Author

Topic: NEW ATI 7000 architecture info (Read 2352 times)

legendary
Activity: 1022
Merit: 1000
Freelance videographer
September 09, 2011, 04:42:27 AM
#18
Sounds like I was misled by an inaccurate article then.I wondered how is it possible to have 6400 shaders up from 3072? At first I thought it was true,then after reading the posts here,this may not be the article I thought it was.

Looks like I shouldn't expect a hashrate that's double the HD6990 that we have now.I wonder if it's possible to have more than 2 GPUs on the same board (yes I know power use will be very high and out of spec in this arrangement,but I ask as I wish to have as much power as possible from a GPU as CPUs don't give me enough power for mining/GPGPU accelerated filmmaking and editing)?

Either way I do look forward to the improvements.I wonder how much a shader unclocked Hd6950 would be worth so that I can upgrade to the new cards if needed and sell off the old one (on the new cards release)?

hero member
Activity: 602
Merit: 500
September 09, 2011, 03:57:50 AM
#17
That videocardz.com piece (of crap) is the worst article ever. Full of errors.

- they quote 12.16 SP TFLOPS which would be 138% more than the 6990, but they say 58% in the text (and both numbers are wrong anyway, see below)
- the 7990 is more likely to be around 8.2 SP TFLOPS with 4096 ALUs @ ~1000MHz, which would be ~61% faster than the 6990
- they write 6400 stream processors in the table, but 3200 in the text (and both are wrong, see below)
- the 7990 is more likely to have 2*2048 ALUs, certainly not 6400 (which would imply a VLIW5 5970-like manufactured in 28nm with a doubling of the # of ALUs)
- typo: s/Steam processors/Stream processors/
- etc

You could have simplified your post by simply pointing out they use a "z" in place of an "s", which invalidates all their writings automatically.

EDIT: To be fair the numbers they are quoting come from: http://news.ati-forum.de/index.php/news/34-amdati-grafikkarten/1815-technische-daten-der-amd-radeon-hd-7990-bekannt

now, ati-forum is not necessarily official afaik, but I imagine they try to put out reasonably good information.
mrb
legendary
Activity: 1512
Merit: 1028
September 09, 2011, 03:48:02 AM
#16
That videocardz.com piece (of crap) is the worst article ever. Full of errors.

- they quote 12.16 SP TFLOPS which would be 138% more than the 6990, but they say 58% in the text (and both numbers are wrong anyway, see below)
- the 7990 is more likely to be around 8.2 SP TFLOPS with 4096 ALUs @ ~1000MHz, which would be ~61% faster than the 6990
- they write 6400 stream processors in the table, but 3200 in the text (and both are wrong, see below)
- the 7990 is more likely to have 2*2048 ALUs, certainly not 6400 (which would imply a VLIW5 5970-like manufactured in 28nm with a doubling of the # of ALUs)
- typo: s/Steam processors/Stream processors/
- etc
legendary
Activity: 1022
Merit: 1000
Freelance videographer
September 08, 2011, 06:23:23 AM
#15
Check this out:http://videocardz.com/28708/AMD-RADEON-hd-7990-how-powerful-can-it-be

New info about the cards.I wonder what MHash/s I can get from that card? I was blown away after seeing those specs :-)
sr. member
Activity: 462
Merit: 250
It's all about the game, and how you play it
September 07, 2011, 08:33:14 AM
#14
Has anyone stoped to consider that with the focus of GCN being on compute power we may see an arcitecture closer to nvidias that is actually worse than the vliw-5 and vliw-4 arcitectures currently available for bitcoin mining?
Have you considered reading the posts above you?

Poor form on my part for reading a thread and coming back to reply later without checking it's updates but thank you for your pretentious answer, more serious thanks for the comparison using NV shader performance comparison adjusting to the number of shaders listed in the new information sleeper unit.
legendary
Activity: 1862
Merit: 1011
Reverse engineer from time to time
September 07, 2011, 07:06:13 AM
#13
Has anyone stoped to consider that with the focus of GCN being on compute power we may see an arcitecture closer to nvidias that is actually worse than the vliw-5 and vliw-4 arcitectures currently available for bitcoin mining?
Have you considered reading the posts above you?
legendary
Activity: 1022
Merit: 1000
Freelance videographer
September 07, 2011, 06:53:26 AM
#12
I like what I'm seeing with those HD79xx cards.Lower energy use and possibly higher performance. My HD6950 card uses ~220Ws at full load (mining) due to shader unlock (which is unnecessary with those cards as they have the right no of shaders to begin with afaics) as well as slightly overvolted (by 0.005V increase) and heavily OC'd (902MHz core clock,320MHZ mem clock for stability and performance).These changes give me ~405MHash/s. I wish I could OC my MBPs GPU and underclock it's CPU to reduce power use and heat issues.Is there a tool that works for OC'ing a MBP GPU?

Thanks
newbie
Activity: 52
Merit: 0
September 06, 2011, 07:42:08 PM
#11
Has anyone stoped to consider that with the focus of GCN being on compute power we may see an arcitecture closer to nvidias that is actually worse than the vliw-5 and vliw-4 arcitectures currently available for bitcoin mining?

  That is what wndrbr3d posted above as being the uncertain aspect still of these cards. It is still quite possible that they do away with that magical offset that makes them such great miners. Anyone able to summize what the worst case scenario is if they go to a more Nvidia like arch but are adding in the much higher SP and lower watt req? What speeds would they manage in the worst case scenario?

I think needing as many shader cycles per hash as my GTX285 did is probably a fair worst case scenario.
GTX285:  240 SP @ 1476 MHz = 60 Mh/s
Rumoured 7970: 2048 SP @ 1 Ghz = ~348 MH/s

Lets hope that the new shader design is better then this. But if they aren't, at least midrange cards like the 78XXs are usually plentiful and reasonably inexpensive.
hero member
Activity: 504
Merit: 500
September 06, 2011, 06:45:48 PM
#10
Has anyone stoped to consider that with the focus of GCN being on compute power we may see an arcitecture closer to nvidias that is actually worse than the vliw-5 and vliw-4 arcitectures currently available for bitcoin mining?

  That is what wndrbr3d posted above as being the uncertain aspect still of these cards. It is still quite possible that they do away with that magical offset that makes them such great miners. Anyone able to summize what the worst case scenario is if they go to a more Nvidia like arch but are adding in the much higher SP and lower watt req? What speeds would they manage in the worst case scenario?
sr. member
Activity: 462
Merit: 250
It's all about the game, and how you play it
September 06, 2011, 06:03:03 PM
#9
Has anyone stoped to consider that with the focus of GCN being on compute power we may see an arcitecture closer to nvidias that is actually worse than the vliw-5 and vliw-4 arcitectures currently available for bitcoin mining?
full member
Activity: 161
Merit: 100
September 06, 2011, 03:09:00 PM
#8
A bit more cores and speed?

The 7970 has 2048 SP running @ 1Ghz.  that is 2.048 trillion ALU OP/s.  Current miner code requires ~3670 OPs/hash.
2.048 trillion ALU OPs / 3670 = 544 MHashes/s roughly 50% more than 6970 while using less power.

Your comparison only works if the 79xx series was also VLIW4, but it's not. We don't have any numbers for comparison for MIMD, and for all we know, it might be no faster than G80+ based nVidia cards. I say this because nVidia has based their cards on MIMD/SIMD/SIMT architecture since the G80. So although the move to MIMD could make ATI cards perform better in gaming... it could also hurt mining performance.

We just don't know yet.

Yeh but still, if this rumors are true new 7970 have 2048 SP compare to nvidia 512 SP in 580GTX just question is how you will able access this 2048 SP....
full member
Activity: 182
Merit: 100
September 06, 2011, 12:54:11 PM
#7
Well it will at least reduce the cost of the existing 6000 series stock, 99$ 6870 might be the new 5830
Its not that it might be it will be. They are the same card basically once you even out that the 5830 gives a little more megahash and the 6870 requires less power. Only thing that makes a 5830 better is the price atm. The two you want to keep your eye on will be the 6870 and the 6950. If the price drops even 20% you will see 6870's at 120 and 6950's at 160.
hero member
Activity: 711
Merit: 500
September 06, 2011, 12:43:48 PM
#6
Well it will at least reduce the cost of the existing 6000 series stock, 99$ 6870 might be the new 5830
hero member
Activity: 914
Merit: 500
September 06, 2011, 12:33:24 PM
#5
A bit more cores and speed?

The 7970 has 2048 SP running @ 1Ghz.  that is 2.048 trillion ALU OP/s.  Current miner code requires ~3670 OPs/hash.
2.048 trillion ALU OPs / 3670 = 544 MHashes/s roughly 50% more than 6970 while using less power.

Your comparison only works if the 79xx series was also VLIW4, but it's not. We don't have any numbers for comparison for MIMD, and for all we know, it might be no faster than G80+ based nVidia cards. I say this because nVidia has based their cards on MIMD/SIMD/SIMT architecture since the G80. So although the move to MIMD could make ATI cards perform better in gaming... it could also hurt mining performance.

We just don't know yet.
full member
Activity: 161
Merit: 100
September 06, 2011, 11:12:04 AM
#4
They decided to keep with VLIW vs. moving to MIMD (which was only rumored)... so performance can probably be a linear increase from the 69xx series to 78xx series.

I think the 78xx series focus appears to be gaming, as with the previous reply, the core counts and core frequencies look about the same with only the jump in memory bandwidth being the major change. This will make memory intensive tasks (like anti-aliasing at higher resolutions) much faster.

For mining, we're going to have to wait for the 79xx series with the MIMD architecture to see any radical change in the positive or the negative. There, the top end card will have 500 more stream processors and a slightly higher clock than the 6970.
There is, in this article, rumored 7970 GCN architecture with 2048 cores at 1000Mhz... + 2 slide with new architecture description.
hero member
Activity: 914
Merit: 500
September 06, 2011, 11:07:33 AM
#3
They decided to keep with VLIW vs. moving to MIMD (which was only rumored)... so performance can probably be a linear increase from the 69xx series to 78xx series.

I think the 78xx series focus appears to be gaming, as with the previous reply, the core counts and core frequencies look about the same with only the jump in memory bandwidth being the major change. This will make memory intensive tasks (like anti-aliasing at higher resolutions) much faster.

For mining, we're going to have to wait for the 79xx series with the MIMD architecture to see any radical change in the positive or the negative. There, the top end card will have 500 more stream processors and a slightly higher clock than the 6970.
full member
Activity: 181
Merit: 100
September 06, 2011, 11:00:23 AM
#2
Well, I took a glance and immediately saw same clock speed and same cores as the current high end at that 1500 range. The upside to the new cards are a WAY higher memory clock and lower power consumption.

With mining, memory doesn't play almost any part, so that'd be wasteful, I'd see someone downclocking that memory down to a few hundred mhz. On the other hand, for same hash performance you'd cost 1/5 the power (more savings with lower memory!?).

So I'd say you would change 5x MHash/Watt for probably 5x MHash/$. If the cards come out to $700 like the 6990, there'll be a lot more profit on the coining due to the lower power underlay (and I imagine less heat because of less watt burning?)

Edit: Woops, didn't see the second chart below, that line has a bit more cores and higher clock speed, don't imagine an insane jump in hashes though, and that line will definately be more expensive
full member
Activity: 161
Merit: 100
Jump to: