AMD: Southern Islands (7*** series) Speculation/ Rumour Thread

Is it still true that higher leakage parts have lower volts while lower leakage have higher voltage? If so what's the voltage for the lower leakage parts? I've been reading posts of people reporting 1.170v but wasn't sure which of the 2 that belongs to.

 

Came across some Bitcoin mining results : https://bitcointalk.org/index.php?topic=57410.0

Check out those chip undervolting results with memory underclocking... DAYUM! Looks almost like AMD won't need to drop the clocks a lot for the dual card - just bin carefully and use lower stock voltage...

PS. If you read the thread further, to around 7th page, you might understand why the 7970's are out of stock.

 

If you mean https://bitcointalk.org/index.php?topic=57410.msg689045#msg689045 they're not 7970's, they're HD5-series (just check the cooler, also, the display outputs)

 

No, I mean this: https://bitcointalk.org/index.php?topic=57410.msg692501#msg692501

 

The mining craze is really confusing. Fixed payouts shared across the network and a constant scramble for faster hardware so you don't get left behind. That's some pretty fuzzy math.

Seems Tahiti needs a new miner to unlock its potential - http://www.tomshardware.com/reviews/radeon-hd-7970-benchmark-tahiti-gcn,3104-14.html

 

Seems Tom's Hardware screwed up.
In that bitcoin forum thread they get 550-600 MHash/s @stock from a HD7970, not only ~415 like Tom's.

 

Oh, that's an alternative explanation That's exactly in line with the flops increase too.

 

I wonder if there'll be a version of SKA (or similar tool) that shows GCN-specific compilation...

Or maybe a way to dump the compiled kernel from the runtime?

 

it would be big surprise if Tahiti wasn't in line with flops increase, there is nothing worse compare to cayman, sp wise.. maybe THG used old drivers, dunno (bitcoin thread date 10Jan, THG 22Dec).. Undervolting is pretty damn impressive,though, i also use my gtx460 at 850mV @675MHz it draws ~100W at games ~120W at furmark.. it's one of the first things i generally do after determining the card's max oc

anyway im pretty sure AMD is gonna release the XTX if they need to.. not saying it's 2304sp or something but it probably will be 1200,1300MHz, with new silicon or binned part.. nearly all cards have vdroop, none of them are actually working when you select 1300mV and voltage drops to ~1210mV under load, altough temps are great..that causes most cards are stuck at 1200-1270MHz, i find it is an artificial limit.. looking forward to custom pcb's whether sitiuation is fixed or continue..

 

There's nothing fuzzy about:

http://blockchain.info/charts/miners-operating-profit-margin

from:

http://blockchain.info/stats

 

If they indeed release XTX it most likey be called HD7980.

Similar HD4870 to HD4890 transition of higher clocks.

 

Thanks. What's the cost input to that margin calculation? If it excludes capital costs then it's not very useful.

Oh found it. They use estimates of $1000 in hardware costs per giga hash over two years. Estimates of electricity and bandwidth costs are thrown in as well. I don't see how that measure can be even close to accurate without factoring in difficulty and network size. The aggregate profit margin might come close to that but a given individual will see very different returns.

 

In OpenCL, you can get the program binary back from the runtime, but, as yet, there are no (public) tools to decode the ISA for HD7970.

 

Difficulty = network size (effectively - difficulty is a function of a smoothed "moving average" of network size).

The graph I linked is a function of difficulty and price, too.

Browse around and you will see lots of interesting stuff on that site that backs-up that graph.

 

Understood. However it means that a given individual is screwed as the network grows since revenue (in bitcoins) is fixed regardless of network size. Aggregate capital and operating costs also increase, reducing overall profit margins. Now consider that some miners are far more productive than others and it's clear that there are many "losers" in this game. I'm looking at this purely from a personal enrichment standpoint.

Ok, I'll check it out. Wrong thread for this anyway.

 

Nothing is fixed in the bitcoin world; sometimes mining makes sense sometimes just buying them outright is the way to go. I bought 20 cards and haven't made a dime on them...but maybe if/when the bitcoin price goes up it will have payed off..but then again I should have just waited for the huge crash in price and bought the BTC's instead of mined them.

If you believe bitcoins are a good idea and will become more popular in the future it makes sense to just buy and hold. If you lack hobbies and want something to heat your house in the winter..and heat it in the summer...go hardcore mining. My 2c.

P.S. If you live in a humid climate and want to run your cards outside in the summer (to save AC cost), moisture WILL corrode the heatsinks on your cards. Mine look like they have been at the bottom of the ocean for a week..but they still work great so I guess I shouldn't complain so much.

 

Oh yes, I was never trying to suggest everyone is profiting to the same extent or that it's guaranteed. If you're mining with a Pentium IV then you're definitely in the "losing" camp

Small time miners will be squeezed out by industrial scale mining (technically in profit but with the barest margin) - if the latter transpires as worth doing, of course.

That graph does hide something like two orders of magnitude variation in the price of bitcoin. So reference to "revenue (in bitcoins) is fixed regardless of network size" is kinda missing the underlying action

 

There is already.

Spoiler

And it works, he also posted some code generated that way.

The dissassembler is actually integrated in the aticaldd.dll togther with the compiler. One just needs to call it on the binary (which isn't well documented). That's how I got some ISA code for GCN half a year ago.

But as the link above shows, the latest version of the SKA can also do it.

 

Yeah the $/bitcoin exchange rate is a huge factor. I havent been keeping up with that though so I'll assume the miners are entering and exiting the network as it fluctuates. Another equalizing force

 

IMHO,

not necessarily.
Let’s just pretend that they can reach 1400MHz with a new PCB layout and 2x8. Therefore such a card would be up to 50% faster than a HD7970. What would you prefer then, such a single-GPU card or a slightly (up to ~20% with perfect scaling) faster dual-GPU card with microstuttering and driver quirks?

IMHO AMD could therefore also bring a card with binned Thaiti GPUs @ up to 1400MHz and name it HD7990.