.... I expected more of a smack down. amd is apparently more focused on beating nvidia at the price game, hence a better card same price though initial offerings suggest big price jumps already, and not producing the king of the single core cards. many, as did I, expected the 6970 to top the 580, so either nvidia has pulled a coup or amd's focus is not interested in top single core gpu. it is a great card for the money.

thanks again for the info.

AMD are more focused on market segment, and producing cards for that segment. From their perspective - a card is a card, and its the performance of the card that matters, no matter the number of GPU(s) on it. The 6990 when released circa Jan 2011 will be the top performing card, and then the whole AMD range will make sense.

At present NVIDIA have a big design problem, they were banking on shrinking the design for Fermi but could not due to TSMC dumping its plans for a smaller die shrink. That has left NVIDIA (like AMD) still on 40nm, and having to resort to rebranding for "new" cards. The design of the NVIDIA GPU is massive, and very unlikely to ever fit 2x580 on a single card. That could change when TSMC's production facility ramps up a smaller die shrink in 2011. However, AMDs design is fundamentally smaller already hence no dramas using 40nm, so NVIDIA will have its work cut out.

The trend to a more focused design is similar to the CPU story going to names rather than focusing on pure clock numbers. It will become even more pronounced in 2011 when the first CPU/GPU integration hits the market, with Intel joining in the fun alongside AMD with its own offering. NVIDIA dont yet have any likelyhood of a combined CPU/GPU aka the AMD FPU, so 2011 is going to be a defining year for many reasons.

By the end of 2011, the way we think about "GPU" performance will change dramatically, and unless NVIDIA have something to pull out of the hat, they are likely to be left for dead. As its unlikely NVIDIA will allow the latter to literally happen ...... its going to be interesting to see what they do. At the moment NVIDIA are way behind the curve in terms of end outcome 2011.

Regards
Zy

The 6990 when released circa Jan 2011 will be the top performing card, and then the whole AMD range will make sense.

zydor, I agree with some of the things you suggest especially long term favoring amd, but not so sure about amd's strategy with the 6970.

which price point to they think they take with the 6970? people dropping ~400 will drop ~500, but $700-800+ is for a much smaller crowd. therefore, the 580 and the 6970 are close enough in price to suggest similar price point and that leaves the 6970 outgunned. the 6990 price will most likely be way out there, heck, the 5970 still goes for 600-700+.

for the range $400-500, right now, nvidia with the gtx580 has done a number on amd. amd wants the price point, but if I am spending $400-500 I want the better card and the 6970 is not it.

I consider the 6970 next to the 580 a grand disappointment but I guess amd know what they are doing and how the 6990 will change our view of things. but if I had to buy a card today as many do for the holidays, it would be the 580 and it's out of stock at several places checked. part of that is timing, had a head start, but surely many are looking at the 580 vs 6970 and picking the better performer as noted in every review I have read.

the 6990 will most likely be the king down the road, so I wait to see if the 6970 take the sales numbers vs. the 580.

5970 @850-1150
WU of type de_separation_17_3s_fix_1* : 130 s per gpu
WU of type de_separation_16_3s_fix_1* : 88 s per gpu
Temp (with Accelero) : GPU = 66, VRM = 75

6970 Crossfire @ 900-1400
WU of type de_separation_17_3s_fix_1* : 101 s per gpu
WU of type de_separation_16_3s_fix_1* : 67 s per gpu
Temp : GPU = 90, VRM = ??

5970 @850-1150
WU of type de_separation_17_3s_fix_1* : 130 s per gpu
WU of type de_separation_16_3s_fix_1* : 88 s per gpu
Temp (with Accelero) : GPU = 66, VRM = 75

6970 Crossfire @ 900-1400
WU of type de_separation_17_3s_fix_1* : 101 s per gpu
WU of type de_separation_16_3s_fix_1* : 67 s per gpu
Temp : GPU = 90, VRM = ??

I think two 5870s also beat a 5970, but the 5970 is still kick*ss.

gpu 90c is a tad hot? actually, that sounds familiar: I remember prior to installing aftermarket coolers (5870) my inference was that the reference fan was blowing the vrm heat across the core. if the 6970 layout similar to the 5870 then your vrms are probably cooler, closer to the fan, than the core gpu temp. 90c is tolerable as long as the vrms are not smoking.

recommend you not oc, if plans therein, until you work better cooling or turn up the fan or something. it will be nice to see 6970 oc times, as stock is 8 secs faster than an oc'd 5870, 17s WUs.

good info, thanks.

..... for the range $400-500, right now, nvidia with the gtx580 has done a number on amd. amd wants the price point, but if I am spending $400-500 I want the better card and the 6970 is not it..... but surely many are looking at the 580 vs 6970 and picking the better performer as noted in every review I have read.

the 6990 will most likely be the king down the road, so I wait to see if the 6970 take the sales numbers vs. the 580.

For sure, right now, if anyone is buying then for raw power the 580 wins, no question. Power wise it will change in January there is no doubt, the 580 only just tops the 5970 (circa +5 to 10%), it will be floored by the 6990 there is no doubt. Then again NVIDIA had little choice but to go for the fastest it could roll out as there is no chance of doubling up on the same card.

The 580 is deliberately pitched between the 6970 & 6990, and is the card that should have come out on Fermi release. Not much else NVIDIA could do. I think they have done enough to maintain share at the top end, but I doubt they will increase it. I think at the price points for the high end cards people will tend to stay loyal to past useage, and not much will change. For new users, the picture I suspect will be different as in effect they have a choice of two AMD cards against one NVIDIA. It will all depend on price - AMD have it in their gift to drop price to a level NVIDIA have no hope of matching as the AMD design is way more efficient to produce.

Could be interesting, and the latter is a good thing, the last thing we need is a clear outright winner at all levels - that would kill off NVIDIA in the consumer market, and we need them in the market else prices will be maintained artificialy high. The real test will be the new upcoming combined GPU/CPU chips in 2011. I hope NVIDIA have something in the back pocket to compete and revive their fortunes, else we will all suffer long term.

Regards
Zy

[quote] ... gpu 90c is a tad hot? actually, that sounds familiar: I remember prior to installing aftermarket coolers (5870) my inference was that the reference fan was blowing the vrm heat across the core. if the 6970 layout similar to the 5870 then your vrms are probably cooler, closer to the fan, than the core gpu temp. 90c is tolerable as long as the vrms are not smoking.

good info, thanks.

The point you raise re VRMs is a good one to watch carefully for. The 5970 has an inherent design fault in that the VRMs are poorly positioned on the card. When o/c a 5970, great care needs taking with the VRMs, its not good enough to just watch the core temp and assume all is well.

A 5970 running at circa 90 degrees core, will be nudging the VRM limits of 120 degrees, usually tipping over causing a downvolting, depending a great deal on cooling solution. When the 6990 comes out, there will be much scrutiny of the VRM positioning on the card to prevent a similar o/c bottleneck. I doubt AMD will get that wrong, its a well known old problem of 5970s that they are unlikely to repeat.

But its a funny 'ol world :) You never know, and bares watching for before buying any 6990 - the VRM position has been the "gotcha" of many-a-5970 burning out, as the user only watched the Core temp and paid little attention to the VRM temp & limits.

Regards
Zy

sure, i don't use the 6970s for milkyway for now
i'll wait till gpu-z can read vrm temp
for the accelero on the 5970, i put some thermal paste on the vrm otherwise temp would raise till 110

..... for the accelero on the 5970, i put some thermal paste on the vrm otherwise temp would raise till 110

Interesting - so thermal paste on VRMs, as well as on the GPUs, will work with the acceleros? Might get my screwdriver out this weekend if thats the case - at present only have it on the GPUs with my two.

Regards
Zy

thermal paste like tuniq2 because there is a lot of space between the vrm and the heatsink

Useful - I'll give that whirl this weekend

Regards
Zy

be careful and verify that there is a contact with the heatsink otherwise you can fry your card
i put for each à drop of paste even on the memory chip because the thermal pad with the accelero are very bad

I understand that some prefer to run with a higher memory speed in case MilkyWay runs out of work and BOINC switches to Collatz. However I would never consider running my HD 5970 with a memory speed of 1150 on MilkyWay or DNETC. The first lot of VRMs get hot enough at 500 MHz in an Australian summer. At 1150 MHz memory and a fan speed compromise that is not deafening I think they would melt towards the centre of the earth.

I'm not having a go here as obviously the Accelero plus extra thermal paste is coping well, but I would be interested in the VRM temps of anyone running a 5970 Accelero combo on MilkyWay with memory speed of 500 MHz. I have one but have never installed it because I read of some who had problems, plus I saw a review that showed higher VRM temps than standard cooler even though the core temps were excellent. Maybe I should have a go at installing it.

The relative selling prices for different models currently vary greatly from country to country so any value conclusion can only be country specific. Many who don't have a NewEgg or similar low priced stores in their country often have different opinions to users and reviewers in America. So I'm not going to comment on relative performance/value of different models/brands for gaming as it not an area I have any experience in. Other than to say that sometimes things are not as they seem and pricing/market positioning and stock availability is done to maximise profit by selling a greater number of mid range cards that are cheaper to manufacture and hence higher profit. Although it has promotional value, it is not only about having the best performing flagship models that are the darling of reviewers and beloved by those who can afford one or more of the most expensive models.

In relation to crunching, it would appear the extra memory and larger die of the Cayman models causes an increase in power consumption which balances performance increases for those whose GPU crunching is constrained by ever rising power bills. Although the extra performance is welcome the full efficiency benefits of the newer architecture will not be realised until the die shrink. A 5970 that is not overvolted is still very hard to beat for efficiency of crunch per watt compared to multiple single core GPUs. Although the crunching performance of the 6970 is good the high temperature of the larger core may concern those who live in a hot climate.

kashi, very nice point considering how the international market might change the position of the cards, and I am certainly not looking in england or germany or australia or wherever to see how they compare for my position. zydor also makes great points. thanks.

not to belabor this too much but my point is a simple one: amd strove for a price point and for the money, at least in the us, have created a great card--but not the best single core gpu. why not leave that mid-market competition with the 6950s and especially 68xx? use the 6970 to go for the king of the single core competition? by lumping their 6970 in with all their other price point grabbers they have created a very good but not as good as gtx580 and with a price that is close enough to the 580 they are left head2head with a superior card. amd's and nvidia's best single core processor head2head. mistake? imo, yes, for amd apparently had no interest in that competition or the 6970 would better compete with the 580 functionally.

it is understood the 6990 will outperform everything to date but its assumed price separates it from this discussion. in short, what they wanted with the 6970 is jeopardized: go for less punch at a lessor price and take that price point.

nvidia might be desperate, but that card is disappearing in our markets just like 5870 did.

I am a fan of function and satisfied with amd--for now, but annoyed with what they decided to do with the 6970.

for the record, I am not dissing the 6970 functionally; it appears to be a great card, but time will tell if this great card gets amd what they seek. I expected more from amd's top single core card, and that I must taper. unfortunately.

again, thanks for the input.

I understand what you are saying. My personal view as someone who lives in a hot country is that the heat and possible noise issues are of greater concern than performance issues. All I was meaning is that the reasoning behind market positioning and pricing is not always what it may seem and may differ from what is stated by some commentators. Having a range of flagship series cards that perform well enough for promotional reasons does not mean that it is an advantage to sell a lot of them in preference to selling more of the higher profit series that use a simpler architecture, smaller die and less costly cooler.

We saw this with the release of the Cypress cards where a relatively small number were available initially and as soon as the more profitable Juniper series became available then stock of Cypress models became so scarce that it caused a rise in price. This shortage persisted for a long time ensuring a large number of more profitable Juniper cards were sold.

Of course this may relate more to the inability of their fabrication plant to switch to the smaller size process. Kind of like a smaller dose of the problems NVIDIA had with the GTX 480. The 4 VLIW architecture was planned for the smaller process so it is not really surprising that a large complicated die does not perform as well as some expected and wished for. AMD may have indeed been aiming to have the fastest single core card but was unable to achieve it because of these problems. Manufacturing something with 2.64 billion transistors ain't just whistling Dixie.

The Cayman double precision change from one fifth to one quarter of single precision is welcome and augurs well for the next generation which should address heat issues. Need more examples to confirm but adjusted to default clockspeed, the figures shown in this thread indicate that the increase in double precision performance in MilkyWay from 5870 to 6970 exceeds that quoted by AMD by a small amount so there may be some other improvements at work there too. HD 5870 = 544 GFLOPS, HD 6970 = 675 GFLOPS, so a 24% increase at default clock is specified, but MilkyWay performance has increased by 26-27%.

I previously commented that a 6970 would use the same or greater amount of power than a 5870 to perform the same amount of work. I based this on AMD power figures. I have since read a review which showed 6970 using slightly less power than 5870 when being benchmarked with FurMark. The 6970 consumption was similar to AMD figures but the 5870 consumption was way higher than the 188 watts quoted. If this translated to MilkyWay it would mean that 6970 is about 30% more efficient than 5870 when being used for crunching here and that is excellent. Of little interest to gamers, performance per watt is becoming of increasing concern where I live. The power retailers have just been sold off and tariffs are likely to rise even faster than in the past. I have already removed my 5870 from my box, my CPU is undervolted and the length of time I can keep crunching with the 5970 is shrinking fast. No regrets, I've enjoyed a good run and hopefully the work done was worthwhile.

I will be interested in how the 6990 tackles the 300W power restriction and the methods used to keep 2 large dies sufficiently cool. Wouldn't be surprised if 6990 introduction is delayed further or if stock is restricted when/if it is released.

I've got 6970.

Stock clocks (880) - 69sec per 213 credits WU and 57-58 secs - max OC'd (950). Easy takes 950 @+20% boost. Now I'm stability on 950, looks fine, for sure can tell 2morrow though. Hot thus noisy. I expected smth else...

I've got 6970.

Stock clocks (880) - 69sec per 213 credits WU and 57-58 secs - max OC'd (950). Easy takes 950 @+20% boost. Now I'm stability on 950, looks fine, for sure can tell 2morrow though. Hot thus noisy. I expected smth else...

WOW!

Temps and fan percentage? What's running on the CPU?

WOW!

Temps and fan percentage? What's running on the CPU?

This rig living in my cool basement (somewhat +18*), so I do not really care about noise. Original fan curve IMO is too low what makes GPU too hot. That's not good for 24/7 crunching, so I increased fan speed in MSI AB. Now on 76% fan speed it's 59-60 *C

I tried to make 4870 work too, but it does not starts (can not see "welcome" screen). 10.12 driver and 10.12a hotfix for 69xx been installed.

I will be interested in how the 6990 tackles the 300W power restriction and the methods used to keep 2 large dies sufficiently cool. Wouldn't be surprised if 6990 introduction is delayed further or if stock is restricted when/if it is released.

kashi, great input; you make wonderful observations, and I look forward to the 6990 specs as well. unfortunately, if the hearsay is to be believed, expect the 6990 q1 to possibly even q2, 2011. yeeepeee. :|

thanks again for the input. good stuff.

oh yeah, it is a pity your electrical concerns; I think those are shared by a growing number. sad. for your sentiment: it is assured your work worthwhile, so enjoy your contribution.

Yes an additional delaying factor is possibly that they would wish to make 6990 OpenCL capable on both cores. If only newer drivers are available for 6990 and the OpenCL dual core bug is not fixed then 6990 may not be able to be used on MilkyWay.

I had a little try of PrimeGrid's OpenCL application. I mistakenly thought that the OpenCL bug on 5970 had been fixed. Didn't realise for some time that all the tasks completed and successfully reported on the second core were invalid after being checked by a wingman. When I came back to MilkyWay the same thing happened and I wasn't aware of it until I noticed a lower than usual daily total and a mass of pending tasks. Didn't know the OpenCL driver would muck up CAL applications as well. Will have to revert to an earlier driver when I get a chance. At least DNETC is working, still freezes occasionally of course on a 5970, but I doubt that will ever be fixed.

I'm a little puzzled about the expectations people have or had regarding the new cards. IMO it looks like this:

Price: in Germany the cheapest GTX580 is 450€, GTX570 is 320€ and HD6970 is 320€. Traditionally these prices translate 1:1 into $ prices (ours include 19% VAT). So the AMD is priced right at the 570 price spot, not the 580. And it's almost a 50% difference, so I see really no point in comparing the 6970 with the 580 (and complaining about the result).

Absolute performance: compared to Cypress AMD is typically extracting 20 - 30% more performance for 24% more transistors. That's a good move. If your starting point has already been very good there's only so much you can improve. In the past we've only seen the 2x increases in GPU speeds per generation because AMD / nVidia were free to double the transistor count, which their are not any more.

Efficiency: it's still built on the same 40 nm process. It's bigger and faster. It's only natural that it needs to consume more power in return. There's no free lunch here.

They didn't want to make 6970 as fast as GTX580: a 2.6 billion transistor chip is really hard to build and AMD wants to keep size and power consumption so low that they can pack 2 of these chips onto one card. Cayman is really borderline in this regard. I suspect we'll see about 800 MHz and a maximum of 1.10V for the chips, yielding a typical gaming power draw of 300W and a power limit at this value. nVidia choose to make their chip so big that they can not realistically do this (at least not with GF100 at somewhat competitive clock speeds and execution units).
Actually AMD pushed Cayman quite hard on the HD6970 to tie the GTX570. Going from 1.10V to 1.175V and from 200W TDP to 250W TDP for a mere 80 MHz more (+2 SIMDs) was probably no easy choice. This high voltage is also what makes the 6970 not very power efficient. HD6950 fares much better in this field.

It's hot: cooling 150W is hard, let alone 250W. In order not to deafen you the automatic fan speed regulation keeps the load temperature pinned at a maximum of 90°C and regulates fan speed / noise to do so. There's nothing surprising about this. Add better cooling or accept higher noise and it will run cooler, just as the other cards (keeping in mind the power draw, which determines the temperature for a given cooling capability).

FurMark: it's not a measure of power draw under real applications. It runs into the power capping of Cayman but drives the older cards way harder than any real software, that's why Cayman seems better here.

@Kashi: you might want to check wether you can lower your GPU voltages via software (e.g. ATI Tray Tools). This way you could save a lot of power for a comparably small hit in clock speeds and productivity.

Personally I'd like a 1 GB HD6950 due to the lower price and better power efficiency than the HD6970. Cooling probably has to be an Accelero Extreme.

MrS

---

Scanning for our furry friends since Jan 2002