Leaked ATI S.I. 6870 benchmark

[dug777]

*slow clap*

Excellent write up, that was largely my understanding of the redesign as well.

Amen to that, nice work Temp

 

[Grooveriding]

***.

Thanks for consolidating all that information.

I thought it would stand to reason to most folks, that considering we are now a year past 5XXX series release and this series is being labeled 6XXX, that we are going to be getting something justifying that moniker.

Not exactly scientific, but good common sense reasoning. That info you've so nicely put together helps to lend credence to this.

We're getting a new series from ATI here, it's been a year since their last. This is not a refresh, it's a new architecture.

This is going to be the same situation as when 5XXX series released and was the best performance you could buy for seven months, there is no guarantee of the the same performance jump percentile wise, but we're getting a new level of performance soon from ATI.

Like it or not, we can expect big changes and significantly more performance from ATI's new 6XXX series cards than anything currently available on the market. :thumbsup::thumbsup::thumbsup:

 

[evolucion8]

Great post Tempered81, so in a nutshell, AMD's approach means that the die is doing more work per mm2 than nVidia's approach. HD 5870 basically has 320 clusters of processors in which each one can execute up to 5 instructions in the best case scenario and much smaller die size competing against for example, the GTX 470 which has 448 processors which can only execute one instruction per CUDA core, and a much bigger die size.

 

[ZimZum]

Amen to that, nice work Temp

+1

 

[tweakboy]

Those are sick Crysis benchmarks. 43fps avg at that res and 4x AA wow.

I hope nVidia can attack back with their 5xx line. But I think nVidia lost for this year. All they will release from now and next year is a 490 GTX card. In this same time AMD will launch the new generation 6xxx cards which cream the 490. nVidia is behind going into next year. The 5xx better be non power hungry and take up 200watts or less. 335watts for a 480 is just ridiculous.

Will nVidia strike back with a DUAL GPU 580 GTX we shall stay tuned to find out. Thanks and gg and gb and gl,

 

[ModestGamer]

You really think AMD is going to get one over on Intel?

I think AMD plans to kick intels ass evrywhere they can for a while. They even completely abonded normal x86 core designs with bulldozer. This tells me they are going for kill.

How soon the rest of the CPU lineup get to see bd cores ??? But they look to be beasts. Whats interesting is that looking at AMD bulldozer info releases we see the same type of rethinking on bottle neck and instruction and intgr distrobution. Its interesting becuase thats where intel is going to be in trouble with there current designs. They can only keep tweaking the same basic core for so long before they hit efficiency problem with the overall architecture.

Its a complicated subject but right now. 2011 is going to be a great year for GPU and CPU tech.

 

[ModestGamer]

Great post Tempered81, so in a nutshell, AMD's approach means that the die is doing more work per mm2 than nVidia's approach. HD 5870 basically has 320 clusters of processors in which each one can execute up to 5 instructions in the best case scenario and much smaller die size competing against for example, the GTX 470 which has 448 processors which can only execute one instruction per CUDA core, and a much bigger die size.

Looks that way. They are taking the same approach to the bulldozer cpu as well or maybe this new line of chips took the bulldozer appraoch to GPU. If so its telling what AMD has up its sleeve going forward. D:

 

[digitaldurandal]

Those are sick Crysis benchmarks. 43fps avg at that res and 4x AA wow.

I hope nVidia can attack back with their 5xx line. But I think nVidia lost for this year. All they will release from now and next year is a 490 GTX card. In this same time AMD will launch the new generation 6xxx cards which cream the 490. nVidia is behind going into next year. The 5xx better be non power hungry and take up 200watts or less. 335watts for a 480 is just ridiculous.

Will nVidia strike back with a DUAL GPU 580 GTX we shall stay tuned to find out. Thanks and gg and gb and gl,

Are you on adderall? Just curious.

 

[PingviN]

Are you on adderall? Just curious.

He's just random.

 

[bryanW1995]

That seems like a very small niche to try and stick the 6850 into in a way. The disparity between 5870 and 470 is so tiny, usually 5fps or so, that slotting a card into that space would almost be tedious.

I'm expecting 6850 to be GTX 480 performance and the 6870 to be about 125-135% of what a GTX 480 is. The rest of the lineup, who knows. 6830/6770 will probably slot in where the 5850 and GTX 460 are currently.

while that would be nice, even the current round of benchmarks only shows ~ 35% improvement over 5870. gtx 480 is, what, 15-20% faster than 5870 right? so you're looking at 15-20% faster than gtx 480 even if the leaks turn out to be true. that doesn't leave much room for 6850 to be faster than 480, especially if amd is really determined to keep better separation btwn 6850 and 6870 as the rumor states.

edit: wow, just read tempered81's post. great, concise, clear post.

 

[solofly]

Thanks and gg and gb and gl,

Ok, gb I think means good bye, gl good luck but what about gg? it can't be good game...

 

[bryanW1995]

I think it makes sense and doesn't mandate major changes. If you are able to tweak the product to produce 10-20% better IPC, and then utilize the the 40nm process more efficiently to allow more transistors within that same die space adding to that the maturity of the 40nm process to allow a clock speed increase in the range of 10-20% increase... The end result could be quite large for just some simple tweaks and updates...

Major would mean a redesign of the architecture of the chip, IMO... Doesn't sound like this is major at all. It sounds more like G80 versus G92 in a lot ways, without the gimped memory bandwidh and ROP decrease that G92 had.

um, yeah, except that g80 was actually, you know, FASTER than g92 for about a year after g92 limped into our lives, remember?

if 4890 was a "minor tweak" and 5xxx was a full step, then 6xxx appears that it will be a 1/2 step. as tempered81 has mentioned, amd is only using 3/4 of a "full" 6xxx gpu this round due to heat/power/size constraints, but the optimizations are still there hence the significant ipc improvements. - edit: ok, maybe more like a full step but amd was tripped by tsmc so they kinda stumbled into a 1/2 step even with the new arch.

 

[ViRGE]

Take this analogy and imagine that only 18-wheelers can use the 5th booth. However, there is only an 18-wheeler passing through this toll-way every 5 mins or so, while there are regular passenger vehicles passing through by the 1000's per minute. So 4 of the 5 booths are at peak efficiency with passenger cars lined up in front of them constantly, while the 5th booth has an 18-wheeler passing through it every couple minutes or so. The 18-wheeler cant pass through booths 1-4, and likewise, the passenger cars can't use the 5th booth. That's an Evergreen shader.

This is a pretty good write-up, but I feel there's something critical you missed here.

The T-unit isn't "18-wheeler only" as you put it. It can do a FP/INT/MOVECMP per cycle just like X/Y/Z/W can; the only difference is that it can also do a Special/Transcendental Function per clock. Thus having the T-Unit capable of SFUs did not waste any space with respect to FP/INT/MOVECMP, because when you weren't doing a SFU on the T-unit you could do a FP/INT/MOVECMP.

The only downside to this approach is that the T-unit can't participate in Double Precision math, where AMD's architecture chains together 2 or 4 of the regular SPs as necessary. So your best-case scenario for FP64 MAD is 1/5th FP32 MAD, because you can't use that T-unit.

Plus assuming all of this is true, I'm not convinced those of us in the peanut gallery even have the reasoning behind the change correct. Keep in mind that AMD doesn't even do FP64 on lower-end parts, so the uselessness of the T-unit alone can't be the reason for the change. Otherwise the only way a 5-wide design can be worse than a 4-wide design is if you can't extract enough ILP to fill all 5 units - moving to a 4-wide design would be tantamount to admitting that they built their shaders too wide. In which case dropping to a 4-wide design would allow them to have more SIMDs overall, trading ILP extraction for TLP extraction.

 

[ArchAngel777]

um, yeah, except that g80 was actually, you know, FASTER than g92 for about a year after g92 limped into our lives, remember?

if 4890 was a "minor tweak" and 5xxx was a full step, then 6xxx appears that it will be a 1/2 step. as tempered81 has mentioned, amd is only using 3/4 of a "full" 6xxx gpu this round due to heat/power/size constraints, but the optimizations are still there hence the significant ipc improvements. - edit: ok, maybe more like a full step but amd was tripped by tsmc so they kinda stumbled into a 1/2 step even with the new arch.

That depends on how you look at it. G92 trumps G80 in shader power.

 

[ShadowOfMyself]

while that would be nice, even the current round of benchmarks only shows ~ 35% improvement over 5870. gtx 460 is, what, 15-20% faster than 5870 right? so you're looking at 15-20% faster than gtx 480 even if the leaks turn out to be true. that doesn't leave much room for 5850 to be faster than 480, especially if amd is really determined to keep better separation btwn 5850 and 5870 as the rumor states.

You might wanna read this post again lol the numbers are all over the place, Im pretty sure this is what you meant:

while that would be nice, even the current round of benchmarks only shows ~ 35% improvement over 5870. gtx 480 is, what, 15-20% faster than 5870 right? so you're looking at 15-20% faster than gtx 480 even if the leaks turn out to be true. that doesn't leave much room for 6850 to be faster than 480, especially if amd is really determined to keep better separation btwn 6850 and 6870 as the rumor states.

 

[bryanW1995]

You might wanna read this post again lol the numbers are all over the place, Im pretty sure this is what you meant:

wow...too much beer this afternoon it seems... I think I'll just go drink my sorrows away!!

 

[badb0y]

+1 to Tempered81's post, clears some stuff up nicely.


We allow cussing in P&N and OT, not in the tech forums.

Moderator Idontcare

 

[MrK6]

while that would be nice, even the current round of benchmarks only shows ~ 35% improvement over 5870. gtx 480 is, what, 15-20% faster than 5870 right? so you're looking at 15-20% faster than gtx 480 even if the leaks turn out to be true. that doesn't leave much room for 6850 to be faster than 480, especially if amd is really determined to keep better separation btwn 6850 and 6870 as the rumor states.

edit: wow, just read tempered81's post. great, concise, clear post.

Not by my math:

Vantage -
GTX 480 stock, GPU score (not total) is 18026: http://hardforum.com/showthread.php?t=1513719
6870 reference bench, GPU score is 24056: http://img291.imageshack.us/img291/3225/vantaget.jpg
6870 appears to be ~33% faster than the GTX 480 in Vantage

Also, here's a run of the GTX 480 in Crysis Benchmark: http://www.tomshardware.com/reviews/geforce-gtx-480,2585-10.html
Without even taking the lower resolution that the GTX 480 was running, the 6870 is already over 40% faster.

Seems like a 6850 that's 20% slower (which seems to be the trend) will still be 10-15% faster than the GTX 480.

 

[SlowSpyder]

Not by my math:

Vantage -
GTX 480 stock, GPU score (not total) is 18026: http://hardforum.com/showthread.php?t=1513719
6870 reference bench, GPU score is 24056: http://img291.imageshack.us/img291/3225/vantaget.jpg
6870 appears to be ~33% faster than the GTX 480 in Vantage

Also, here's a run of the GTX 480 in Crysis Benchmark: http://www.tomshardware.com/reviews/geforce-gtx-480,2585-10.html
Without even taking the lower resolution that the GTX 480 was running, the 6870 is already over 40% faster.

Seems like a 6850 that's 20% slower (which seems to be the trend) will still be 10-15% faster than the GTX 480.

IF these numbers are true, then I think these are some very impressive parts coming out by AMD. I think what makes them so impressive (again IF these numbers are correct) is that they're building them on the same process that the 5870 is built on. I would expect large gains from both camps on 28nm, but for another 40nm part these look great.

If they are able to launch over the next few months it would appear that AMD was able to get a new architecture out without too many bumps in the road. Fermi was late and had a few glaring issues that made it look somewhat rushed to me. For whatever reason AMD seems to be able to get their redesigned architecture out with out so many hiccups.

 

[Kenmitch]

Fermi was late and had a few glaring issues that made it look somewhat rushed to me.

I think some of the delay on Fermi can be blamed on the 5xxx series of cards. I'm thinking that nvidia under estimated ATI's abilities and had to scramble to produce a card that was faster....Then took the who cares how much power it takes and how much heat it'll produce stance on it!

It'll be nice once we start seeing these new cards hit the market. Can't wait to see some real reviews of them....That and the fanboyism involved with the release of a new video card series is kinda fun to read

 

[Grooveriding]

I think some of the delay on Fermi can be blamed on the 5xxx series of cards. I'm thinking that nvidia under estimated ATI's abilities and had to scramble to produce a card that was faster....Then took the who cares how much power it takes and how much heat it'll produce stance on it!

It'll be nice once we start seeing these new cards hit the market. Can't wait to see some real reviews of them....That and the fanboyism involved with the release of a new video card series is kinda fun to read

Maybe nvidia will just take a pass when 6870 drops and continues ATI's DX11 domination.

That or release a small thermonuclear reactor that is 15% faster... seven months later, so... September 2011 or something... ()

Get your 1500W power supplies ready!

 

[evolucion8]

Maybe nvidia will just take a pass when 6870 drops and continues ATI's DX11 domination.

That or release a small thermonuclear reactor that is 15% faster... seven months later, so... September 2011 or something... ()

Get your 1500W power supplies ready!

Oh noes, I had a hope that I could hold my current PSU a little longer... Jumps from a cliff

 

[tviceman]

Oh noes, I had a hope that I could hold my current PSU a little longer... Jumps from a cliff

Your signature link is broken, kinda like your comparisons.

gtx460 ($190 AR) vs. hd5850 ($260)

and

gtx470 ($280 AR) vs. hd5870 ($355)

I'd hope that more expensive cards would outperform less expensive ones - which may or may not make them "better" buys. BTW, those are all the cheapest current prices on Newegg.

EDIT: Thanks happy, fixed the prices. The 5850 is 37% more expensive than the gtx460 and the hd5870 is 27% more expensive.

 

[happy medium]

Your signature link is broken, kinda like your comparisons.

gtx460 ($218) vs. hd5850 ($260)

and

gtx470 ($279 AR) vs. hd5870 ($355)

I'd hope that more expensive cards would outperform less expensive ones - which may or may not make them "better" buys. BTW, those are all the cheapest current prices on Newegg.

there is a Palit gtx460 for $190 AR also.

http://www.newegg.com/Product/Produc...-075-_-Product

it's nice to see ATI prices dropping though, finally. If the 5850 reaches 220$ I'm in for 1.

 

[evolucion8]

Your signature link is broken, kinda like your comparisons.

gtx460 ($190 AR) vs. hd5850 ($260)

and

gtx470 ($280 AR) vs. hd5870 ($355)

I'd hope that more expensive cards would outperform less expensive ones - which may or may not make them "better" buys. BTW, those are all the cheapest current prices on Newegg.

EDIT: Thanks happy, fixed the prices. The 5850 is 37% more expensive than the gtx460 and the hd5870 is 27% more expensive.

Fixed, now you can see the comparisons, the HD 5850 can be bought for $239.99 AR.

http://www.newegg.com/Product/Produc...82E16814127500

Can we justify the price difference due to its performance? I think that depends of the tastes.