Official AMD Radeon 6000 Series Discussion Thread

[yogurt_21]

What is exactly the benefit of having 4D shaders?

ati uses a 1+ 4 config in their current and previous gens 1 complex, 4 simple. this means that on the 5870 you have 320 shaders capable of complex items and 1280 capable of simple items, most of the time 320 are maexed out in usage while the others sit idle OR most of the 320 complex are in use but some simple shaders have to be utilizes and you can only have 1 item transmitting at a time from that 1+4 cluster so you could have a scenario where complex shaders are needed but are idling because some clusters are workign on simple shader tasks.

so in reality when you see 1600, you're looking at 320 possible operations at any 1 given time.

if they were to have the same number of shaders at a 4d arch that would of course jump to 400 possible operations at any 1 given time. Increasing performance by default. Also there is talk that the 4d arch is supposed to be 4 moderate complexity rather than 3 simple + 1 complex. This would make each cluster capable of quickly processing anything it's given and then transmitting it. This as I've mentioned before won't affect peaks frams as much as minimum frames.

 

[HalfAHertz]

ati uses a 1+ 4 config in their current and previous gens 1 complex, 4 simple. this means that on the 5870 you have 320 shaders capable of complex items and 1280 capable of simple items, most of the time 320 are maxed out in usage while the others sit idle OR most of the 320 complex are in use but some simple shaders have to be utilizes and you can only have 1 item transmitting at a time from that 1+4 cluster so you could have a scenario where complex shaders are needed but are idling because some clusters are workign on simple shader tasks.

so in reality when you see 1600, you're looking at 320 possible operations at any 1 given time.

if they were to have the same number of shaders at a 4d arch that would of course jump to 400 possible operations at any 1 given time. Increasing performance by default. Also there is talk that the 4d arch is supposed to be 4 moderate complexity rather than 3 simple + 1 complex. This would make each cluster capable of quickly processing anything it's given and then transmitting it. This as I've mentioned before won't affect peaks frams as much as minimum frames.

If you look back all the way to the HD2900 Reviews, you'll find a deeper explanation of the architecture. This is a good example.

Edit - this is my summary:

There are five scalar units per block, each able to co-issue a floating-point multiply-add instruction per clock.

This means that all SPs both simple and advanced can do basic calculations, so whenever an algorithm uses those, all SPs will be used as long as there isn't another bottleneck somewhere else.(As is the case with the hd5800 vs the hd6800; shaders were reduced but performance is still very competitive.) If you want to do double point calculations (which games almost never do) then you are limited to the advanced 320SPs plus in some cases another special kind of 320SPs(which are basically the 4 simple SPs grouped together, a bit more on that further down)

The problem of the [4s+1a] (4 simple and 1 advanced) design is, that when you need to do a Vector calculation, you need to group all 4 simple SPs and use them all at the same time. And in those case the texture unit looks more like (1v+1a) to the driver.
Most of the times a game code is translated by the driver as a garbled mess of simple scalar calculations and moderately advanced vector calculations. The cool thing about ati's design however is that each SIMD(a group of 80 SPs) can be assigned to either vector calculations or scalar calculations. So that means that the total SP cound constantly carries between 320vector SPs(worst case scenario) and 1600 scalar SPs(best case scenario) or somewhere in between(f.e. 10SIMDs are in Vector mode providing 160vector SPs and 10SPs are in scalar mode providing 800 scalar SPs, resulting in a total of 960 mixed SPs)

Now because of the way games work the game code is very erratic and cannot be predicted. After all happens in real time and is being generated as fast as possible. That means that the game code always varies (You know that there will always be a bush rendered over there and a rock 3 feet away from it. But will the enemy getting ready to shoot at you decide to stand next to the bush or behind the rock? Will he shoot or suddenly decide to throw a grenade at you? What if his buddy wants to call him back for a cup of tee, trying to talk some sense into him and force him to forfeit his life of crime and misery?)
That all means that the code can never be completely predicted and that most of the time, cannot be distributed evenly amongst the SIMDs, thus leading to underutilization. This also explains why games are never as stressful as benchmarks and tests such as OCCT or that donut test.

And that's pretty much it. To be honest I really don't see how moving to a 4SP desing will solve this problem, at the very best it will only negate it.

 

[TheMailMan78]

I want a 580. My mind is made up.

 

[dir_d]

I want a 580. My mind is made up.

How? we have 0 official specs from both sides.

 

[TheMailMan78]

How? we have 0 official specs from both sides.

Just an educated guess. ATI I think is up to some shanniagns. I mean Ill wait and see mind you but as of right now I'm getting a 580.

 

[pantherx12]

I tell you what if the 580 came out when it was supposed too ATI would of been buggered.

( 580 is what the 480 was gona be before fab issues)

 

[cadaveca]

I want a 580. My mind is made up.

Your 5850 isn't fast enough for ya, that you need another card? Or have you got the hardware bug?

 

[Tatty_Two]

I tell you what if the 580 came out when it was supposed too ATI would of been buggered.

( 580 is what the 480 was gona be before fab issues)

OR in essence, NVidia took so long to release fermi, they edged their bets, more time was clearly needed to fully develop the potential of their top end card, they ran out of that time so instead delivered a "stunted" version of what they really wanted, thus giving them more time to fully develop their "flagship"......... to be fair to NVidia, they probably had no choice at the time and it has worked out reasonably well for them, what with the 460 coming along also, they have recovered some sales and IF, and only IF, the 580 produces the goods in comparison to their rivals new flagship offering, they may just make up a little more ground,

I always would caveat though with the point that these high end enthusiast products never have a huge impact and are often seen more as marketing tool for the more mainstream offerings, which in the case of ATi would appear at least for now to be in the position to potentially dominate the sector again with the 68XX series..... dunno, just my thoughts.

 

[mastrdrver]

I want a 580. My mind is made up.

If you change your mind like your avatar then I give it a week.

Anyone remember what the original TDP of the GTX 480 was before it got changed a lot?

Also remember that nVidia has to cover the market Tesla does with their GPUs while AMD doesn't since they have CPUs. So the GF104 was more of the PC market GPU while GF100 is more of a HTPC gone gaming GPU........or something like that.

 

[CDdude55]

Anyone remember what the original TDP of the GTX 480 was before it got changed a lot?

Wasn't the TDP like 220w for the 480?

 

[mastrdrver]

Originally? I don't think so. I was thinking more like high 200s, 280 or maybe that was the rumor and something like 250 was the initial. I know it has been revised down a couple times......somehow.

 

[a_ump]

HD 6XXX architecture explained!?

HD 6XXX series info

I wasn't sure if anyone else had seen this n i went back n checked hte last 5 pages which dates before this was posted on those forums and found it to be a great read. Idk enough personally to say whether its true or believable sounding bullshit but the guy made it sound good and probable lol. enjoy

 

[wahdangun]

Wasn't the TDP like 220w for the 480?

no its 250 watt but can reach over 300 watt if using furmark

 

[macks30]

ati 6xxx

re; ati cards 6xxx series I cann't see any big jumps just a number change !!

 

[DriedFrogPills]

re; ati cards 6xxx series I cann't see any big jumps just a number change !!

compared to what models of the 5xxx series? cause you do realise that the 6850 and 6870 are the replacements for the 5750 and 5770 over which they have a decent bump

 

[pantherx12]

re; ati cards 6xxx series I cann't see any big jumps just a number change !!

You must be what I call "trippin' balls".

You may also apparently have you back turned to nvidia who are renaming their 480 to 580.

 

[HalfAHertz]

HD 6XXX series info

I wasn't sure if anyone else had seen this n i went back n checked hte last 5 pages which dates before this was posted on those forums and found it to be a great read. Idk enough personally to say whether its true or believable sounding bullshit but the guy made it sound good and probable lol. enjoy

Think you can quickly summarize the important bits in English for us? Google just gives me a garbled mess of spanglish when i try to translate it.

 

[motasim]

think you can quickly summarize the important bits in english for us? Google just gives me a garbled mess of spanglish when i try to translate it.

+1

 

[H82LUZ73]

I have some questions to the guys who own these cards,Well my specs are at the side .Anyways I just bought myself an Onkyo Home theater in a box (HT-S5300) it does 3d so Wizz did you try hooking it up to a AV system and see if the card really does pass the DTS-HD MASTER over it so my AV can decode it.

 

[Anusha]

compared to what models of the 5xxx series? cause you do realise that the 6850 and 6870 are the replacements for the 5750 and 5770 over which they have a decent bump

if that is the case, shouldn't they be around the same price as the 5750/5770 release prices? ($149 and $199 weren't they?)

 

[Tatty_Two]

if that is the case, shouldn't they be around the same price as the 5750/5770 release prices? ($149 and $199 weren't they?)

Possibly, but with seriously greater performance, the company may have felt they could justify the higher prices, performance ultimatly dictates placement and pricing.

 

[Anusha]

Possibly, but with seriously greater performance, the company may have felt they could justify the higher prices, performance ultimatly dictates placement and pricing.

imo, they didn't hurt nVidia's sales with that price. they must have hurt their profit though.
the GTX460 1GB still seems to be a very good purchase for its price. 6870 at $199 price point would have done terrible things to nVidia imo.

 

[Tatty_Two]

imo, they didn't hurt nVidia's sales with that price. they must have hurt their profit though.
the GTX460 1GB still seems to be a very good purchase for its price. 6870 at $199 price point would have done terrible things to nVidia imo.

I hear what you are saying but look at it this way, on Newegg the cheapest 6870 is $239.99 from what i can see and the cheapest 1GB 460 is $199.99, the 6870 is 15% faster across all resolutions so the price difference for AMD is justified (I say for AMD because I personally am not commenting on the differences), however, the HD6850 is 2% faster than the 1GB 460 across all resolutions and can be had on Newegg for $20 cheaper than the 1GB 460 and therefore represents excellent value for money in performance terms against the competition.... and therefore I think they probably will hurt 460 sales.

Edit: Taking the value angle one step further, are you aware that across all resolutions, the HD 6870 is only 3% slower than a GTX470? and in my resolution 1920 x 1200 it is only 1% slower, now compare the prices between those 2 .....

 

[Benetanegia]

Think you can quickly summarize the important bits in English for us? Google just gives me a garbled mess of spanglish when i try to translate it.

Nothing new there actually. He is just talking about the change to 4 VLIW architecture and how that affects the theoretical performance. He is using a statistical aproach to do so.

In the first few paragraphs he is explaining how VLIW architecture works and how that affects actual performance rather than peak performance, depending on how many of the 5 ALUs can be used at any given time, mostly based on interdependence to each other: if function B depends on function A, function A and B cannot be executed in the same clock, etc.

He then goes on to calculate peak (5 out of 5 ALUs used) and minimum (1/5) FP capabilities of Cypress and the alleged Cayman with 1920 SP. It's worth mentioning that he thinks that SIMDs are going to be 32 SP wide, rather than 16 wide like in previous gens. Apparently he thinks so based on rumored die size and relation to TMU number, etc.

The conclusion to the calculations is that peak performance would be 20% higher than Cypress and minimum performance (when only 1 out of 5 ALU is working) would be 50% higher. Actual expected performance increase will obviously be somewhere in the middle.

As a personal note from me, I think that actual performance is going to be significantly closer to the 20% figure than 50% figure, unless we assume that R600->Cypress architecture was absolutely innefficient which is not only very unlikely (they would have changed it ipso facto) but stupid on AMD's part.

He does not take into account performance gains/losses that could be derived from the fact that the fat ALU has been "replaced" by adding medium compexity to all 4 ALUs, mainly because he is not even sure if AMD trully did that or they only eliminated one of the simple ones.

Again as a personal note, he is just exposing the fact that statistically, by only decreasing the ALU number per SP, efficiency goes up. He is not even trying to figure out other posible improvements, which is smart considering we are all just playing with rumors.

 

[a_ump]

Nothing new there actually. He is just talking about the change to 4 VLIW architecture and how that affects the theoretical performance. He is using a statistical aproach to do so.

In the first few paragraphs he is explaining how VLIW architecture works and how that affects actual performance rather than peak performance, depending on how many of the 5 ALUs can be used at any given time, mostly based on interdependence to each other: if function B depends on function A, function A and B cannot be executed in the same clock, etc.

He then goes on to calculate peak (5 out of 5 ALUs used) and minimum (1/5) FP capabilities of Cypress and the alleged Cayman with 1920 SP. It's worth mentioning that he thinks that SIMDs are going to be 32 SP wide, rather than 16 wide like in previous gens. Apparently he thinks so based on rumored die size and relation to TMU number, etc.

The conclusion to the calculations is that peak performance would be 20% higher than Cypress and minimum performance (when only 1 out of 5 ALU is working) would be 50% higher. Actual expected performance increase will obviously be somewhere in the middle.

As a personal note from me, I think that actual performance is going to be significantly closer to the 20% figure than 50% figure, unless we assume that R600->Cypress architecture was absolutely innefficient which is not only very unlikely (they would have changed it ipso facto) but stupid on AMD's part.

He does not take into account performance gains/losses that could be derived from the fact that the fat ALU has been "replaced" by adding medium compexity to all 4 ALUs, mainly because he is not even sure if AMD trully did that or they only eliminated one of the simple ones.

Again as a personal note, he is just exposing the fact that statistically, by only decreasing the ALU number per SP, efficiency goes up. He is not even trying to figure out other posible improvements, which is smart considering we are all just playing with rumors.

thanks for explaining it to everyone else lol. i understood it but wasn't quite sure how to go about explaining it to others.