AMD Radeon 7000-Series 28nm (Southern Islands) | 7990 7970 7870 7770 | Discussion

[tviceman]

the 7970 is not looking so so good here. apoppin acts like it could be slower the 6970 in some cases. it looks like the gpu division might have followed the cpu division's handbook. :biggrin:


http://alienbabeltech.com/abt/viewtopic.php?f=6&t=23899

Yeah he thinks parts of Tahiti will be disabled even in the hd7970. Admittedly I don't know if he's right or not, but my guess is that hd7970 is a fully functioning Tahiti GPU - it just wasn't able to hit the target core operating frequency within the TDP that AMD wanted to stay under. In which case, AMD will likely do a deep respin of Tahiti sometime next year, allowing a significant boost in core frequencies while staying at the same TDP and re-release it as hd7980 or something like that.

But yeah, what Apoppin is saying to me confirms what I thought about this Dec. 22nd launch. Paper release to limited reviewers for bragging rights of the fastest GPU in 2011. No real availability until Jan. 9th or 10th, and unless Nvidia is having similar yield problems, Kepler is going to leapfrog Tahiti version 1 quite easily. I hope AMD doesn't price themselves too high and end up losing sales. Nvidia has definitely shown in the past few years that while they will definitely charge a small premium for similar performing parts, they are more reactive to price shifting than AMD is.

 

[3DVagabond]

the 7970 is not looking so so good here. apoppin acts like it could be slower the 6970 in some cases. it looks like the gpu division might have followed the cpu division's handbook. :biggrin:


http://alienbabeltech.com/abt/viewtopic.php?f=6&t=23899

He also states that AMD didn't give him one. We'll see for sure soon enough, but it could be feeling scorned by AMD. We don't know how many cards they have. ABT just might not be high enough up the list to get one.

 

[tviceman]

Looking forward to reading some reviews from people that have an actual card. That thread was just more fud.

It really isn't. He's a respectable hardware reviewer and probably has the largest benchmark suite amongst all GPU reviewers. He is not at all biased in any of his reviews and, if anything, always puts a positive spin in review summaries even if products end up being disappointing. I have never once seem him post fud about anything and claim it to be straight up fact.

 

[lavaheadache]

Apoppin is scaring me into keeping my 580. I was planning on selling it to fund the purchase of a 7970 but I really don't want to buy a second rate product.

 

[Silverforce11]

It's ~380mm2, if they have serious yield issues, so will NV, and it will be worse since their dies are much bigger typically.

I don't think its crippled, because the architecture layout (even from way back at AMD fusion summit) doesn't really fit with extra CU clusters, each extra CU requires the corresponding 16 TMU and 4 ROP along with cache to be functional. If they have 2 CU disabled, thats 2176 1D shaders, 40 ROPs and 160 TMU, seriously. Can it even fit on such a small die given their 1D shaders are likely to be quite huge? They have no reason to cripple their top end card, AMD needs the perf crown and their pricing indicates its the king of the hill deserving of $550. BD may be crap, but it doesn't sell for ridiculous prices beyond its perf.

Their previous architecture has ROPs linked to the memory controller. This architecture looks like there's a global interface to/from memory contr, so ROPs are independent (linked to CU clusters instead).

 

[Firestorm007]

He also states that AMD didn't give him one. We'll see for sure soon enough, but it could be feeling scorned by AMD. We don't know how many cards they have. ABT just might not be high enough up the list to get one.

Agreed. It sounds like he's letting his emotions get the best of him. We'll find out soon enough how good the 7970 is... Maybe AMD just doesn't like Apoppin...^_^ I'll wait for the real benchmarks.

 

[Saico]

I don't think apopin deserves one.

 

[Silverforce11]

MOAR slides! The last one is actually awesome
yes.. it will be 50% better for mining

This slide is extremely telling. VLIW4/5 typically get full utilization for bitcoin mining already (SHA). 7970 beats it by 50% (35% more SP, faster clock speed = ~50%). This confirms 1D shader is flexible and powerful and their new architecture is fully efficient in terms of utilization similar to NV.

In everything else where VLIW was at typically 70% efficiency (including games), GCN 1D shaders gives it a spanking.

If this slide is legit, i'm confident we can see >50% Cayman perf.

 

[RussianSensation]

If what apoppin says is true, specifically:

1) HD 7970 is only 20-30% faster across the board than the GTX 580.
2) And if you run DX9 games, the HD 6970 is likely to win a few benches against the HD 7970

If you put those two statements together, it appears that the largest improvement in 7970 is going to come from Tessellation. This is why it will be barely faster than a GTX580 and perhaps explains why the only performance slide showed games/benches that use specifically use Tessellation. It is interesting that BF3 or Witcher 2 weren't present.

If this is true, AMD is going to be in serious trouble once a full-fledged GTX680 launches. The chance of GTX680 only 30% faster than a GTX580 is 0. The last time AMD introduced a brand new architecture, it was HD2900XT, and that didn't go so well.

If you have to ask that if the 32 ROP figure is true, WTF where they thinking? They've had 32 ROPs since HD5870 serires in 2009...

 

[toyota]

If what apoppin says is true, specifically:

1) HD 7970 is only 20-30% faster across the board than the GTX 580.
2) And if you run DX9 games, the HD 6970 is likely to win a few benches against the HD 7970

If you put those two statements together, it appears that the largest improvement in 7970 is going to come from Tessellation. This is why it will be barely faster than a GTX580 and perhaps explains why the only performance slide showed games/benches that use specifically use Tessellation. It is interesting that BF3 or Witcher 2 weren't present.

If this is true, AMD is going to be in serious trouble once a full-fledged GTX680 launches. The chance of GTX680 only 30% faster than a GTX580 is 0.

Nvidia can just hold back the clocks on their next gpu and still easily beat them.

 

[Silverforce11]

If what apoppin says is true, specifically:

1) HD 7970 is only 20-30% faster across the board than the GTX 580.
2) And if you run DX9 games, the HD 6970 is likely to win a few benches against the HD 7970

If you put those two statements together, it appears that the largest improvement in 7970 is going to come from Tessellation. This is why it will be barely faster than a GTX580 and perhaps explains why the only performance slide showed games/benches that use specifically use Tessellation. It is interesting that BF3 or Witcher 2 weren't present.

If this is true, AMD is going to be in serious trouble once a full-fledged GTX680 launches. The chance of GTX680 only 30% faster than a GTX580 is 0. The last time AMD introduced a brand new architecture, it was HD2900XT, and that didn't go so well.

If you have to ask that if the 32 ROP figure is true, WTF where they thinking? They've had 32 ROPs since HD5870 serires in 2009...

1. Apoppin doesn't have review samples, his remarks are speculation.
2. The leak benches have games where no Tessellation is used or not used to the point of bottlenecking Cayman, only Crysis 2 on that list fits your criteria.
3. BF3 numbers look too good to be true.
4. 32 ROPs but who knows what else they've done to tweak it? Just the revised layout alone could boost performance.

 

[RussianSensation]

This slide is extremely telling. VLIW4/5 typically get full utilization for bitcoin mining already (SHA). 7970 beats it by 50% (35% more SP, faster clock speed = ~50%). This confirms 1D shader is flexible and powerful and their new architecture is fully efficient in terms of utilization similar to NV.

Actually based on your own math, it doesn't really show that GCN is much more efficient than VLIW-4.

For example, if you simply replaced the rumored 7970 specs with VLIW-4, you'd have:

- 50% more memory bandwidth
- 40% more texture performance
- 40% more shader performance (2048 SPs @ 925mhz using VLIW-4 tech)

It looks like the geometry engines got a huge boost, which is why the slides show much improved performance in Tessellation. But they could have improved these engines in VLIW-4 architecture as well.

So what you have then is:

HD6970 (100) => GTX580 (115) => HD7970 (144-150)*based on apoppin's estimate of 25-30% faster than a GTX580.

That means HD7970 would be about 44-50% faster than an HD6970, which is actually in line with what a VLIW-4 architecture might have given us if you simply increased HD6970's specs to 7970's levels.

If GCN was WAY more efficient, we'd be seeing a 65-75% performance increase over HD6970. I think HD7970 is primarily about GPGPU compute and performance per watt increase for games is minimal, if not flat.

And here is the best part: Instead of putting GCN "fluff" such as 32 CU units that are useless for gaming, you could have increased the ROPs to 48. A 2048 SP VLIW-4 card with 48 ROPs and everything else the same would actually cream the HD7970 since we know Radeons are lacking pixel fill-rate vs. Fermi.

 

[Silverforce11]

Actually based on your own math, it doesn't really show that GCN is much more efficient than VLIW-4.

No you don't get my maths. If it was 2048 VLIW4 SP, it would be 50% faster at SHA256 and also 50% faster for other instructions (slide has it 70% to >100%, VLIW chokes with transcendental instructions). SHA256 is near 100% effective on VLIW thats why bitcoin flies on AMD radeons. Everything else perform at around 3.4/5 or 3.2/4 per VLIW shader.

Thus, 2048 1D shader performance should be taken as if x VLIW 4/3.2 = ~2560 VLIW shaders effective (because its 100% utilized). Makes sense?

 

[Grooveriding]

A lot of worthless conjecture in that thread, on a site that generally raves over anything nvidia puts out, whether that is physx, 3d or even the laughable review about the gtx 560 448 that was met with lukewarm reviews from most respectable review sites.

Sounds like a lot of sour grapes about not being on the short list to get a 7970 like the more respected sites have. I think we all can be certain anand and Ryan Smith have one in house an will give us a good, impartial and balanced review. ABT has a worthless benchmark suite of archaic games used to give flawed overall review conclusions. Most of the games benched in those reviews could be maxed out on DX10 class hardware. I have no interest in seeing what card gets 200fps vs another getting 150fps @ 1080P. Useless data.

I want to see how the card does in BF3, Crysis 2,Civ 5, Metro 2033 and perhaps Crysis 1 just because it still puts a hurting on hardware. Sadly, most other games out there have been beaten down by the hardware we already have.

I can believe the 7970 will be AMD's GTX 480; in it being hot, loud and due for a refresh down the line to give it a 20% performance boost.

Remember this: http://www.anandtech.com/show/2977/...tx-470-6-months-late-was-it-worth-the-wait-/9

The 480 was only 30% faster than the 285 for the most part at release as well. The 7970 being 30% faster than the 580, not the 6970 which is slower, is awesome.

I would expect the 680 to be faster than the 7970, the same way the 480 was faster than the 5870. It would be a sad state of affairs to take 6 months longer to release your new part than the competition and not be faster. That's how it goes with GPUs, as time goes on, they get faster. What made the 480 so underwhelming performance wise compared to the 5870 was the 15-20% performance advantage it had was unimpressive in the context of it taking 6 months more to develop and release.

The 7970 being 30% faster than a 580 will be a success and will sell. The people who buy these $500 GPUs will do so happily. Like I said in the past for myself personally, if the 7970 comes out with a proper performance boost over 40nm cards and nvidia has no solid launch date and zero anouncements, I'll be buying them. If anything this is even more of a draught on anything from nvidia, when the 5870 launched they were already starting the rumor mill and they had nothing even close to ready.

It's irrelevant if the 680 will be faster or not, I'm sure it will be. But if you're buying one of these cards rather than running commentary on them, who cares, when the 680 could be released in May or June 2012. The 7970 will be here in days/weeks and will kick the ass of everything out there.

There is always something new coming, it's only relevant if it's just around the corner, not some hazy unknown. These sorts of arguments are irrelevant though, like that sour grapes thread on that forum. Rather than looking forward to seeing what the new card brings, whether good or bad, sour grapes lead to trying to pooh-pooh on a new release with an unconfirmed and launch-date lacking unknown product.

 

[tviceman]

It's ~380mm2, if they have serious yield issues, so will NV, and it will be worse since their dies are much bigger typically.

I like how you immediately turned AMD's potential problem into Nvidia's bigger problem. You need a jump to conclusions mat from the movie Office Space. Silverforce, I know you know that there are so many other factors just as relevant as die size that affect yields. Transistor density? Transistor composition? Design flaws?

If die size is as big a factor in yields as so many people like to claim when talking up AMD and talking down Nvidia, then Nvidia would have gone belly up with G80.

Back on topic, I think apoppin's reasoning is sound. Release now, even if it's not hitting the speeds they want within the TDP they aimed for instead of waiting for a respin and potentially coming to market at the same time or even later than Kepler. So instead of just respinning, AMD will have a product out now and can go in and make bigger tweaks to the design that isn't possible with quick respins.

Sounds like Tahiti isn't bulldozer, but definitely isn't Cypress either. It's coming early to get AMD a head start, but it still needs a little TLC.

 

[apoppin]

Agreed. It sounds like he's letting his emotions get the best of him. We'll find out soon enough how good the 7970 is... Maybe AMD just doesn't like Apoppin...^_^ I'll wait for the real benchmarks.

Emotions? It is business and it is no biggie. AMD PR is gutted. Many of my former AMD contacts are now working for Edelman (who is handling the HD 7970 launch). And i'll get it eventually from one of AMD's partners at CES and ABT will get a lot of traffic from the article that will leak more information tomorrow.

FIRST of ALL, the HD 7970 is released the 22nd. Paperlaunch. The the 7950 is released at CES on Jan 9 and there should be availability at retail then.

i believe AMD's supply is really limited and frankly, they don't want me to test 30 games. Just the DX11 games.

i don't know for sure if the chip is crippled or not - like GF100 was. But the low clockspeed, large die size and poor performance indicate it might be
http://www.chiphell.com/thread-329802-1-1.html
More or less translation: "Beats GTX580 by 20% on average, not 3DMark. Beats GTX580 by around 30% in Crysis. Expected goal has been reached. There has been no goal of doubling the performance of previous generation cards as speculated."

i'm going to write a preview tonight. Look for the details there. i cannot be here and at ABT forum also. i also see plenty of slander from Grooveriding, but then he has a real (hidden) reason to do so. Anyone who knows ABT, knows we use more modern DX11 games than any one else. And we bench at 2560x1600 and 5760x1080.

Peace and aloha

 

[DarkKnightDude]

Much looking forward to the reviews now. Speculation leads everyone everywhere.

 

[Silverforce11]

i believe AMD's supply is really limited and frankly, they don't want me to test 30 games. Just the DX11 games.

I often check benches on your site. But could you please add more newer games to your bench list. There's a lot of older games where their popularity has waned out long ago.

Definitely add BF3, Skyrim, Anno 2070 and even Starcraft to your list, all very popular games that heaps of gamers play.

Edit: You know its a paper launch when Napoleon @ Chiphell doesn't receive a sample for review. heh

 

[RussianSensation]

2. The leak benches have games where no Tessellation is used or not used to the point of bottlenecking Cayman, only Crysis 2 on that list fits your criteria.

Not true. In Lost Planet 2 and here, Unigine (extreme tessellation) and Crysis 2, HD6970 gets slaughtered. So that's 3 out of 5 benchmarks where HD6970 tanks because of Tessellation. How convenient that AMD chose to cherry pick those very same benchmarks where GTX580 is already 30-45% faster?

3. BF3 numbers look too good to be true.

I must have missed it, but where do you see BF3 benchmarks for HD7970?

Thus, 2048 1D shader performance should be taken as if x VLIW 4/3.2 = ~2560 VLIW shaders effective (because its 100% utilized). Makes sense?

No, it doesn't make sense to me. Those benchmarks are related to some specific instruction sets that have nothing to do with games. I don't care how fast the card is in Mandelbrot or in AES256. In MilkyWay, my 6970 has 99% usage and creams my previous GTX470 by a factor of 3x. I am not going to use that example to project its gaming performance, am I? Let's stick to real world benchmarks, not theoretical programs that have little to do with gaming code.

Let's wait for benchmarks before proclaiming that GCN 2048 shaders = 2560 of VLIW-4 style SPs. No architecture is 100% utilized since you always have a bottleneck somewhere (whether it's TMUs, SPs, ROPs, memory bandwidth, etc.).

 

[3DVagabond]

If what apoppin says is true, specifically:

1) HD 7970 is only 20-30% faster across the board than the GTX 580.
2) And if you run DX9 games, the HD 6970 is likely to win a few benches against the HD 7970

If you put those two statements together, it appears that the largest improvement in 7970 is going to come from Tessellation. This is why it will be barely faster than a GTX580 and perhaps explains why the only performance slide showed games/benches that use specifically use Tessellation. It is interesting that BF3 or Witcher 2 weren't present.

If this is true, AMD is going to be in serious trouble once a full-fledged GTX680 launches. The chance of GTX680 only 30% faster than a GTX580 is 0. The last time AMD introduced a brand new architecture, it was HD2900XT, and that didn't go so well.

If you have to ask that if the 32 ROP figure is true, WTF where they thinking? They've had 32 ROPs since HD5870 serires in 2009...

20%to 30% faster than the gtx 580 is way out of the 6970's league. I don't know how a card that fast is going to lose in anything to the 6970. We also don't know what benches were run where those figures came from. It could be benches at resolutions that the 580 has a 25% lead in rather than the typical ~13% overall.

While the gtx 680 will likely outperform that level, it's rumored to not be coming for many many months.

source

We could have AMD's next gen/major refresh by then. All these people quoting 24mos. between gens, that's nVidia's time table with a refresh in between. AMD said new gen every ~12mos.

 

[Silverforce11]

Not true.
I must have missed it, but where do you see BF3 benchmarks for HD7970?

I was referring to the leak slide with a lot of games. ~1.5 to 1.6 faster, not the unigine/crysis 2 slide. LP2 is so TWIMTBP i don't even know if its tessellation thats bottlenecking the radeons. Only certain of Unigine extreme and Crysis 2.

BF3 was included in my previous slide link.

As for the maths issue, it's pretty damn simple.
VLIW in games operated at 3.4/5 OR 3.2/4 average in optimized games. Its never operating at max efficiency. Key factor here.
That means for Cayman's 1536 4D shader, its average performance is ~1228 SP, not the full 1536 (where it only happens in a few cases, ie. SHA256/bitcoin). With GCN 1D shader, the performance relative is 2048 / ~1228. The slide clearly demonstrates this, 50% advantage in SHA256 but a lot more with other instruction sets. Just to infer from that, it means the 2048 1D shader will operate at 100% efficiency. So in gaming performance, 2048 1D shaders perform comparable to ~2560 VLIW4/5.

Further inference: 2560 / 1536 = 66% more "effective" SP performance in games. If drivers are good, it could very well be close to 6990 levels.

 

[apoppin]

I often check benches on your site. But could you please add more newer games to your bench list. There's a lot of older games where their popularity has waned out long ago.

Definitely add BF3, Skyrim, Anno 2070 and even Starcraft to your list, all very popular games that heaps of gamers play.

Edit: You know its a paper launch when Napoleon @ Chiphell doesn't receive a sample for review. heh

Thanks. i dumped a few older games and added a bunch of new games recently.

Starcraft is a very difficult game to bench where it is representative of an actual game. And it takes me quite awhile to create a representative benchmark - i usually play the game from start to finish. Eventually, i will add Skyrim

Here is my shortened current benching list; in bold is brand-new:

• Vantage
• 3DMark 11
• Heaven 2.5
• Wolfenstein
• Batman: Arkham City
• Left4Dead 2
  
• Serious Sam, Second Encounter HD (2010)
• Mafia II
• Crysis
• Far Cry 2
• Just Cause 2
• World-in-Conflict
• Resident Evil 5
• BattleForge
• Alien vs. Predator
• STALKER, Call of Pripyat
• Metro 2033
• F1 2010
• H.A.W.X. 2
• Lost Planet 2
• Civilization V
• Crysis 2
• Total War: Shogun II
• Dirt 3
• Deus Ex: Human Revolution
• BF3

i have no axe to grind against AMD for not sending me a HD 7970. Either way, it is traffic for my site. i wish them well in their new strategy with HD 7970. Their CEO appears to be pretty savvy with the launch timing (moving it up a week to take advantage of CES and world attention).

And we are "equal opportunity" at ABT. If Nvidia doesn't send me a GTX 680, we will leak the specs .. we are a Media site, after all ... NEWS; and what i publish as speculation is identified as such.
:biggrin:

 

[toyota]

apoppin, quit posting and get that preview of yours up.

 

[Silverforce11]

Starcraft is a very difficult game to bench where it is representative of an actual game. And it takes me quite awhile to create a representative benchmark - i usually play the game from start to finish. Eventually, i will add Skyrim

Here is my shortened current benching list:

That's not short.

Starcraft 2, download a replay with lots of action. Load replay, win.

You can probably remove older games like Far Cry 2 and a few others off your list to make room for newer games. Id hesitate to remove Valve games since so many people play their games on the same engine. But stuff like LP2 and Hawx2, do ppl even play that these days in any meaningful numbers?

 

[Zed03]

apoppin, quit posting and get that preview of yours up.

he didn't get a 7970.. therefore no review?