Bitcoin Forum

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute (http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute)

To anyone who knows stuff about programming and computer circuits: Are the next generation of Radeons going to be just as good at mining as their predecessors were? Or will we see outrageous prices for used 5xxx and 6xxx cards in a few months?

Horrible. The 7 series will be the worst cards at mining period (AMD will actually block the SHA-2 family). Better save your money and get that Matrox...

whaaaaa? really?   ??? 

Also only comes with DX9 support since all games are console ports anyway.

"At no point has AMD specified when a GPU using GCN will appear, so it’s very much a guessing game". It is rumored and extremely likely that the 7xxx GPUs, which have reportedly taped out and are quite possibly in production, are NOT this "future generation" AMD GPU. They will be based on a die-shrunk and likely tweaked 5xxx/6xxx-like GPU, meaning they should have similar performance characteristics to older generations while likely improving some workloads due to evolutionary changes, and improving performance/watt because of the die shrink.

No one knows how they will perform in any given situation of course, but I would expect them to be comparable to previous generations.

+1 to this guy. Sadly AMD has not released any 7xxx info basically. Well except the 7xxx mobility, showing proposed die-shrink, but that doesn't tell you much about the full scale versions. if they are tweaked 5/6xxx that will likely not be a good sign looking at the 6xxx vs 5xxx in terms of $/MHash (though good in $/Watt).

Since these cards ARE primarily gamer cards, expect a lot of focus on non-mining features like tessellation and DOF shader effects and other things that take up PCB space and don't contribute to hashing.

It seems like someone could write new hashing software that will utilize what a certain card is good at. Perhaps not though.
Would be kinda cool if AMD would release a special mining card; would be certain to sell out.

Rograz, why are they blocking SHA-2?

New slides show they are going into the same direction as nvidia. This means support for c++ and most likely less shader, but certainly not more. Recent mobos all come with 3 slots, so i guess we are suppose to use them  ::)

Yep, they are going for more general purpose computing. In the presentation they even stated that one of the few applications that thrives on their current architecture is hashing.. that's not to say the 7-series will be bad for hashing, but they will certainly not be as optimal for it as the current generation.

7xxx will perform similar to 58xx and 6xxx in terms of hashing, it'll be terrible for mining due to its "newness" and price will be pretty high. 58xx is still the best deal if you could find any at a good price point.

Without getting too technical, that architecture (SIMD) will be less efficient for mining. The question is whether we will see it in the 7xxx series, and while unconfirmed - my guess is yes. They will probably reuse their current strategy - low end/mainstream parts using tried and true architecture (VLIW4 in the current 69xx) for the <79xx parts, while going to the SIMD (which is worse for mining) on the 79xx parts. The 4+1 architecture found in the 5xxx series is the most efficient mining architecture we will ever see, period. But, keeping in mind that VLIW4, in the current high end 6 series may not be as efficient, its still pretty good, combine this with a very large die shrink (40nm --> 28nm) we will still have some very efficient cards for mining, more efficient at cost and watt per hash rate than anything out today. Density may be the only real problem if my guess is correct, as we wont be seeing the dual gpu cards with the mining efficient architecture for the next generation.   

I don't know where you guys are getting this info from that the 5000 series give better hash rates as they don't. The equivalent 6000 series cards gives a couple of more Mh/s.

As for the 7000s, if it's just a die shrink, we'll see a major increase in hash rates if the GPU stays the same size, as there would be over a billion more transistors. A 28nm 6870 with a GPU size of a 40nm 6870 would offer similar performance to a 6970. On the other hand, the size of the GPU could be reduced, keeping the same number of transistors as before. This would provide a decrease in power consumption.

With smaller transistors, you can pack in more SPs per mm^2, and lots of SPs is what makes these cards good for mining. If the 7000 series is simply a die shrink, then we'll see faster hash rates, decreased power consumption, or both. Just look at the difference between the 55nm 4000 series and the 40nm 5000/6000 series.

Define "equivalent"? 58xx vs 68xx? 5870 gets 430-450MHash/sec, 6870 gets 300-340MHash/sec. i don't see that as the 6xxx gettng more mhash/sec? if you are suggesting equvalence of price it still doesn't work out, a 5870 gets 430-450 again, and a 6950 2GB unlocked shaders (actually more expensive usually, but anyway) gets about 385 - 410 if you're lucky. Those numbers are both from personal experience and backed by https://en.bitcoin.it/wiki/Mining_hardware_comparison

i'd be interested in what data you're using to suggest that 5xxx series don't hash better.

Anyway, just because die sizes have been shrunk doesn't mean they're going to be throwing on transistors and streamprocessors out willy nilly. Depends on the gaming focus AMD chooses to take. in theory the vliw4 config allows more SP / graphic core because it takes less die space, but we saw a reduction of SPs in the 68xx series (clocked higher).

Again, we don't really know jack, but it is unlikely the architecture will be superior to current 5xxx cards.

In terms of efficiency they are, 6970 is larger and uses more power than a 5870 and still the 5870 can outperform it, show me a 6970 pulling 445MH/s at 1ghz core like the 5870 does.

The 6870 is not equivalent to the 5870 though, it's the 5830s replacement. If you look at the specs of those 2 GPUs side by side, you'll see that the only difference is the increased clocks and twice as many Z/Stencil ROP Units and Color ROP Units. The 6950 replaces the 5850, the 6970 replaces the 5870 and the 6990 replaces the 5970. The majority of people make this mistake at first. The 6000 series don't use the same numbering scheme. They don't even use a consistent numbering scheme.



The mining hardware comparison chart. The Caymans (6950 and above) use a different architecture though so the same rule doesn't apply to them.


Like I said, they can either reduce power consumption, increase performance or both. The variable is the GPU size.


The 6800s dont use VLIW4 they use VLIW5 the same as the 5000 series, it's only the 6900s that are using VLIW4, and they don't have less SPs. The 6870 has the same number of SPs as the 5830, which is 1120. A 28nm VLIW4 GPU would destroy the equivalent 40nm VLIW5 GPU in terms of hash rate.

Give me a 6970 and I'll show you one.

Have a look at the mining hardware comparison site. There's a 6950 @ 1GHz with 432 Mh/s. The best value listed for a 6970 is 423 Mh/s. You can tell instantly that those cards haven't been optimised because if they had, they would be getting better results than the 6950. Now look at all the other cards, starting with the 6870 vs the 5830. The 6000s get a couple of Mh/s more, which comes as no surprise to me since they're just tweaked versions of the 5000. Some of them aren't even tweaked, just rebranded.

My stock sapphire 6970 get 420 MH/s, I haven't OCed them since they are call crammed together, I'm still waiting on riser cables from HongKong.

my xfx 6970 blows chunks

This is why i asked you to define equivalent, if you arbitrarily define equivalency by no metric but what "seems" like it should replace what i guess that makes sense. Except that the price, heat, power consumption and so on are not at all equivalent. A 6970 costs far more than a 5870, consumes more power, is often bigger and bulkier, and i suppose you are technically correct, that for those tradeoffs if you can manage to overclock it farther than a 5870 you can get a few more MHash potentially (non-equiv clocks). That's not a very good measure of equivalence though.

It's not my definition, it's AMD's. That's how they marketed the cards. Changes nothing though. A 6870 is still just a 5830 with twice as many stencil ROPs and colour ROPs and a 6770 is just a rebadged 5770. When comparing the 5000s to 6000s, count the SPs.

The 5770 and 6770 have 800 SPs (same card, different sticker).
The 6850 has 960 SPs and hasn't got an equivalent 5xxx card.
The 5830 and 6870 have 1120 SPs.
The 5850 has 1440 SPs, the 6950 has 1408 SPs.
The 5870 has 1600 SPs, the 6970 has 1536 SPs.

That's how you compare these video cards, not by using some arbitrary naming sheme but by their technical characteristics.

The particulars on AMD's next gen GPU are still unknown, so, any opinion on whether it'll be good at mining is just speculation.

http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute (http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute)

There are a lot of info here. Though I guess it's possible that this is referring to the generation after the 7000.

6970 and 5870 isn't really a fair comparison at all. As was said, those 6970's on the chart aren't even optimized. Not to mentioned drivers in general have yet to be optimized for the entire 6 series. Here's the thing, the 6970 has no direct equivalent. It's an odd ball. I mean, people can say it was made to replace the 5870, but the 6950 whips the 5870 in pretty much every game. Let's not forget these are all gaming cards, not mining cards. I have a feeling the 6970 is really a monster waiting to be unleashed with the right drivers.

Ive not tested any 6xxx cards myself but even if it possible that you can get 6970 to be the best single card in terms of pure MH/s you will never beat 5870 in terms of MH/s per watt, and let's not forget the fact that cayman is larger than cypress. Your comparison with 6950 vs 6970 isnt really valid either since that 6950 is more than likely unlocked so in essence no different from a 6970.

The only rebranded parts in the desktop area are the 6750 and 6770 and how can you compare a 5830 that is a horrible version of the cypress core to a full blown barts (6870).

sounds they are firing up the 58xx layout again to profit from the mining rush...

Anyone seen specs of the 7xxx? Other than those slides from the Fusion summit that officially have nothing to do with 7xxx?


This does not account for the stock clock speeds. 5830 and 6870 have 1120 SPs, but 5830 runs at 800MHz while 6870 runs at 900MHz by default (architecture being almost the same).

Comparing 69xx is more difficult, it's VLIW4 and you get near 100% ALUPacking unlike with 5xxx. That can give you up to 5-10% performance advantage on the same code. In fact, 6970s tend to be slightly faster as compared to 5870s despite the reduced SP count (almost compensated by the higher clock speeds though). The reason why people often report better speeds with 5870s as compared to 6970s lies basically on unoptimized kernels. And of course, 6970 is much less power-efficient.

That's an unlocked 6950 (equivalent to a 6970), next time read the comments listed next to it.

Of course it is, otherwise the hash rate would be far lower. The point is, that the highest 6950 listed has a higher hash rate than the highest 6970 listed. The 6970 has a higher core clock and the unlocked SPs on the 6950 don't scale with clock speed. So, either those scores for the 6950s are badly exaggerated or the results for the 6970s aren't optimised.

So back to the original question:

1) SHA-2 capability has been removed from the upcoming 7xxx series, thus, won't help miners

2) The 7xxx specs hasn't been specified yet so we do NOT know if it will outperform the 6990 yet?

Can someone please verify that the AMD HD7000 Cards won't have the Int32 Right Rotate Operation (very useful for SHA-256).

I can't find it in any rumor about the upcoming HD7000 series.



http://www.hardwareboard.eu/topnews/geruchte-zur-radeon-hd-7000-serie-kommt-definitiv-noch-2011-in-28nm-update-8-release-monat-bekannt-379/ (http://www.hardwareboard.eu/topnews/geruchte-zur-radeon-hd-7000-serie-kommt-definitiv-noch-2011-in-28nm-update-8-release-monat-bekannt-379/)

This homepage (german) says, that the upcoming HD7xxx chip will have 3200 Shader Cores and about 6.08 TFlop/s. But unless AMD releases official informations, we should be careful with this informations.


https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU (https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU)
And finally in this wiki there is an explanation why ATI is better for mining than Nvidia. If the HD7000 really has 6,08 TFlops and does support INT32 Right Rotate, we will see about 750 MHash/s.

Why do i make this prediction?
Simply a not overclocked HD5850 has 250 MHash/s with about 2 TFlop/s. A GTX480 reaches 105 MHash/s with 1.4 TFlop/s. The INT32 RR Operation gives ATI an advantage of 1.9 for SHA-256 hashing. Together with the higher Peak Performance of the HD5850, the 5850 should be about 2.6x faster than a GTX480, which is nearly true.

I know the HD7000 cards will use a newer architecture similar to NVIDIA, but Mining performance seems to depend linear to the peak performance of a video card. The reason why ati cards are more efficient seems to be just the INT32 Right Rotate Operation, which can be performed in one operation on ATI cards, but needs 3 operations on Nvidia cards. So it's very important to know wether the HD7000 support this operation in order to predict the Hashrate.

Here:
http://img19.imageshack.us/img19/38/watxb.jpg

I have it at only 950core/1375Memory doing an average of 460+MH/s (values fluctuate between 457.8, 460.5, 468.7). Still, a 5870 is far more efficient at mining since the GPU is smaller by half a billion transistors, it has half the ram clocked at a lower speed and at lower voltage plus the PCB is less complex (less layers).

how you geting that many Mh/s out your card? my unlocked 6950 at 975 core only get 430ish

WTF.

What SDK? Can you post all the flags you're using? That's insane.

Besides using Vista, I've never been able to get that on my 6970s let alone on that clock. I have to push it to 1ghz to get even close and I'm still 20 mh/s short. Getting 460 with aggression at 7?

I'm extremely confused. I've tried a zillion combos on win7 and linux I don't understand how you get that speed. Could Vista really boost it by that much?

And you won't find confirmation anywhere because it has not been removed, v_alignbit_b32 is still there.
Here is a partial dump of the Southern Island's instruction set: http://forum.beyond3d.com/showpost.php?p=1561886&postcount=365

Under vector instructions, there is v_alignbit_b32

The specs posted on that site are pure speculation. No one knows the actual specs of the new cards, but we do know the underlying architecture and we know (via the instruction set dump from the drivers) that AMD's GCN (Graphics Core Next) does support the INT32 Right Rotate Operation.

No, Vista is largely irrelevant. I just put that in there for full disclosure. My performance is achieved through the various kernel tweaks posted on the forums. These tweaks appear to have a greater benefit on the VLIW4 architecture of the HD69xx (Cayman) cards than the VLIW5 architecture of the HD5K (Cypress) family.

Considering that the 7xxx was due out Q3 2011, I'm kinda surprised the official specs haven't been released yet.

Specs are usually released at 6 months ahead of launch schedule.

Does this mean that they will be delayed? We are already in Q2.


I heard a rumor from a wholesale supplier about the shortage of 6990 cards. He said something to the effect of:
"In general, the manufacturers don't produce too many high end video cards like the 6990 because they aren't for the mass market but only for high-end gamers. Combine this with the earthquake in Fukushima, and there is a shortage of high-end transistors that can ONLY be found from the Japanese companies"

If so, I'm only guessing that:
1) natural shortage of high-end cards to begin with
2) shortage of transistors due to Fukushima
3) bit mining demand surge


If all factors of the above are true, it makes sense why the 6990 card is a bitch to find.

If we combine this factor with the possibility that the 7xxx card may not even include SHA-2 (still unconfirmed) and thus, not ideal for mining, we could be facing a HUGE short-squeeze in 6990 cards for at least another year.

Any comments to the above?

Capacitors. :-)

Is there any way to turn this junk off in RBE to save energy?

It's pointless for AMD to remove it ;) Unless you are paying them to change the architecture of the new serie :P