-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD's 6870 coming in November - Early silicon doing very well
- Thread starter An Exception
- Start date
unfortunately for nvidia (and maybe my wallet) you're probably right
Maybe Metro 2033 will be able to be maxed with great frame rates and all bells and whistles (at awesome resolutions) 
(not aware of a single card capable of doing such at the mo, not that I extensively researched this...so maybe I fail)
(not aware of a single card capable of doing such at the mo, not that I extensively researched this...so maybe I fail)
Last edited:
Soldato
Looks like NDA could be up on the 21st.
http://obrovsky.blogspot.com/2010/09/radeony-hd-6800-vime-vse.html
http://obrovsky.blogspot.com/2010/09/radeony-hd-6800-vime-vse.html
Soldato
Last edited:
Soldato
AMD HD3D technology
o Stereoscopic 3D display/glasses support
o Blu-ray 3D support
o Stereoscopic 3D gaming
o 3rd party Stereoscopic 3D middleware software support
Did the 5xxx series have that?
o Stereoscopic 3D display/glasses support
o Blu-ray 3D support
o Stereoscopic 3D gaming
o 3rd party Stereoscopic 3D middleware software support
Did the 5xxx series have that?
Associate
Looks promising but new tech does too me 
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Thats really daft, can't fake the bios version? no one out of official AMD and select testing partners know the bios number, they could make up ANYTHING. The only reason to hide the bios number is its yet another way for AMD to spot whose leaking info.
As for the numbers, people need to stop thinking the 6870 will be double the specs/performance of the 6770, at minimum the 6770 is going to be around 240mm2, doubling everything and you'd have an almost 500mm2 part, AMD will catagorically not make a core that big.
The difference in shaders from top end to bottom end increases and allows space for more varients, last gen the 128/256bit bus barrier, and high end only having 1600 shaders meant there just wasn't room for a cost effective 3/4 speed part, this gen there is. Though its not likely to be 3/4 the speed maybe 60-66% the speed of the 6870.
As for the numbers mentioned, yes a 4way shader should indeed be more efficient, but it won't automatically be 25% more efficent, that calculation is incredibly, and MASSIVELY wrong. It assumes that its easy to get 4 calculations out of the 5 way shader, and that the 5th is the only one being wasted, its not even close to that.
Of the 5 way shader, its averaging somewhere between 2 and 3 instructions per clock, the 4 way shader should help, even more so that they will be more complex than the 4 single shaders, but it will not, in any way, max out 4 instructions every clock, in incredibly optimised, incredibly predictable gpgpu data theoretically it should be easier to design code to fully(or close to) fully utilise the 4(or current 4+1) shader setup, in reality and in gaming where what the scene needs changes from frame to frame, you don't get close to theoretical performance. Remember AMD's theoretical performance puts it miles, and miles and miles ahead of Nvidia, theoretical is based on every shader being maxed every clock. nvidia's architecture is essentially very basic, but very easy to max out every shader every clock, which makes it very easy to leverage the power out of.
25% more throughput from a 4 way shader design is beyond pie in the sky wishful thinking, 5-10%, maybe.
If only 1% of clocks manage to max out the whole current 5 shaders, and only 10% can use the full 4, but 79% use 3 or less shaders, then moving to a 4 way isn't going to improve much. The increased complexity of the 4 "bigger" shaders than 4 ickle basic ones and one complex is likely where the big increase comes in. When you did need to do something on the more complex shader, you were limited to 320 shaders, meaning the other 4 would be wasted if they were waiting on data from the 5th, the throughput of complex instructions should go up dramatically, however, the downside is, the vast majority of instructions or simple or fairly simple over complex.
We should see higher minimum fps, max fps, and general performance though given the same shader count with the 4 way shaders vs 5way, but not that big a jump.
Likewise in a lot of situations doubling the shader count didn't double performance, though in a few games like Metro the 5870 gave double the performance of a 4890 which was matched to the frame by the 5770. In other games, 20% more shaders won't give 20% more performance.
In the very best case scenario, in heavily optimised situations, in the right game performance increases should be very impressive, in general performance across all games, the 50% more performance from 25% more shaders and 25% more throughput from 4 way shaders is incredibly, incredibly unlikely.
I thought they were going for more banks of shaders, so instead of 1600 shaders = 320x5, it'd be 1600 = 400x4, or if it's a 1920 shader part, 480x4.
Same amount of shaders, or an easier way to boost the overall shader count by a small amount, but result in larger throughput potential. Surely that's the whole point of going for banks of 4 instead of 5, as it's easier to max 4 out rather than 5.
That'd mean at worst, if the cores are only managing 1 instruction per shader block, the 6870 would still be getting 50% more work done than a 5870.
Last edited:
This too, if they're correct, then the 6770 will have 320x4 shaders, while the 6870 will have 480x4 which works out at 2/3 or 50% more.
No one whos faking it is gonna risk making one up when someone in the know could prove they were just making it up.
Unless they are batch or vendor/feature specific BIOS they usually use generic numbers anyhow so not very specifically identifiable. Granted it could be someone testing an AIB specific BIOS but unlikely.
My first thoughts on that was AMD could very easily use the bios number to identify the code, assuming of course that it's real. It's not unreasonable to assume they could use specific bios numbers on sample cards as a means to identify them really, while it's all pre-release, you just don't know.
Drunkenmaster i have enjoyed your posts in the past but i think you have your maths a bit wrong.
When you say this
"As for the numbers mentioned, yes a 4way shader should indeed be more efficient, but it won't automatically be 25% more efficent, that calculation is incredibly, and MASSIVELY wrong. It assumes that its easy to get 4 calculations out of the 5 way shader, and that the 5th is the only one being wasted, its not even close to that"
I think you are missing the point. Assuming that by moving to a 4 way part you are effectively getting room for 25% more shaders units for free (320x5 to 400x4, 400/320 =1.25), and assuming from your example only 1% of the 5th part of the old shader was ever being maxed out (ie effectively nothing - your numbers) then on the new 4way shader these will supply 25% more power for the same area on the die.
Essentially the new 6970 (if the new naming scheme suggested is correct) you'e getting 50% more shader power as previously noted by other people.
By saying the 5th shader unit is never used, then no matter how efficient the other 4 are say "x", by multipling the number of shaders groups by 50% you are still getting 1.5x.
50 percent more shader power for 20% more area (Total shaders 1920/1600 = 1.2
And this power will express itself in the worst of situations the strongest ie mins, as can be seen with Nvidia's more simplified shader designs.
Which is what kylew said but in less words...(oops)
When you say this
"As for the numbers mentioned, yes a 4way shader should indeed be more efficient, but it won't automatically be 25% more efficent, that calculation is incredibly, and MASSIVELY wrong. It assumes that its easy to get 4 calculations out of the 5 way shader, and that the 5th is the only one being wasted, its not even close to that"
I think you are missing the point. Assuming that by moving to a 4 way part you are effectively getting room for 25% more shaders units for free (320x5 to 400x4, 400/320 =1.25), and assuming from your example only 1% of the 5th part of the old shader was ever being maxed out (ie effectively nothing - your numbers) then on the new 4way shader these will supply 25% more power for the same area on the die.
Essentially the new 6970 (if the new naming scheme suggested is correct) you'e getting 50% more shader power as previously noted by other people.
By saying the 5th shader unit is never used, then no matter how efficient the other 4 are say "x", by multipling the number of shaders groups by 50% you are still getting 1.5x.
50 percent more shader power for 20% more area (Total shaders 1920/1600 = 1.2
And this power will express itself in the worst of situations the strongest ie mins, as can be seen with Nvidia's more simplified shader designs.
Which is what kylew said but in less words...(oops)
Last edited:
Soldato
October release now or somebody getting the wrong end of the stick? http://www.digitimes.com/news/a20101001PD208.html
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Actually, again I was simplifying, the 5th shader wasn't often used, largely because it was the "complex" shader, the new shaders are said to be less complex, when there is complex things to be done that 5th shader could dramatically speed things up, with the less complex shaders, the inbetween shaders it would seem it will likely either do an ultra complex calculation using maybe 2 of the 4 shaders, maybe all 4, maybe it can only run one instruction on the 4 per clock and its either 4 simple or one complex, in which case in those situations the 4+1 where it can do 4 simple and 1 complex will be much much better.
Theres ALWAYS tradeoff's in performance, this is ignoring things like only one thread per cluster, the fact that there simple IS NOT in any way a 50% increase in shaders, unless its a 2200 shader part, it didn't happen and pretending it has is really really not accurate.
It COULD BE like a card that has 50% more shader power IN SOME SITUATIONS, in others it could well work as a card with less shader power.
Depending on how it handles more complex instructions, you might find it can only run 400 high complexity instructions per clock with NO simple shader functions where the current 4+1 could only run 320 but could run at least 320, and up to 1280 simple instructions at the same time.
There are always positives and negatives. Some minimums are likely to be higher with more 4way shader clusters, other minimums where a balance of complex and simple instructions might have lower minimums.
We don't know, at all, no one knows anything about how it handles different instructions, but as I said, claiming a 50% performance increase with such incredibly basic and inaccurate reasoning is just that, inaccurate, and misleading.
More shader clusters could simple prove to be overwhelming to the on die bandwidth/communication, maxing out more clusters could prove a problem, likely will.
The higher number of clusters could simple leave more clusters waiting for data longer. As I said before, EVERYTHING is a tradeoff, more shaders, woo, more clusters, yay, same on die bandwidth as before feeding more shaders, more clusters, more data faster.... not at all great. More clusters, no increase to on die bandwidth, not great, more on die bandwidth but no more clusters, not great.
Have they managed a good bump in internal bandwidth along with the new rop engine, who knows, these are all questions, they'll ALL effect performance, some will hold back performance and some will bump it up massively. Almost never is every increase in anything 100% positive nor will give linear increases. if the 4 way shaders offered 50% more shader power, its really this simple AMD would have started off with that ages ago.