Archives Discussions

[avk]

Why not?

Let's look at the current core frequencies:

4850: 625 MHz;
4870: 750 MHz.

As you can know, almost every hardware vendor striving to release its own designed cards, and the non-reference frequencies on them are not so seldom. So why not to let AMD/ATI to make more money by releasing another models of RV770?
Let's do some calculations :

750 (as 4870) / 625 (as 4850) = 1.2 times = 120%. What about these new models?:

4860 = 687.5 MHz (110%);
4880 = 812.5 MHz (130%);
4890 = 875.0 MHz (140%).

Of course, those new frequencies should not to be exactly (say -5%/+5%). Of course, the models with higher frequencies (4880, 4890) will be much more rare than ones with lower, but they will cost more money for the consumer and make more profit for the vendor!
You see that nVidia will release GeForce GTX 270 (yes, a middle of 260 and 280) very soon, and, IMHO, it will be a clever decision to breed more 48xx. Don't you think?

[avk]

Seems that TSMC's 40 nm technology has not so good yield . So why haven't you, guy, release 4890X2? I understand that this model is very complicated one to design, but who knows when TSMC's 40 nm will be fine? Who knows how long should you wait to produce RV870 in large quantity?

About Catalyst. You, guys, are capable to design the state-of-art chips like RV7xx, but, alas, Catalyst's quality is not perfect. I understand that it is extremely difficult to support several driver versions (I mean: WinXP-32, WinXP-64, WinVista-32, WinVista-64, Win7-32, Win7-64 and lots of Linux/Unix), but you should definitely do something. Do everything you want, hire for Terri Makedon's team the best specialists in the world (I know, it will costs some money), but, please, make Catalyst perfect (I mean: no graphics issues and superb performance). I think it is shame to release so named Catalyst Hotfixes every time. Perhaps, you should change Catalyst's monthly schedule to 2 or even 3 months. Please, remember what was happened with S3 and Matrox - both of them are outsiders now because of nasty drivers.

BTW, please, emphasize gamer's attention on Radeon's Edge-detect CFAA. This technology makes PC games' visuals pretty appeal, so tell gamers about it everywhere you can, at least at AMD Game.

[avk]

It seems that RV840 and RV870 are well designed chips. But we know that nVidia, sooner or later, can release a rivals. Although I'm a complete ignoramus in the graphics chip design, I'd like to wonder about, say, RV850 and RV880 .
First of all, I think that RV740's design was...more proportional rather than RV770's, although I don't know why I'm thinking that. Maybe because of the fact that RV740 and RV840 are almost equal in the DirectX9&10-performance ?
So, I dream of the RV850's and RV880's specifications using RV740 as a basis. What if just to scale RV740 (640 SPU, 32 TMU, 16 ROP, 128-bit bus) by:

1.5 times for RV850 (960 SPU, 48 TMU, 24 ROP, 192-bit bus) with possible VRAM: 0.75 and 1.5 GB;
3.0 times for RV880 (1920 SPU, 96 TMU, 48 ROP, 384-bit bus) with possible VRAM: 1.5 and 3.0 GB?

These formulas, of course, will lead to the somewhat die increase (~30% against of RV840 and RV870, I guess), but I hope that the TSMC's 40-nm technology will let to produce the largest of these chips, sooner or later. And I think that they will be faster by 30-40%% than their predecessors, what will let them to compete with the nVidia's upcoming rivals.

[avk]

Alas, the TSMC's 40 nm yield is about 50%, even after more than a year! This is a bad omen, if you ask me. I think it would be nice to redesign Cedar and/or Juniper to make them more cheap and/or powerful?

What would happen if we would try to overclock Radeon HD 5850 to the 5870's frequencies (core 725->850 MHz, memory 4000-4800 MHz)? You can see this. But hey, we can see that "Overclocked" 5850 is just by 2-3% slower than 5870! Does this mean that "1440 SPU and 72 TMU" of 5850 could be reduced to "1280 SPU and 64 TMU" of yet not existed 5830? I think yes. You, AMD, could redesign Cedar from "1600 SPU and 80 TMU" to "1280 SPU and 64 TMU", which could lead to the 10-11% die reduce (~300 sq.mm) and to the ~5-6% speed reduce. This new chip could be less power-hungry, and if you clock it at 900 MHz, it will be speed-equal to the current 5870.

As for the Juniper redesign I could say: that chip is not powerful to compete with 4890 and sometimes even with 4870. Is it possible to make Juniper's bus more wide, say, 192 or 256 bits? If yes, then please do it right now! If not, maybe it is time to change the chip's formula to "960 SPU, 48 TMU and 24 ROP?" After this the die size will be increased to 220-225 sq.mm, which is good, IMHO. This new chip will compete not only Radeon 4870&4890, but GeForce 275&285 too.

[bridgman]

I don't think Juniper was intended to compete with HD4870 or 4890 - that's what the 5850 and 5870 are for. AFAIK Juniper is a replacement for 4670 and 4770 with more shader power and DX11 support.

[avk]

A replacement for 4670/4770? Maybe. But what for? IMHO, AMD/ATI should strive to compete not its own previous models, but an nVidia's ones. Yes, Radeon HD 5850 is good enough to compete with GeForce GTX 285, but perhaps it's too good. I believe that not existed Radeon HD "960 SPU, 48 TMU and 24 ROP" on 192-bit bus could compete GeForce GTX 285 too. And this chip could be smaller than Cypress by 1.5 times (225 vs. 334 sq.mm) and, of course, cheaper.

Of course, I may be mistaken, but I'm waiting for some enthusiasts able to disable several units inside Radeon HD 5850/5870 to my formula in order to proof my theory. Although all we know that Radeon HD 4890 is about equal to GeForce GTX 275, so Radeon HD "960/48/24" will definitely defeat GeForce GTX 285.

[avk]

AMD, I think you shouldn't create Radeon HD 5830 as you have made it. 16 ROPs is not enough, and I think you do perfectly know this. Please change your mind and redesign Radeon HD 5830 - I mean please increase its ROPs from 16 to 32 or at least 24.

What about Radeon HD 5890? Pack it with 2 GB of GDDR5 at 5 GHz, overclock Cypress GPU to 950-1000 MHz and voila - this would be a GeForce GTX 480-killer! Of course its TDP will be increased from 188 W to, say, 225 W or even 250 W. But who cares? Some PC gamers want to buy most powerful hardware, so why not to earn some money on them? Yes, Radeon HD 5970 is still most powerful graphics card ever made, but some gamers still are afraid of multi-GPU technology.

[avk]

As you can see, nVidia GeForce GTX 460 is slightly faster than Radeon HD 5830. So why should you take any advantage to nVidia? What about Radeon HD 5840 with a formula like this: 1280/64/32? Set its frequency slightly lower than 5850's one, and you'll get a nice 460-killer.

[avk]

There is no doubt - 6850 & 6870 are very good cards, but their model numbers are not, IMHO. AMD, you should have name them 6830 and 6850 in order to comply the rule "the higher number, the better."

As for 6850 with 1120 shaders: please let it exist. Just name this model 6860 and put its price between 6850 and 6870.

[avk]

Both 6950 and 6970 are good enough cards, but their performance suffers, IMHO, due to not so high memory bandwidth. Yes, I know that 5.0 and 5.5 Gb/s are quite high numbers, but it seems to me that it is not enough to compete with GeForce GTX 570 & 580.

What about to raise their frequencies to 5.5 and 6.0-6.4 in a near future? Of course, this means that the TDP of both cards will raise, but I hope not so high. The new models (let's name them 6960 & 6980) can have a slightly higher price than their 6950 & 6970 counterparts. You can even eliminate 6950 & 6970 right after the 6960 & 6980 start, if you think that the second ones can be a rivals for the first ones.