AMD HD7*** series info

[AtenRa]

If you clicked my links, most of the benches I linked were actually 1920x1080, not 2560x1600. Either way, to show true representation of performance difference between a modern generation a new one, you should be testing them in high resolution, not in "all resolutions" averaged as you keep linking from TPU. Both 4870 and 5870 were intended for 1920x1080 at launch since they were high end cards.



That's because TPU loves to include older games in their reviews - and we didn't have a ton of modern games when they did that initial 5870 review. As more modern games came out, HD5870 was really able to show its prowess over the slower 4870. And that's the point: AMD designed HD5870 to perform well in more modern games, not in older games.

If you look at a lot of the games they tested, an HD5870 would not be able to show its true performance since the games they tested were not intensive / ancient that reduced the true performance delta:

- UT3
- Dawn of War 2
- Call of Juarez 2
- ET:QW
- Far Cry 1
- Quake 4
- Prey
- Stalker 1
- TF2

^ Who upgraded from an HD4870 $130 to a $350 HD5870 to get more performance in these games? TPU's decision to include these old games is diminishing the performance advantage that HD5870 has.

Not to mention they include 3dMark as part of their averages and those aren't games.....

If you tested a bunch of old games with minimal AA at lower resolutions today with GTX580 vs. GTX460, you also will see much lower performance delta than if you were testing modern games for which GTX580 was intended in the first place.

The last graph i posted is only for 1920x1200 and shows HD5870 only 43% faster than HD4890.

From TPU

Each benchmark was tested at the following settings and resolution:

1024 x 768, No Anti-aliasing. This is a standard resolution without demanding display settings.
1280 x 1024, 2x Anti-aliasing. Common resolution for most smaller flatscreens today (17" - 19"). A bit of eye candy turned on in the drivers.
1680 x 1050, 4x Anti-aliasing. Most common widescreen resolution on larger displays (19" - 22"). Very good looking driver graphics settings.
1920 x 1200, 4x Anti-aliasing. Typical widescreen resolution for large displays (22" - 26"). Very good looking driver graphics settings.
2560 x 1600, 4x Anti-aliasing. Highest possible resolution for commonly available displays (30"). Very good looking driver graphics settings.

At the time HD5870 was released (22/23 September 2009), most of the games in TPUs review where the latest or they had the best IQ and features (DX-9/10). It is only logical (heh Spock speaking ) they benched those games with HD5870, no DX-11 game was available at the time.

Again, at the initial release date, HD5870 was not more than 40-50% faster than HD4890 at 1920x1080/1200.

On more thing, modern games (2010-2011) were coded for AMDs HD58xx and HD69xx architectures and they better have more than 50% performance than HD48xx series

 

[96Firebird]

All this talk about the new 28nm makes me want to apply for a position at the new GloFo Fab 8 facility in NY. Brand new facility, the only thing I wouldn't like is wearing the clean room attire. 7 recent graduates from the college I just graduated from already have a job there...

http://fab8update.com/

 

[Silverforce11]

The 5870 vs 4870 debate is stupid, because the 5870 was designed for dx11, to measure its performance gains using old games or all resolutions is irrelevant. The fact that it can deliver >75% gains in modern non-dx11 games shows its strength.

But realistically, Tahiti is not going to be anywhere near >75% gains over Cayman because the TDP is 190W. If the TDP was ~225W, then we can be more confident in expecting ~75% gains. A die shrink is good, but its not magically going to improve your perf/watt by ~170%.

But then again, maybe AMD decides to do TDP like NV and under-report the average load rather than max load. Who knows. The fun comes when AIBs get their hands on the cards.

 

[RussianSensation]

But realistically, Tahiti is not going to be anywhere near >75% gains over Cayman because the TDP is 190W. If the TDP was ~225W, then we can be more confident in expecting ~75% gains. A die shrink is good, but its not magically going to improve your perf/watt by ~170%.

Imagine AMD cooks up a 1-2 BIOS switch that allows you to raise TDP from 190 to 250W and those 2048 ALUs turn into 2560 ALUs @ 1ghz? :awe:

Or maybe they'll launch HD7970 initially at 190W and leave headroom for 7980 250W part in case Kepler GTX680 smacks them. That would be a cool strategy: leave headroom to respond and to appear to NV as if your 2048 part is all you could squeeze, while in the short-term making nice margins on higher yielding 2048 ALU chips. And then have 2560 chip ready in the wings to counter and beat GTX680!

 

[Silverforce11]

Well if the NV delay rumors are true, AMD is going to have ~6 months lead this time. They can launch a "refresh" 6 months down the road if they need to combat kepler.

 

[Grooveriding]

Imagine AMD cooks up a 1-2 BIOS switch that allows you to raise TDP from 190 to 250W and those 2048 ALUs turn into 2560 ALUs @ 1ghz? :awe:

I feel there is indeed something AUSUM that will be released...

 

[tviceman]

Well if the NV delay rumors are true, AMD is going to have ~6 months lead this time. They can launch a "refresh" 6 months down the road if they need to combat kepler.

Want to place a friendly bet that AMD's 28nm high end parts will be no more than two months ahead of Nvidia's 28nm parts? I'll bet you a $20 steam game on it....

 

[Silverforce11]

I feel there is indeed something AUSUM that will be released...

Making a huge die, cripple it deliberately so it performs at 75% of potential.. flick a switch and it goes 100%.. NOT LIKELY to happen under any scenario.

There is a more likely scenario:
Make a medium die, cripple it with core/ram clock to lower TDP, flick a switch and it gets TURBOBOOSTtm by 25% in clock speeds and reach a much higher TDP.

But either is way too far fetched atm.

 

[Silverforce11]

Want to place a friendly bet that AMD's 28nm high end parts will be no more than two months ahead of Nvidia's 28nm parts? I'll bet you a $20 steam game on it....

With volume or do paper launches count?

 

[tviceman]

With volume or do paper launches count?

Good question. I'd like to base it on initial product availability. I would like to wager a $20 steam game with you that Nvidia's next gen high end card will be in-stock available on the market to purchase within 2 months (60 days) from the first day AMD cards are in-stock available to purchase.

You're saying Nvidia will be six months behind

Well if the NV delay rumors are true, AMD is going to have ~6 months lead this time.

So this really should be a no-brainer bet for you if you truly believe what you say.

P.S. I'll go for a more expensive game if you want, but I didn't want to scare you off from taking the bet.

 

[WMD]

Making a huge die, cripple it deliberately so it performs at 75% of potential.. flick a switch and it goes 100%.. NOT LIKELY to happen under any scenario.

There is a more likely scenario:
Make a medium die, cripple it with core/ram clock to lower TDP, flick a switch and it gets TURBOBOOSTtm by 25% in clock speeds and reach a much higher TDP.

But either is way too far fetched atm.

The latter is more likely but it won't be in the form of a switch. More likely they will release a new card with increased clocks and voltage + better cooling. Much like 4870 - 4890 or 5850 - 6870.

 

[happy medium]

Nvidia's next gen high end card will be in-stock available on the market to purchase within 2 months (60 days) from the first day AMD cards are in-stock available to purchase.

Any 7000 series AMD card? or just the higher end?

 

[RussianSensation]

Good question. I'd like to base it on initial product availability. I would like to wager a $20 steam game with you that Nvidia's next gen high end card will be in-stock available on the market to purchase within 2 months (60 days) from the first day AMD cards are in-stock available to purchase.

You should bet him High-end HD7970 vs. High-end GTX680. AMD may launch HD76/78xx derivatives this year and post-pone HD79xx series into Q1 2012. So if you bet "any" HD7000 series, then you are going to lose big time.

 

[AtenRa]

The 5870 vs 4870 debate is stupid, because the 5870 was designed for dx11, to measure its performance gains using old games or all resolutions is irrelevant. The fact that it can deliver >75% gains in modern non-dx11 games shows its strength.

But realistically, Tahiti is not going to be anywhere near >75% gains over Cayman because the TDP is 190W. If the TDP was ~225W, then we can be more confident in expecting ~75% gains. A die shrink is good, but its not magically going to improve your perf/watt by ~170%.

But then again, maybe AMD decides to do TDP like NV and under-report the average load rather than max load. Who knows. The fun comes when AIBs get their hands on the cards.

Sorry but no, Evergreen (HD5870) was not designed for DX-11, ATI at the time, continued from HD4890 by just doubling almost everything (kept VLIW5) and they just added a single Tessellator unit for DX-11.

The Evergreen design was successful because it gave a big fps boost in all DX-9 and DX-10 games at the time of release over previous generation video cards.

On the other hand, Fermi and Cayman were designed with DX-11 Tessellation in mind by having a multi-core (tessellator units) Architecture design.

AMD's HD7970 is a GCN multi-core (Tessellator Units) architecture design, focused on GPGPU and DX-11 Tessellation. I suspect that in DX-11 Tessellation Games it could have more than 50% performance than Cayman (HD6970), it could even reach close to 100%.

The more Tessellation the more the performance advantage of GCN architecture over Evergreen/Cayman but here comes another problem, current Game Evolved games don't implement a lot of Tessellation and performance in those games will not scale more than 40-50%. On the other hand, HD7970 could be faster that 70-100% in TWIMTBP DX-11 Games against Cayman due to heavy Tessellation implemented in those games.

I could be wrong

 

[Silverforce11]

You should bet him High-end HD7970 vs. High-end GTX680. AMD may launch HD76/78xx derivatives this year and post-pone HD79xx series into Q1 2012. So if you bet "any" HD7000 series, then you are going to lose big time.

I'm not a gambling man, regardless of the easy win.

NV's tapeout is ~5 months behind AMDs, even if everything works out, they will be very late.

 

[Silverforce11]

Sorry but no, Evergreen (HD5870) was not designed for DX-11, ATI at the time, continued from HD4890 by just doubling almost everything (kept VLIW5) and they just added a single Tessellator unit for DX-11.

There's more to DX11 than just tessellation.

If you claim recent games are better coded for the newer card, i can claim the reverse. That on release, the new card was not optimized to run on the older games as well. Which is true. After several driver updates, the 5870 series has increase its lead over the 4870.

IF the 5870 was just a pure "double everything" of the 4870 without supporting new dx11 features, you could expect almost doubling of performance. But it wasn't.

 

[3DVagabond]

Another reason TDP is low could be that HPL is clock limited. It's likely that at the "sweet spot" size running @ ~1GHz a GCN card will only draw 190W more or less by default. Once HP comes out and clock speeds can be increased we're more likely to see cards with higher TDP. The "sweet spot" there is 225W. If they actually go with this XDR2 RAM that reduces TDP as well.

 

[AtenRa]

IF the 5870 was just a pure "double everything" of the 4870 without supporting new dx11 features, you could expect almost doubling of performance. But it wasn't.

You don't get double the performance with double the transistor count (Pollack's Rule), not even with lower lithographic process.

Evergreen really was almost double the HD4890 in SP count, TMU's, ROPs, same VLIW5 architecture, with a single Tessellator engine and DX-11 support.

It was faster than NVs high end single die video cards at DX-9/10 games, at lower power usage and that made it a huge success, not the DX-11 and Tessellation support.

 

[AtenRa]

Another reason TDP is low could be that HPL is clock limited. It's likely that at the "sweet spot" size running @ ~1GHz a GCN card will only draw 190W more or less by default. Once HP comes out and clock speeds can be increased we're more likely to see cards with higher TDP. The "sweet spot" there is 225W. If they actually go with this XDR2 RAM that reduces TDP as well.

TSMC's HPL is not suitable for more than 40-50W chips, so i find it very doubtful AMD will use HPL for Desktop chips not to mention High End.

 

[WMD]

5870 is double of a 4890 except for memory bandwidth and geometry throughput. It can never reach 2x 4890 performance. But Russian's estimate of 75%-100% increase from 4870 is right. On my rig, Resident Evil 5 benchmark went from 78fps on a 4870 oc to 153fps on a 5850 oc. Metro 2033 also saw close to double performance.
Earlier reviews shows lesser gains due to older games and slower cpu bottlenecking the cards.

 

[Silverforce11]

5870 is double of a 4890 except for memory bandwidth and geometry throughput. It can never reach 2x 4890 performance. But Russian's estimate of 75%-100% increase from 4870 is right. On my rig, Resident Evil 5 benchmark went from 78fps on a 4870 oc to 153fps on a 5850 oc. Metro 2033 also saw close to double performance.
Earlier reviews shows lesser gains due to older games and slower cpu bottlenecking the cards.

He's right, the difference is near double in newer games and as drivers mature. I know, because my old GPU was a 4870, i still have it laying around collecting dust.

 

[tviceman]

You should bet him High-end HD7970 vs. High-end GTX680. AMD may launch HD76/78xx derivatives this year and post-pone HD79xx series into Q1 2012. So if you bet "any" HD7000 series, then you are going to lose big time.

Yeah I didn't word my statement properly; that is exactly what I meant. I should have said Nvidia's next gen high end available to buy in 60 days or sooner from the point AMD's next gen high end is available to buy.

I'm not a gambling man, regardless of the easy win.

NV's tapeout is ~5 months behind AMDs, even if everything works out, they will be very late.

So I take it no then? You don't want to put your money where your mouth is? You sure do talk a big game and don't like to back it up! That's okay though, with all the racket you stirred up about the 6900 series - and were ultimately wrong about - I wouldn't be taking bets (no matter how good they seemed) if I were you either!

 

[AtenRa]

Problem is that Russian says that 79xx series will be (or must be) 75% faster that 69xx series and not against 58xx.

My argument is that 79xx will not be more than 50% faster than 69xx when it will be released, except perhaps, as i have said before in DX-11 tessellation.

 

[WMD]

Problem is that Russian says that 79xx series will be (or must be) 75% faster that 69xx series and not against 58xx.

My argument is that 79xx will not be more than 50% faster than 69xx when it will be released, except perhaps, as i have said before in DX-11 tessellation.

I think their performance target wont be based on their previous generation at all but at NV's current offerings. 2 possibilities are likely for their high end single GPU:

a) Barely outperforming a GTX580 by 5-10% and priced just under it. Much lower power.

b) Outperforming a GTX 580 by ~30% and similarly priced.

This has always been AMD's practice in the past. 100% increase in performance at dirt cheap price (4870) doesn't happen unless they are far behind their competition.

 

[RussianSensation]

Problem is that Russian says that 79xx series will be (or must be) 75% faster that 69xx series and not against 58xx.

My argument is that 79xx will not be more than 50% faster than 69xx when it will be released, except perhaps, as i have said before in DX-11 tessellation.

I said 75% because I expect Kepler to be fast, very fast. Also, historically every 18-24 months, AMD increased performance by 75-100%, except 9700Pro-->9800Pro (refresh really) and X1950XTX --> 2900 series.

But otherwise, 75-100% is actually the norm for them in the last 10 years.

Considering HD6950 ~ HD5870 = 100%, HD6970 = 112%, GTX580 = 121% @ 1920x1080 4AA/16AF, if HD7970 is only 75% faster than HD5870, you'll have:

HD5870 / 6950 = 1.0
HD6970 = 1.12
GTX580 = 1.21
HD7970 = 1.75

So HD7970 would be 1.75 / 1.21 = 45% faster than GTX580.

What are the chances GTX680 is only 45% faster than a GTX580? AMD needs more than 75% speed increase unless they plan to compete on price again. Or it will be again a situation where GTX670 competes with HD7970 at $350 and GTX680 takes $500 price point.