News - ATI Radeon HD 5870 hits the rumour mill

[kalniel]

what i meant was compared with the GT300 rumours which are promising the start of a shift towards cGPU this is nothing particualy stunning.

All dx11 class cards will be capable of such a shift, it's not nVidia specific.

First DX11 chip ok il give you that but from the specs it seems to be pretty much just an upgraded 4 series architecture.

You don't go from dx10 to dx11 without some fairly significant architecture changes.

[Biscuit]

All dx11 class cards will be capable of such a shift, it's not nVidia specific.

You don't go from dx10 to dx11 without some fairly significant architecture changes.

Fair enough. Guess i dont really know enough about it all to really argue the point.

[kalniel]

This is a fairly good introduction to DirectX 11 and what it involves on the architecture front.

http://www.anandtech.com/video/showdoc.aspx?i=3507

I think it's going to be easier for ATI because they already have some of the features in hardware already (Dx10.1 stuff, tesselation unit), but it's still a big change.

[borandi]

Question is, how about double precision? Currently only x8xx cards do double precision calculations. Fast is good, but we'll wait until OpenCL comes into effect. Though atm two 4850s are doing me fine, more speed always helps

[GheeTsar]

Looks like the Peak Gtexel/s (FP16, bilinear) doesn't scale very well with the x2 model. Both values also seem to be lower than the GTX295 too which surprises me.

[alwayssts]

Short version: 40nm and 32nm could put nvidia/ati back on par; give or take.


My personal theory on GT300:

GPU: 700mhz
Shader: 1750mhz
Memory: 4xxx (5ghz gddr5)
Memory Bus: 512-bit
Mem Size: 2GB
Fab Tech: 40nm
DX/SM: DX11, SM5
Shaders: 512sp, consisting of 2 dual-issue madds + MUL (5 flops per clock)
TMUs: 128
ROPs: 64
Peak Gflops: 4,480
Fillrate: 48
Texal: 96


Not sure if that's general consensus or not, but making the madd dual-issue would make sense with the MIMD structure. It would also explain why the X2 is clocked 50mhz higher: At that clockspeed it would edge out GT300 in a couple key area by 1-2% in raw spec. Sure, unless they have some crossfire/sideport/MCM miracle-scaling (hydra?) mechanism up their sleeve GT300 would prevail, but it's always nice when you can claim the raw theoretical higher number.

Also, we can expect 32nm from ATi/nvidia H1 next year. I would think ATi Q1 and Nvidia Q2. GF (old AMD fabs) is taking designs in Q409, and you can bet your ass ATi is going to make their gpus there ASAP.

This should bring mid-range (800sp/32tmu) and low-end (400sp/16tmu) parts. The former being clocked at 950mhz, 1520 gflops. The later clocked at 625mhz on the low-end, giving .5TF to the IGP market, or in other words, R600/rv670 spec.

I would expect ATi to go from 12 to 16 arrays on 5970 (1600 shaders, 64 TMUs) based on how 32nm should scale. This would put a 950mhz model at just over the magic 3TF, and 800 just over 2.5TF. See how this year's rv870 with 1200sp would slip in there? In a perfect spot to retain sales, even when it drops in price compared to the new part.

Quite interesting when you see that rv870 is aiming for 900mhz and 750mhz for the 40nm 5870 and 5850 parts respectively. A 50mhz jump, just like rv790->rv870.

This would mean AMD's market would go from .5TF (400sp,625mhz) to 6TF (5970x2), with the three chip segments completing a full trickle-down from their counterparts (rv670/730/810-.5TF, rv790->rv830-1.28-1.52TF, rv770x2->890-2.56-3.04TF/x2 - 5.12-6.4TF).

Nvidia you can expect to shrink their 500mm2+ 3B transistor monster and ramp the clockspeed, I would think 800c/2000s. This would mean 5120 Gflops. The same as an 800mhz 5950x2.

I would think that 5ghz will end up going to the lower-end (gtx365, HD4790, 5670, HD5950) and 7ghz gddr5 will go to the top end parts (GTX385, 5970,x2).

Comparatively speaking on this hypothetical high-end; The 5970x2 at 1000mhz could compete with the GTX385 on even keal, if it could scale @ 1.8x versus a single chip 5970...Very realistic, as that is about what we currently see crossfire scale.

How you like them speculative apples?!

[kalniel]

Short version: 40nm and 32nm could put nvidia/ati back on par; give or take.

What do you mean back on par? They were on par last generation (after nVidia's price realignment). I think your specs on the GT300 chip are way out too, sorry

[TooNice]

I wonder how well nVidia would cope with a new architecture + a die shrink simultaneously if that's what they are planning. When was the last thing it was done? Paraphrasing (probably badly) a quote from Intel in an article I read some time ago, you either do a die shrink or go for new architecture. Doing both at the same time is asking for all sorts of trouble.

[dangel]

I thought the 4 series ATI cards where only better power consumption than the GTX series when you dont count idle power.... which in my book is a bit silly

Really? Personally I'd say i'm pretty interested in not having it eat loadsa-powa when i'm just running desktop apps!

[TooNice]

^ Aren't you two implying the same thing? Biscuit said that ATI is only better when you exclude idle power consumption which is silly. And you say that you would be interested in knowing how much power a GFX card gobs up when you are using desktop apps (more or less idling).

[dangel]

I thought ATI were better when you INCLUDE idle consumption (it's less than nVidia) oui?

My 280 eats power on the desktop which is silly really.

[Biscuit]

no Nvidia has better idle consumption than ATI but when under full load ATI is better performance per watt

[dangel]

Ah I see - ta!

[Methanoid]

no Nvidia has better idle consumption than ATI but when under full load ATI is better performance per watt

Horse manure....

http://archive.atomicmpc.com.au/foru...c=7&t=9354&p=0

Although this doesnt have the VERY latest cards from both sides it does cover 4670-4850 for example (ie mainstream) and Nvidia is POOR on power consumption.

There was just so long I could go letting this inaccuracy slide...

[Biscuit]

is better performance per watt

I admit overall the ATI cards seem to have overall better idle power consumption than the Nvidia 'equivalents' but that chart doesnt take in the above quote at all and the main cards which are compared (as they have extremely similar performance) are 4870 and GTX260 where my statement holds true.

[kalniel]

Horse manure....

http://archive.atomicmpc.com.au/foru...c=7&t=9354&p=0

Although this doesnt have the VERY latest cards from both sides it does cover 4670-4850 for example (ie mainstream) and Nvidia is POOR on power consumption.

There was just so long I could go letting this inaccuracy slide...

You're not talking about the performance sector here, but that's what we're interested in, and it's a very different story purely because of the DDR5 in the 4870 upwards which can't be put in a low voltage idle state. Therefore ATI cards have high idle power compared to the nVidia's, but a lower 100% load draw.