I agree with you on the TDP part. 280 W of TDP for a reference card is just plain insane.

Here we go again.

Leaked slide: 190W
Official slide: 250W

So, HD 8970 will have 320W as minimum. Don’t forget XRD2 memory, lol.

You might be right about the TDP.
KitGuru today mentions June for the launch, which in my opinion is not true. February was proposed for AIB partners earlier. My source is quite good, I’m not sure about theirs.
http://www.kitguru.net/components/graphic-cards/jules/amd-radeon-hd-8970-and-8950-launch-plans-revealed/

AMD 5000 series was best, besides tesselation perfomance.
AMD 6000 the FLOP.

http://tomshw.it/cont/news/amd-radeon-hd-8950-e-8970-con-gpu-venus-al-massimo/39886/1.html
8950 decrease ROPS to 32, comparing 48 from 8970, that’s a big step back, but i think is enough to give Venus Pro(8950) 7970 GHz performance consuming less power.

According to given specification it seems to be only 25-30% better performance compare to HD 7970, but with less TDP. So, I waited more.

Thanks :)

i’m taking about the max power usage not declared TDP, they’re very different things like gas mileage on a car

Look what I´ve found:
http://www.youtube.com/watch?v=Pt1j6vdypsk ;)

HD8XXX in DEC 2012 HD 9XXX IN JUNE 2013 AND PROBABLY HD 10XXX IN DEC 2013 …..but still the loser remains a loser in every way

That doesn’t make sense. AMD HD5000 series was the one with terrible tessellation that HD6000 series later fixed. Also, HD6950 for $250-300 unlocked into a 6970 and delivered 90% of the performance of a $500 GTX580. There is no way you can call HD6000 series a flop.

Sea Islands is not on AMD’s roadmap until 2013. Therefore, retail launch is coming in 2013, not 2012.
HD9000 series will be Volcanic Islands I bet, made on 22nm node. That’s not coming until 2014.

The leaked slide is fake, from March 2012, with spelling mistakes too, and says HD8970 will have just 20% more compute than 7970 series. HD7970 already has 14% more compute than 7970 series (1050 vs. 925mhz). So if that slide was true, there would be no reason to release the 8970 with just 6-7% faster performance?

AMD 5000 was 1st to give us opportunity to play DX11 games, but yes tessellation was terrible comparing to GTX400.
I know what he means, performance jump from 4870 to 5870 was incredible, vs 5870 to 6970.
4870×2 +/- 5870, 5000 flagship on pair with 4000 Dual-flagship, of course don´t include 4890.
6970 is not near 5970 performance.

HD6000 and GTX500 were ‘Refresh’ generations, especially GTX500 which was just a GTX485 basically with minor tweaks to Z-culling and FP16 texture performance. If you had at all followed the 32nm manufacturing node issues, you would have known back then that HD6970 and GTX580 were going to be 15-20% faster. AMD even said they had to split up their 1 generation into 2: Northern and Southern Islands because of the 32nm node failure.

There are plenty of games that you can play on 6970 that you cannot play on the 5870 with tessellation and VRAM:

1. Skyrim with ENB mods won’t work on 1GB 5870 vs. 2GB unlocked 6950. You can see here any GPU with less than 1.28GB of VRAM tanks in performance in Skyrim:
http://www.anandtech.com/show/5625/amd-radeon-hd-7870-ghz-edition-radeon-hd-7850-review-rounding-out-southern-islands/13

Also, Total War Shogun 2 needs > 1GB of VRAM for 1080P maxed out.

2. In modern game with heavy tessellation, HD5870 is 50-75% slower than 6970. Look at Batman AC or Crysis 2:
http://images.anandtech.com/graphs/graph5699/45143.png

Regardless, HD6970 and GTX580 were refresh series, not a brand new generation of GPUs. Anyone who expected the same increase that happened from 4870 to 5870 to apply to 5870 –> 6970 didn’t follow GPUs then or what was happening in the fabrication industry at TSMC.

Poor show 6990 :)
I know they were refresh generations, the point is: came from 4870 to 5870 was a better upgrade than 5870 to 6970.
5870 was bad with some titles/games in tesselation, but for some reason AMD 5000 was one of best selling generation, even they had less Vram, compared to 1.5GB GTX 480.

There are some cases than 5870 wins GTX 560 TI, and almost reach 6950, but AMD don´t give a damm about previous generation cards, talking about drivers. HD 4000=DEAD

Vram isn´t everything, many test show that 6950 1GB or 2GB perform same as 1080p.

No i don´t know what happening with TSMC, LOL, and really don´t care, like what is actual AMD or Nvidia CEO what matters?

We need fast and lower TDP card´s, fabrication process are always a problem when a new nanometer tech is out, but with time they improve yields…

Just hope this new generation will comes with fair prices and less driver problems.
http://forums.guru3d.com/showthread.php?t=364498 Every new generation people create this kind of topics.

Sapphire 7950 OC hands down, especially if it’s the 950mhz version. Those overclock to 1100-1150mhz = GTX680/HD7970 speed!

I agree that 5870 was a great upgrade and 6970 was not. The same can be said for GTX285 –> 480. 580 was a lousy upgrade from the 480 as well since 480 could overclock to 580 speeds. Even with the power consumption reduction, the 580 still used a fair amount.

HD4000 is not dead. AMD still releases drivers for it but not monthly, instead quarterly. This makes sense since HD4000 series is not DX11 and cannot really play modern games. In other words, the drivers focus on bug fixes, not performance improvement. The same for GF8/9/GTX200. There is no performance increase left in those old parts.

I never said VRAM is everything mate. What I meant is that if a person called HD6900 a flop, so would be the GTX500 series. The performance increase from 5800/480 series was equally lackluster for both camps.

I don’t need we need lower TDP parts everywhere though. I want NV and AMD to go back to 250W TDP flagship parts since I want more performance on the desktop. If I cared about power consumption, I’d get a Wii U as it uses under 75W, or PS3 even. AMD and NV should make lower TDP parts for low-end and mid-range desktop cards. For high-end $500 flagship, I want as much performance as possible, even if it means 250-300W cards. That’s why we have 1000W PSUs. PC gaming is not cheap. 2-3x $500 flagship cards is $1,000-1,500. For high-end PC gamers, electricity costs are not a factor. And for gamers who want only a 140-150W part, AMD and NV have the mid-range series every generation.

yea it was 950mhz… and you just posted my thought…i was also gonna do that..lets just hope this offer stays for more 15 days…then i can buy that hd7950 ooh…

Exclusive: AMD HD8970 Venus XTX Block Diagram

http://www.arabpcworld.com/?p=19424

ROPs and TMUs don’t provide any performance themselves. SIMD Engines providing it. And HD 8970 will be the same as HD 7970, but with 8 additional compute units and 8 additional ROPs. (40 units, 64 SIMDs per unit: 40×64 = 2560 Cores). It means +25% to SIMD count, +50% to ROPs count and +25% to TMUs count. With better microarchitecture and higher clock all it gives 25-40% better performance in games compare to HD 7970 1 GHz.

KitGuru has 2133 SPs for 8950. How did they post an odd number of shaders in their article and managed to claim their info is credible with a straight face? ;)

Oh, I meant 16 ROPs…

It’s system setup power consumption. Not only card.

they use furmark for the max power so it have no relation with real gaming
gtx 480 peak+average consumption =real consumption when u play a game so always check the average and the peak consumption in TPU reviews

10% is realistic. Figure the best ROP/sp (including sfu) for 48 is around ~2800 (just like 32 is in-between 1792-1920). That’s means likely 12 SMX for geforce, more-or-less equating to 2688 units from AMD. Throw in slightly more texture power per flop (224 shaders + sfus/16tmus vs 64sp/4tmus) and whatever other little efficiencies that typically make up around 5% per clock, and 10% seems realistic if neither are clock/bandwidth restrained within 300w. Again, sounds feasible…if clock is not limited by tdp. AMD could make up ground with higher clock potential within the same tdp. It might not be 10%, but knowing AMD the parts will make sense given the value equation.

While one would think 6 months late would make sense for such a big chip, I think the unbalanced inherent design (3360 units with 48 rops…err no) for gaming is a key tip-off that 780 will use harvested chips to make the part an actual and cost-efficient reality, while fully-working chips will go into workstation parts running lower clocks and having larger buffers. The trade-offs make sense. The ratio of yields needed per market makes sense. In theory it is a good idea on their part.

As far as this analysis goes, it’s hard to disagree with most of it, but it doesn’t stake a claim on the thing that is realistically arguable…’3 or 4gb’ is avoiding the issue completely. Obviously 2560-2688sp makes the most sense. Given AMD’s love of round and even numbers that are a factor of 8, as well as the CU count that fit the expected hierarchy of SI/CI (4/6/8/10/12/14/16/20/24/28/32/36/40) 40 makes more sense than 42. That said, I also think 64/512-bit (5gbps) may make more sense than 48/7gbps given what we’ve seen from AMD in the past. We know a 7gbps/384-bit memory controller would be bigger than a smaller 512-bit one. The die size insinuates there is enough pad space. 3 Setup engines would be very odd compared to 4. Etc etc etc. Not saying 384-bit isn’t the most plausible, just saying I wouldn’t be surprised if AMD capitalizes on the strengths of a larger bus (cheaper ram, running slower for tdp space, 2gb as a value/1080p option, 4GB for single-card extremely high rez, 8gb for workstation cards, perhaps cutting down ROPs/bus width/buffer size etc for salvage skus while still having similar bw to gk110 models.)

Does anyone else miss the days when the dual GPU cards were $499 (GTX 295 for example)? These 28nm price increases really suck for the consumers. Looking at the specs of the 8970 vs 7970, I don’t think it stands a chance vs GK110 if NVIDIA gives us a fully unlocked GK110 GeForce card with a 512 bit Bus and 2,496-2,880 Cuda Cores (7.1B Transistors). That is a big IF though, a fully unlocked GK110 would have a hefty price tag I’m sure, especially considering the 28nm pricing.

I used this article:
http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_480_Fermi/30.html
However Zotac improved power consumption by a lot, so the more likely consumption was 220 W in heavy gaming. http://www.techpowerup.com/reviews/Zotac/GeForce_GTX_480_Amp_Edition/27.html

I use Windows 7 32-bit in my Laptop, that has a old Ati X1100, DX9.
What i ask is if AMD could provide drivers for old card´s when Windows 8 will official launch, i dont talk about games. Driver updates are necessary to stretch screen(HDMI), change color profile/brightness etc and CCC is helpfull, besides media players bug fix.
There are many DX9, DX10 native games that i like, example: Crysis 1, BF2, Counter strike, Company of heroes…
I remember when brought this laptop, i had a problem with VLC that appeared grey stripes and froze computer, after driver updates back in 2009, problem was fixed, so drivers are not only for gamers.

AMD 7870(mid-range) is faster than 6970(high-end) and consumes lot less power.
So, why waste money now on 7970, when rumors of 8000 series start appear, and for example tell us that, 8870(mid-range) will have a 160W TDP, less 15W than 7870, and performance should compare to 7970.

Your laptop can actually play those old games with 1 core, 1GB of video ram and X1100?

So here is the question then, why would you spend $ upgrading such a dinosaur to Windows 8? Windows 7 is a great OS to begin with. Sounds like you are creating problems for yourself. If you want Windows 8, you’ll benefit greatly from a laptop with high PPI as well since W8 scales resolution well. Maybe you should consider upgrading at some point!

1. I must have been thinking about the old rumors of the GK110 having a 512 bit bus like this site itself has said (as did many others): http://videocardz.com/31650/geforce-gtx-685-gk110-features-4gb-512bit-memory when I typed 512, I knew the K20 is 384.

2. to quote myself, “That is a big IF though” and the GTX 480 sold pretty well despite its TDP.

3. The 680 also uses around 70W less at the wall outlet and you are comparing a stock NV card with an AMD OCd 7970 which is achieved purely through binning and BIOS voltages, hardware is identical. I’m sure if NVIDIA bumped voltage and the clock speeds across the board through binning they could achieve the same thing. Hell an OCd 7970 GE uses more power than the stock dual GK104 GTX 690 in both Metro 2033 and OCCT, probably why we aren’t going to see a 7990.

http://images.anandtech.com/graphs/graph6025/47618.png
http://images.anandtech.com/graphs/graph6025/47619.png

The PowerColor 7990 Devil13 showed it using 3x 8 Pin, talk about ridiculous. (http://videocardz.com/34576/powercolor-announces-radeon-hd-7990-devil13-6gb)

4. I was simply making an observation on the 7970 vs 8970 rumored specs compared to the GK104 vs GK110 rumored specs. 5-10% faster can be achieved through OCing right now with the GK104, especially considering the 7970 GE OCs barely, if it all, faster than 7970 OCd cards even though it has more voltage:

http://www.anandtech.com/show/6025/radeon-hd-7970-ghz-edition-review-catching-up-to-gtx-680/17

“Depending on how you want to count this overclock amidst the presence of
the boost clock this is either 25MHz better than our best 7970 card, or
75MHz better. In either case our 7970GE definitely overclocks better
than our earlier 7970 cards but not significantly so, which is in-line
with our expectations.

With AMD now binning chips for the 7970GE we’d expect to see some
stratification among the 7970 family such that high overclocking chips
that would previously show up in 7970 cards will now show up in 7970GE
cards instead. For penny-pinching overclockers this is not good news,
but for more hardcore overclockers this is nothing new as AMD’s partners
have been doing something similar with their factory overclocked cards
for some time now.

Meanwhile our memory overclock isn’t significantly different from what
we could pull off with the reference 7970. The limitation is the memory
bus or Tahiti’s memory controller, neither of which has changed. After
around 6.4GHz errors start catching up and performance gains become
performance losses.”

5. Inflation alone doesn’t explain the top of the line pricing over the last 5 years:

2007 – 8800 Ultra – $999 (90nm)
2009 – GTX 295 – $499 (55nm)
2012 – GTX 690 – $999 (28nm)

(used one brand’s top of the line for obvious comparison reasons)

Keep in mind my original post is based on rumored information from both sides, so everything we typed could be meaningless when we get verified specs. Until then, I’ll debate computer hardware all day long : D

There 3 reasons:

- If you don’t want to wait 4-5 months until February 2013 and want to play games starting today
- In North America we can use the 7970 to make $ every day until 8970 launches.
- If you can find a 7970 for a very good price ($312 on Newegg for example:
http://www.newegg.com/Product/Product.aspx?Item=N82E16814103201)

But ya, if you can wait 4-6 months, of course GPUs get even better :)

1. Inflation is 1 factor but not all. If consumers are willing to buy $500 cards that are 30-45% faster than last generation, then AMD and NV don’t need to increase performance by 75-100% like they used to. Supply and demand.

2. NV and AMD have both complained about rising wafer costs and they haven’t passed on those costs to use realistically for many generations. 8800 Ultra was $830 in the US and that card doesn’t even count because it was a limited run. 8800 GTX was $599. Also, keep in mind both the 680 and 7970 use less power than the 580. Both AMD and NV traded some performance for lower wattage, especially NV.

3. Using reference 7970 GE for any metric outside performance = meaningless. Noise levels, power consumption, temperatures = all worthless unless you plan to flash your 925mhz 7970 with the June 2012 Boost BIOS. After-market 7970 GE and normal 7970s overclock very very well. Not sure what you mean that they barely overclock. You can hit 1150-1200mhz on most 7970/7970 GE cards and good ones reach 1300mhz.
http://www.legitreviews.com/article/1962/15/

Not sure why you even brought up the 7970 GE overclocking. My main point was 2-way:

–> GK106 and Pitcairn XT have very similar performance/mm^2, so if you add dynamic scheduler, compute, 384-bit bus, double precision to GK104 you have GK110. Now we know that GCN and Kepler are very similar in performance per clock (but actually GCN is about 10% faster per mm2 since 7870 is slightly faster than the 660 despite only being clocked at 1ghz vs. 660 that boost to 1150mhz+ in real world). So that means a GK104 chip with all the added features of Tahiti XT would weigh in at 350-365mm2 but not be any faster (because we still didn’t add any functional units).

Ok now start adding functional units to Tahiti XT and to this hypothetical 365mm2 compute GK110 => 410-420mm2. Remember GCN is 10% faster per mm2, that means GK110 410mm2 would be slower than HD8970 at 410mm2. That’s my point that 680 is 5-10% behind 7970 GE 1.05ghz despite 680 boosting to 1100-1150mhz as well.

This is why even if Kepler GK110 is 520-550mm2, realistically it would only be 20-30% faster, not much more and that’s if it boost to 1100-1150mhz like 680 can.

Do I believe 780 will beat 8970 if NV launches a full-blown GK110, sure. Do I actually believe NV will release a 60-75% faster single-GPU card than 680 next year? No, for TDP and marketing reasons. NV could launch a 2880 SP GK110 @ 1ghz but it’ll probably top 250W in no time. Secondly, such a massive performance increase leaves very little room for a fanfare launch of Maxwell. NV needs only to beat 8970 by 10-15% and it could easily command a $100-150 price premium. Why would NV go balls to the wall like they did with GTX480 and have a hot running and dust buster flagship card? All they need to do to reclaim the performance crown is beat 8970 by 15% solid and they can price it at $650.

I’ve read some claims how people think GK110 will beat 8970 by 50, 75, 100%…that’s like fanboy talk :)

I pretty much agree with all you said there. I never meant the GK110 was going to beat the 8970 by 75-100%, you are right, that is fan boy talk. The reason I showed the 7970 GE OC was to show that the vanilla 7970 could already hit those speeds and further and to show the insane power draw an OCd 7970 GE pulls (topping the 690 by 55W). Needless to say, NV has a lot of room to play with the TDP of Kepler because of how well Kepler does in watt vs performance, unlike the 7970 GE as stated/shown above in the charts.

I didn’t even think about Maxwell’s role in GK110 launch/performance. You make a very good point there, NV doesn’t need a huge jump over the 8970 due to Maxwell coming (probably my next card). Kepler was really just a refined version of Fermi if you read both of their Whitepapers (see bottom of this post) and Maxwell will be a completely different architecture, like GT200b vs GF100. It is really nice to have a site you can discuss hardware without a backlash of fan boys, I know they exist on the site, but they don’t bother to go in depth like you or I are (and others of course). About the 8800 Ultra, at launch it did retail for $999-$830 depending on model (if you remember some Ultras were actually above $1000), the 8800 GTX was ~$650 at launch: http://www.anandtech.com/show/2222

The GTX 680 can be found all over NewEgg for $479 right now w/ Borderlands 2 (which is an insanely popular $60 game). The 7970 you linked is discontinued and it was $394 after MIR, the avg 7970 price is still $420 (avg of all 7970 on NewEgg and I didn’t include the insane 6Gb models or water block editions) and the GE are avging about $435 and these all include Dirt Showdown, which is much less popular than Borderlands 2 of course (some people take free games into consideration with prices, so I included them).

One thing I noticed about the 7970 GE on NewEgg is there is only 6 models available for sale, while you have 30 GTX 680 models available.

NewEgg 7970 GE models:
http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&DEPA=0&Order=BESTMATCH&N=-1&isNodeId=1&Description=7970+ghz+edition

NewEgg 680 models:
http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&DEPA=0&Order=BESTMATCH&N=100006662&isNodeId=1&Description=GTX+680&x=24&y=21

Going back to 7970 GE OCing, I wasn’t saying the 7970/7970 GE didn’t OC well. What I was saying is the 7970 released in 2011 OC just as well as the Ghz Edition produced in 2012 (which they should since the hardware is identical), that is all I was saying. The other reason I brought it up was the 7970 GE is just an OCd 7970, so in essence you are comparing an OCd 7970 to a stock GTX 680. Which is why I said you can make up that 5-10% today with a mild OC on a stock 680. The main problem with the 680 is that 256 bit interface, I couldn’t believe it when I read it for the first time. I mean hell, the GeForce FX 5900 (2003) had a 256 bit interface… That is why the 7970 spanks the 680 at 2560×1600+ resolutions in some games. Nice chatting with ya : )

For anyone interested in the whitepapers on Kepler and Fermi:

Fermi: http://tinyurl.com/8d9mlsw
Kepler: http://tinyurl.com/c3drc6d

There is actually a GTX 680 whitepaper and a GK110:

GK110: http://tinyurl.com/c3drc6d

Brilliant article but why should I even regard it as accurate when you do not even have enough regard for convention to use the correct symbol as a decimal place, making the whole thing counter-intuative and confusing?

Make money how? that joke bitcoins?

Great post LedHed!!! I enjoying coming here since we just talk about hardware and try talk about various GPUs, coolers, SKUs, future GPU specs, etc.

I think the other point I wanted to make: Remember in 5-7 years ago the node shrinks were much easier and NV and AMD/ATI were able to increase performance dramatically every new major generation (GeForce 5–> 6 –> 7 –> 8, etc.). As we are slowly approaching the physical limit of something like 5 atoms across, it’s becoming harder and harder to shrink transistors. This is why TSMC had issues at 40nm and NV and AMD decided to say well under 400mm2 die for safety reasons too. For NV, they won something like 300+ contracts for mobile Kepler. So imagine if they sold 520-600mm2 GTX680 — it would have brought a strain on their ability to meet their other client obligations. These things are probably planned 6-9 months in advance and I bet NV knew TSMC would be wafer constrained for much of 2012 since TSMC themselves discussed that production would ramp up to good volumes only by Q3-Q4 2012. I don’t think NV intended to launch GK110 this round not only because it was too risky, but also because it didn’t go well with their plan of winning more mobile contracts.

Also, as we saw in the past if you try a large die on a brand new node, it often gives you a lot of problems. AMD tested out HD4770 on a newer node and then launched the flagship later. I think instead of launching GPUs 75-100% faster like GeForce 6800U was over 5900 or X1800XT was over X800XT, both AMD and NV are going to split the generations into 2 parts: and each delivers 30-40% more performance. This gives them more room for error and less chance of ending up with a GTX480 style delay and heat. GTX680/7970 gave 35-45% over their predecessors. If HD8970/780 can also deliver 30-40%, we would end up with the same 75-100% over 2 years. The difference is in the past they would do a major upgrade, and then a small 10-15% bump (and you’d wait 2 full years for a huge increase). It seems now they are spreading out the risk of newer node but still delivering enough to get enthusiasts to upgrade. The real boredom has set in in CPU space. Even if GPUs increase 30-40%, that’s still “overkill” right now with console ports!

Seems sometimes is difficult to explain my point of view, especially in English language.
I hope Windows 7 driver will works with Windows 8, like Vista works with 7.

I tested PCIe 3.0 cards in older boards for fun, have a 650i by Nvidia even and it takes a 7970. Right now what we need at 1 TB SSDs for $200 and next generation games. The motherboard and RAM have become the most bland components since they hardly contribute to performance nowadays. I mean the top board might net you 150mhz more on the CPU but it’ll cost $150 more for that. I bet a $130 Asrock Z77 Extreme 4 is 98% as good as the Asus Maximus $350 Z77 boards for overclocking.