waterblocks I'll need is a tad intimidating though 

I wouldn't trust GPU-Z numbers and not really sure about the idle numbers, although 7970 does have zerocore and low idle consumption otherwise too. However if running multiple monitors or unofficial overclocking, those idle features may not work. That 7.5MH/J number would be pretty awesome, but I doubt you're getting anywhere near that...


Yours are watercooled though right? That could make a big difference due to lower temps/power consumption. Or were those numbers on stock cooling?

GPU-Z must be inaccurate then.

I did find this post by ArtForz:

https://bitcointalksearch.org/topic/m.709254


You seem to have a much lower undervoltage than he does, so maybe that's why your card matches 7970s in efficiency. ArtForz is showing 20% efficiency advantage to 7970.

Can't wait to see the 7990 efficiency numbers.

That's what it says on GPU-Z.  Actually, switching between cards (I didn't notice I could do that), I see some of the GPUs are at 37.5A.

Edit: And yeah, that's at roughly 277 MH/s per core.

28W*2 for 550 MH/s?

That's 19 9.5 MH/J, which is FPGA efficiency. Are you sure that's not idle?