So what's the difference between the 6870 shaders and the 5870 shaders?
-
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.
ATI cuts 6950 allocation
- Thread starter An Exception
- Start date
So what's the difference between the 6870 shaders and the 5870 shaders?
Very useful summary table
Where did you get it from out of interest?Fundamentally, the shaders in the 5870 were arranged into blocks of 5, with four "simple" shaders, and one "complex" shader:
The complex shader [t] is designed to perform a transcendental operation (computing a number like, say, pi, square-root, or a sine/cosine, where an iterative refinement process is required). Barts uses this same configuration. The other shaders [x, y, z, w] perform simple MADD (multiply-add) operations. Since the vast majority of operators are multiply-add, the 5th unit gets used much less than the other four, and also takes up more transistors (and so die-space).
With Cayman, the SPs are instead arranged into blocks of four. Each SP is identical. The SPs are certainly capable of performing MADD operations. We don't yet know exactly how transcendental operations are going to be handled, but there are many options. One is to use a single SP and force it to do several iterations (i.e. use several clock cycles to compute the transcendental). Another is to use just a few clock cycles and use all four SPs.
Whichever way you look at it, the "4D" arrangement of Cayman should be faster at performing MADD operations (by up to 25%, since the "dead" 5th shader is not in use), but far slower at computing transcendentals (by how much, we don't know). Since the vast majority of computations required are MADD and not transcendental, it's reasonable to assume that the 4D arrangement should be faster.
It's actually a little more complex than the above, since it all depends on the size of the vector which is passed to the shader group (which tends to be four or less - hence the 5th shader being "dead" in most cases), but it works as an illustration of the concept.
edit - crap... I misread your question... The answer I give above is for how the 6970 shaders will differ from those in the 6870 (and 5870). i.e. the difference between the "4D" and "5D" arrangement.
Last edited:
Associate
- Joined
- 23 May 2005
- Posts
- 812
- Location
- Hove
None the Barts shaders are the same as Cypress, Juniper etc.
So how does that work then?
The 6870 on par with 5870 is some games and benchies although it has 1120 sp v the 1600 of the 5870?
5970 still is the fastest card in the world.
been King for a long time.
6990 be the new king.
6970 or 6950 in crossfire likely will scale "much" better than nvidias sli and bring tremendous value to us as a consumer.
As far as my technical knowledge about the workings of a GPU, it seems AMD sprinkled ground unicorn hearts on the PCB to make it work better.
From what has been said, the layout of the processors have been rearranged and they use a new front end (? whatever that is) to overall increase the efficiency of how the GPU tackles a job and therefore achieves more with less.
While the shaders have stayed fundamentally the same, AMD has improved a lot of the control logic (updated scheduler / dispatch processor etc). They have also re-balanced the GPU, offering slightly less compute power, but more texturing power. The clockspeed has also risen slightly...
Combined, these effects have given a nice performance boost to Barts over Cypress.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
This is what people are forgetting, barts has a new front end, similar/same to what Cayman should have, this increased efficiency dramatically.
From the benchmarks up there in the tables(which aren't very accurate, compare results, some sites have Dirt 2 10-15% in the 5970 favour, some have it 10% to the 580gtx, Crysis Warhead swings from review to review, Call of Juarez 2 DOES work in crossfire and the 5970 is 55% ahead of it and the 5870 beats a 480gtx and is much much closer than the review would have you believe.
Anyway just assuming they are right the 6870 is sub 50% behind the 580gtx, the 6970 has a circa 70% effective increase in shaders over barts, do the math.
one 4d shader should provide about the performance of one 5d shader, so shader for shader they are about 25% faster. Meaning 1536 shaders is equivilent to 1920 shaders, but with BARTS front end, that meanst 1120 up to 1920 = 71% increase in shader power, roughly.
A 1536 shader Barts would be gaining around 40% shader power, I'd frankly expect a 1536 Barts to beat a 480gtx comftably and be not far off a 580gtx anyway, potentially beating a 570gtx(with a increase in memory clock speeds to bump bandwidth with more shaders). Cayman could be potentially adding another 25% from shader performance with better shaders, more 4d shader clusters meaning better performance in worst case scenario's.
But other limitations might mean that much shader power on a 256bit bus, and similar internal bandwidth means it might not scale up with shader power perfectly. Theres also the possibility it could scale better with more tweaks to the memory controller and front end. I think the 1536 should be close to a 580gtx, potentially 10-15% faster anyway, if its got 1920 shaders, and the 4d shaders do give a 25% performance boost, it could utterly destroy the 580gtx.
Associate
- Joined
- 23 May 2005
- Posts
- 812
- Location
- Hove
If 1536 shaders then it will be close between the 580 and Cayman XT. I will go for an AMD win at around 5-10%.
If 1920 shaders then Cayman XT will be the fastest. A margin of 30% would not be unreasonable.
If 1920 shaders then Cayman XT will be the fastest. A margin of 30% would not be unreasonable.
- Joined
- 18 Oct 2002
- Posts
- 4,422
- Location
- 2nd City - Manchester.
Heart says I want AMD to have it - head says NVidia will still have it - but not by much, will cost more (as is their right) but also take more power and have more heat.
I can't see the 6970 being at 1920 shaders (or it would murder the 580) - I think that comes from the 5D vs 4D confusion multiplied by the number of ROP/Clusters etc - but I hope to be wrong.
Last edited:
I cba to read what you guys posted cause it's mostly rolled over thread over 9000 times already but I personally don't see why anyone would doubt that Cayman XT has 1920 shaders and Cayman Pro has 1536 shaders with one or two clusters cut off (not sure how it works but it's 20% off). That will put Radeon 6950 very slightly ahead of GTX570 and should be competitive price-wise (bear in mind, it's price is dropping to £249-259 now) and HD6970 approximately 15% ahead of GTX580. I don't see why any of this wouldn't be possible.
It's quite obvious that Nvidia is working on a dual-gpu card as well and AMD has HD6990 to counter that. Both companies did pretty well in maxing out 40nm potential IMO, it's all down to process shrink by the end of next year me thinks.
It's my table, obviously.
It's quite obvious that Nvidia is working on a dual-gpu card as well and AMD has HD6990 to counter that. Both companies did pretty well in maxing out 40nm potential IMO, it's all down to process shrink by the end of next year me thinks.
It's my table, obviously.
Last edited:
No reason why that would be obvious... But nice work anyhow. It's good to have a summary of the findings from each review site for a range of cards.
Well, it's been around for quite some time now. Gotta pay attention to what is being posted in the Forums
Plus I would obviously reference it if it wasn't mine
I think that at the end of the year the ATI 5970 will remain the fastest card available, closely followed bye the Nvidia GTX580 then the AMD 6970, Nvidia GTX570 with the AMD 6950 bringing up the rear.
now of course very early next year both AMD and Nvidia will bring out dual GPU cards and then it will get very interesting as to what would be faster dual Nvidia GTX570 derivatives or AMD's 6950/70 derivatives.
now of course both of these will be severely hampered initially by the power and thermal issues but I'm sure they will be able to fly once these things are opened up by the drivers control panels or third party software.
I'm very confused right now, is there any evidance that Cayman XT is a 1536 shader card? I think it was claimed that Cayman Pro is the one with 1536 this morning?
