Leaked ATI S.I. 6870 benchmark

[Grooveriding]

Agreed that 5870 overall > 470. However, the summary you provided was not for DX11 games was it? Plus what's the point of adding all the resolutions into your tally? How is that even relevant?

Add these games and see how 5870 fares: Metro 2033, STALKER:CoP (sunshafts), Lost Planet 2, AvP, BattleForge, Just Cause 2. 5870 is almost $90 -100 more expensive and can't really beat 470 in any of these games.

Plus your point was related to superior texture performance of 4870 vs. 3870. However, even from the benches you linked, 5870's double the texture throughput is hardly an advantage compared to 470/480, unless we are talking about 2560x1600. Then where did this advantage disappear to? Well compare 470 to GTX285 in older OpenGL games (mostly texture limited like Quake 4, Wolfenstein, etc.) and you'll see that the 285 is at least as fast if not faster!!! In current games, 285 has no chance because they are shader limited, where texture advantage is less relevant.

"If" 5870 was less than $300, it would be recommended :thumbsup:



I don't recall a single person in this entire thread disputing 30% performance improvement for HD6000 series as a possibility. The discussion is mainly revolving whether or not HD6770 will = 5870 (i.e., 100% performance improvement from HD5770).

Don't be surprised if 6770 is as fast or very near to a 5870. We'll find out when its launched next month.

To make sense it will have to be faster than the GTX 460 1GB, or else it would not be worth releasing. So at a minimum as fast as a 5850 is. The gap between the 5850 and 5870 is not that vast, so if it can pull ahead by even 10% of a 5850, it will basically be as fast as a 5870. With rumors pointing to it being 2/3 of what the 6870 is going to be that would put it around 5870 performance with the 6870 looking to be about 35% faster than a GTX 480.

AMD knows the only bracket they are getting challenged at right now is the mid-range. They also know the only card NV has right now that has any sort of curb appeal is the 460. The 6770 is going to be directly targeted at the 460.

The 6 series is AMD's brand new architecture, it's 2nd generation DX11 against 1st generation DX11 from NV. Expect every 6 series Radeon released in the main brackets, ie: mid-range 6770, single gpu leader 6870 and dual gpu performance crown 6970 to be faster than their NV counterparts, 460 will be slower than 6770, 480 will be slower than 6870 and nothing at the top end I guess as NV still does not have a halo part that has the performance crown, so the 6970 will just take over where the 5970 has left off.

This should come as no surprise. Making the process argument that because 5870 was on 40nm and 6870 is on the same process there can't be a significant performance jump is inaccurate. There are precedents for this; 2900 to 3870. 9800GTX to GTX 280. Those were all done on the same process and gave performance leaps of 40-60%.

There has never been a launch under a moniker of the next generation that has not given significant gains in performance.

We'll see bigger dies consuming more power and putting out more heat I bet. And they'll be performance leaders in their respective brackets, likely until NV manages to get out Fermi II on 28nm late next year. I can't see NV doing anything but a refresh at this point, if they were going to do a completely new generation on 40nm like AMD, by the time they could get it out there, 28nm would be ready. NV is going to be pulling a turtle until they release a new generation on 28nm next year.

 

[Daedalus685]

There are precedents for this; 2900 to 3870. 9800GTX to GTX 280.

The 2900xt was on 80nm, the 3870 on 55nm which was a pretty huge shrink. The large jump on the same node was the 3870 - 4870.

 

[RussianSensation]

I will add to Daedalus685's point regarding GTX280 vs. 9800GTX. The comparison is invalid since 9800GTX is simply a refresh of G80 architecture, not a new generation. The GTX 280 is the first real reworking of the G8x architecture, rather than G92. The previous generation 8800GTX which was in fact only on 90nm with the same 128 shaders as the 9800GTX. The 9800GTX is then just a 65nm die-shrunk G80 90nm with some improvements to the GPU core. No wonder that 9800GTX is very similar in performance to the 8800GTX.

GTX280 consisted of 1.4 billion transistors covering a 576mm2 die surface area built on a 65nm process. The 9800GTX does not differ much from 8800 GTS 512, just overclocked. The core remains the same (G92), just a new revision, with transistors at around 700 million covering a 334mm2 die surface area.

This easily explains why GTX280 was such a major performance jump from G80/G92. GT200 is 72% larger in size than 9800GTX is!!! http://hardforum.com/showthread.php?t=1325165.

So unless HD6770's die suddenly 50-70% larger in size compared to the HD5770, it's probably unrealistic to think that it will equal the HD5870 at $200!

 

[Skurge]

AMD has done this kind of voodoo before. The 4870 is one example, but the X1900XTX was another, it was on the same 90nm process as the X1800XT and had the same architecture, but I would be anything from 20 to 70% faster.

 

[RussianSensation]

AMD has done this kind of voodoo before. The 4870 is one example, but the X1900XTX was another, it was on the same 90nm process as the X1800XT and had the same architecture, but I would be anything from 20 to 70% faster.

It is never any voodoo

Like I said, without a process node change, ATI is unlikely to be anywhere near 100% performance increase unless they increase the die size beyond 400mm2.

X1900XT to X1800XT was almost a 20% increase in die size on the same 90nm process node! The performance increase was only about 30-35% for the most part.

If we apply that, we would get a 400mm2 high end HD6870/90 series with about 35% faster performance than HD6870. So how in the world is HD6770 going to be as fast as 5870 then?

There was no vodoo either with the 4870 vs. 3870. The die size increased a whooping 37% from 190mm2 to 260mm2.

All I am saying is most performance improvements in the past have all come from either a serious die size increase, or a node shrink, or both (which allowed more transistors /shaders/ ROPs, etc. to fit into the same space).

This is why as SilverForce11 noted, if ATI intends more than 50% performance increase, they are going to increase the die size from 334mm2.

We must not forget that 5870 increased in die size 28% relative to 4870 as well (334mm2 vs. 260mm2).

To conclude, we thus far had GPU die increases from 3870 @ 55nm 190mm2 --> 4870 @ 55nm 260mm2 --> 5870 @ 40nm 334mm2. In each of these cases, the transistor count also increased due to a GPU die increase and/or a node shrink. This mostly explains why we saw such massive performance increases in these generations.

This is why I am more inclined to believe that HD6000 will be 50% faster, not 100%. I'll be thrilled if I am wrong!!!

 

[Grooveriding]

Like I said, without a process node change, ATI is unlikely to be anywhere near 100% performance increase unless they increase the die size beyond 400mm2.

X1900XT to X1800XT was almost a 20% increase in die size on the same 90nm process node! The performance increase was only about 30-35% for the most part.

If we apply that, we would get a 400mm2 high end HD6870/90 series with about 35% faster performance than HD6870. So how in the world is HD6770 going to be as fast as 5870 then?

There was no vodoo either with the 4870 vs. 3870. The die size increased a whooping 37% from 190mm2 to 260mm2.

All I am saying is most performance improvements in the past have all come from either a serious die size increase, or a node shrink (which allowed more transistors /shaders/ ROPs, etc. to fit into the same space).

This is why as SilverForce11 noted, it makes sense that if ATI intends more than 50% performance increase, they are going to increase the die size from 334mm2.

It's not a 100% performance jump to put the 6770 at 5870 performance levels though.

They are treating the 6770 differently than they did the 5770. Using a scaled back 6870 core, putting it on a 256 bit bus. This is a different approach than what the 5770 was to the 5870.

There are no claims that 6870 will twice as fast as a 5870, probably 150% as fast. But 6770 being treated more the way the 5850 was will change the amount of performance it delivers. I don't think it will cost $200, probably more, especially if hits the 5870 level of performance. But even at $250 , that will still be an amazing amount of performance for the price.

 

[RussianSensation]

It's not a 100% performance jump to put the 6770 at 5870 performance levels though.

They are treating the 6770 differently than they did the 5770. Using a scaled back 6870 core, putting it on a 256 bit bus.

It would be 100% increase from 5770 though. Do we have any reliable source which showed that HD6770 is going to be treated differently than 5770 was?

Again, I think there is a confusion between a chip codenamed 6770 (which is rumored to be a HD6870 card) and HD6770.

 

[Skurge]

You keep overlooking the fact that the architecture is different. We already know the shaders are different, who knows what they have done to the rest of the chip, the ROPS or even clocks.

If AMD can sell a 334mm2 5830 at $190, surely the can make the make the 6770 about 300mm2 at $200. Im sure at 300mm2 with improvements all over the core, it wouldn't be hard to match a 5870. That would also fit rumors saying barts is 1/3 of caymen if caymen is 400mm2 and 35% faster than 5870 unlike a 5770 that is EXACTLY half a 5870 even down to the die size.

You also make it sound like they are building from a 5770. The 6770 is whole new architecture, so getting 100% performance improvement from a 5770 isn't so far fetched.

 

[RussianSensation]

That would also fit rumors saying barts is 1/3 of caymen if caymen is 400mm2 and 35% faster than 5870 unlike a 5770 that is EXACTLY half a 5870 even down to the die size.

Again, you and Grooveriding are using Barts XT interchangeably with HD6770 videocard. Barts XT chip is rumored to be called HD6870 videocard, not HD6770 . While Barts XT (i.e, 6870) may in fact be 2/3 of Cayman XT (i.e., HD6970), this tells us nothing about HD6770, the replacement for HD5770.

This is what I've been trying to communicate for days now. HD6770 videocard and a chip codenamed 6770 (i.e., Barts XT) may not be the same videocard. Does this make sense now?

If HD6770 videocard is a rebadge of the HD5770 with tweaked clock speeds, it is no laughing matter because that will suck!

 

[Skurge]

Again, you and Grooveriding are using Barts XT interchangeably with HD6770 videocard. Barts XT chip is rumored to be called HD6870 videocard, not HD6770 . While Barts XT (i.e, 6870) may in fact be 2/3 of Cayman XT (i.e., HD6970), this tells us nothing about HD6770, the replacement for HD5770.

This is what I've been trying to communicate for days now. HD6770 videocard and a chip codenamed 6770 (i.e., Barts XT) may not be the same videocard. Does this make sense now?

OKay, so we'll just call it Barts XT than. My point still stands though, Barts could easily be just as fast as a 5870.

 

[RussianSensation]

OKay, so we'll just call it Barts XT than. My point still stands though, Barts could easily be just as fast as a 5870.

How can your point stand? Barts XT does not appear to be a mid-range videocard though. Grooveriding always keeps talking about how "6770 with 256-bit bus" is launching next month. However, this card is the 6870 and has nothing to do with HD6770 which IS launching in October.

I must be confused myself because I thought we are on the same page

 

[Skurge]

How can your point stand? Barts XT does not appear to be a mid-range videocard though. Grooveriding always keeps talking about how "6770 with 256-bit bus" is launching next month. However, this card is the 6870 and has nothing to do with HD6770 which IS launching in October.

I must be confused myself because I thought we are on the same page

Huh, what?

So what was the discussion again?

 

[RussianSensation]

Huh, what?

So what was the discussion again?

You and Groove were talking about HD6770 = 5870. I said no. What in fact both of you were talking about was Barts XT being as fast as the 5870 (which is to say HD6870 = HD5870). I wouldn't disagree with that statement. :thumbsup:

 

[Skurge]

You and Groove were talking about HD6770 = 5870. I said no. What in fact both of you were talking about was Barts XT being as fast as the 5870 (which is to say HD6870 = HD5870). I wouldn't disagree with that statement. :thumbsup:

Ahhh, So we were right... and wrong at the same time. lol

 

[GaiaHunter]

You and Groove were talking about HD6770 = 5870. I said no. What in fact both of you were talking about was Barts XT being as fast as the 5870 (which is to say HD6870 = HD5870). I wouldn't disagree with that statement. :thumbsup:

http://www.nordichardware.com/news/...ts-launches-in-october-as-radeon-hd-6800.html

According to nordich, the 6800 is the new midrange and so it is replacing the 5700 series, or if you prefer the Barts XT and Pro are the replacements of the Juniper XT and Pro.

The 6900 is the new performance and direct replacement of 5800 series or if you prefer Cayman XT and Pro replace Cypress XT and Pro.

The 6990, Antilles is the dual card and replacement of Hemlock (5970). If there is an Antilles XT and an Antilles Pro it is going to be interest what each will be called (6990 and 6995 or 6980 and 6990?).

So Barts XT should be SMALLER than Cypress but about the same speed as a 5850.

So, yes, what Groove and Skurge were talking about was the 6870 (which was previously supposed to be the 6770) but their argumentation about size and performance were correct if the rumours are true.

 

[Grooveriding]

How can your point stand? Barts XT does not appear to be a mid-range videocard though. Grooveriding always keeps talking about how "6770 with 256-bit bus" is launching next month. However, this card is the 6870 and has nothing to do with HD6770 which IS launching in October.

I must be confused myself because I thought we are on the same page

Yeah I am totally referring to Barts XT

The only thing I've seen referencing that whole new naming nomenclature is a post by someone on Chiphell. I just ignore it, makes it too messy. We'll know the real naming scheme next month with the first 6 series launch.

I can't see them launching a rebadged 5770 as 6770 next month. There are multiple references even here in recent Anand reviews stating they are launching Northern Islands this year.

I doubt they will start the launch with a renamed 5770. My thought is that it is Barts XT we are getting next month. Only a month till we find out!

 

[RussianSensation]

GaiaHunter cleared this up for me better than ever! Thanks man

Yup, now I can definitely see how you guys were saying that 6870 will slot between 5850 and 5870. Somewhat misleading naming convention from AMD though since 6870 would normally imply that it's faster than the 5870 hehe

 

[bryanW1995]

It would be 100% increase from 5770 though. Do we have any reliable source which showed that HD6770 is going to be treated differently than 5770 was?

Again, I think there is a confusion between a chip codenamed 6770 (which is rumored to be a HD6870 card) and HD6770.

remember though that 5770 was 1/2 of 5870, while barts xt (6770 or 6870 or whatever they end up calling it) is 2/3 of cayman xt and both are on 256bit bus.

 

[PureHazard]

http://www.nordichardware.com/news/...ts-launches-in-october-as-radeon-hd-6800.html

According to nordich, the 6800 is the new midrange and so it is replacing the 5700 series, or if you prefer the Barts XT and Pro are the replacements of the Juniper XT and Pro.

The 6900 is the new performance and direct replacement of 5800 series or if you prefer Cayman XT and Pro replace Cypress XT and Pro.

The 6990, Antilles is the dual card and replacement of Hemlock (5970). If there is an Antilles XT and an Antilles Pro it is going to be interest what each will be called (6990 and 6995 or 6980 and 6990?).

So Barts XT should be SMALLER than Cypress but about the same speed as a 5850.

So, yes, what Groove and Skurge were talking about was the 6870 (which was previously supposed to be the 6770) but their argumentation about size and performance were correct if the rumours are true.

This new naming scheme is unfortunately misleading.

Hopefully, this will drive the prices of the 5850s down to under $200 so I can grab another for Crossfire. That is, unless the performance 6900 series (6950?) will give me roughly the same 5850 Crossfire performance for $350 with lower power usage.

 

[Silverforce11]

Its a matter of sales department ppl deciding on the naming scheme. I think its a stupid idea to even name cards based on a mid-range GPU 68xx, it breaks the mold. Barts XT is only ~=5850 perf-wise, to name it 6870 is very out of place. Whatever name they go with will be final because packaging, manuals and drivers are going to have to be set a month prior to release.

 

[RussianSensation]

I think its a stupid idea to even name cards based on a mid-range GPU 68xx, it breaks the mold.

I agree too. Also with 6870, 6890, 6950 and 6970 probably occupied, what are they going to call their refresh line? Or I guess the new trend is no refreshes at all and just new generations every 12-15 months?

 

[Lonbjerg]

Most people will remember the points in time that Intel and AMD traded blows for the performance crown. Back and forth they went for some time until the point in time when Intel took the crown and hasn't looked back ever sense.

You memory must be bad, as they only traded blows once, just around the "NetBurst" era, all the time before that and all the time after after that AMD has always been subpar to Intel in performance.

 

[Timorous]

http://www.nordichardware.com/news/...ts-launches-in-october-as-radeon-hd-6800.html

According to nordich, the 6800 is the new midrange and so it is replacing the 5700 series, or if you prefer the Barts XT and Pro are the replacements of the Juniper XT and Pro.

The 6900 is the new performance and direct replacement of 5800 series or if you prefer Cayman XT and Pro replace Cypress XT and Pro.

The 6990, Antilles is the dual card and replacement of Hemlock (5970). If there is an Antilles XT and an Antilles Pro it is going to be interest what each will be called (6990 and 6995 or 6980 and 6990?).

So Barts XT should be SMALLER than Cypress but about the same speed as a 5850.

So, yes, what Groove and Skurge were talking about was the 6870 (which was previously supposed to be the 6770) but their argumentation about size and performance were correct if the rumours are true.

The problem is this rumour is stupid. If barts XT is about the same performance as a 5850 and they are going to call it the 6870 dont you see a massive flaw with that argument. There is no way that AMD's new 68xx generation is going to be slower than the 58xx generation.

Here is my view based on past launches and the current rumours.

Antilles, 2x cayman xt, 6970 --> faster than anything out at the moment

Cayman xt, 6870 --> about 25% faster than GTX480
Cayman pro, 6850 --> about GTX480 level

Barts xt, 6770 --> about 5850 level
Barts pro, 6750 --> about GTX460 level

Earlier I suggested barts was about 190-200mm^2 but after looking at the rumours again this is probably incorrect and I think a die size of about 240mm^2 is more likely. At this die size AMD will likely charge a bit less for a 6770 vs a 5850 to increase perf/£ but at the same time they can increase their own margins.

The only rebadge I can see is possibly 5770 --> 6670 but if AMD can make a smaller die than juniper at that performance level they will probably go that route instead.

 

[Caribbean Geek]

There was no vodoo either with the 4870 vs. 3870. The die size increased a whooping 37% from 190mm2 to 260mm2.

All I am saying is most performance improvements in the past have all come from either a serious die size increase, or a node shrink, or both (which allowed more transistors /shaders/ ROPs, etc. to fit into the same space).

This is why as SilverForce11 noted, if ATI intends more than 50% performance increase, they are going to increase the die size from 334mm2.

We must not forget that 5870 increased in die size 28% relative to 4870 as well (334mm2 vs. 260mm2).

To conclude, we thus far had GPU die increases from 3870 @ 55nm 190mm2 --> 4870 @ 55nm 260mm2 --> 5870 @ 40nm 334mm2. In each of these cases, the transistor count also increased due to a GPU die increase and/or a node shrink. This mostly explains why we saw such massive performance increases in these generations.

This is why I am more inclined to believe that HD6000 will be 50% faster, not 100%. I'll be thrilled if I am wrong!!!

The HD 4870 die size is 33% larger than the HD 3870 and offered 2.5 times of everything at the same node process, which explains why it was often two times faster easily than the HD 3870, but the HD 5870 had only 2 times of everything because the performance increase never scales linearly with hardware. The HD 5870 had a slightly bigger die size than the HD 4870 and yet, it was near twice faster, never was able to match the HD 4870X2 in quite many scenarios (HD 5x00 shaders are less efficient compared to HD 4x00 shaders), but if the HD 6970 becomes 33% larger than the HD 5870 because its stream processors are smaller, its possible to have easily a performance boost of 50% with little effort, and if its true that the HD 6x00 series has faster stream processors and better IPC, that can boost the performance up to 70% in average, may be more depending of the core clock configuration.

 

[Caribbean Geek]

I want a shot.

6670 released Oct 24th 5% faster then a gtx 460
6850 released Dec.1st ? ~ = to a gtx475
6870 released Dec.1st ? 5% faster then a gtx 480 512sp
dual 6770 card released Dec.1st ? = to a gtx495

The GTX 480 512 in a preview benchmark showed only 5% faster performance in overall, 32 more stream processors and an additional TMU will not make a big difference at all, I think that a full GF104 chip with faster core shows more promises. Since the HD 6870 appears to be Barts XT, I would be surprised if it reached GTX 480 performance which I doubt seriously, I think its inline with the HD 5850.