More R350 Details

[Tahir2]

What's the point of asking questions you don't know the answer to? Of course there is no point, that is plain silly! (See.. I only asked this question because I knew the answer already).

Anyways the thermal characteristics of DDR I compared to DDR II are... [censored..censored..censored]....and that's why DDR II has lower voltages as well. Hope that makes it clear for all of you.

 

[martrox]

CMKRNL said posted that 4 days ago.....

DDr II has higher latiency than DDr, and has higher PCB requirements...


The GFFX heatsinks covers both sides of the card, covering all of the DDr II chips ..... why?

talk about denial....... basicly, I doubt there's anything you could be shown that would sway your opinion, is there?

 

[Tahir2]

Right without any knowledge on the subject you would expect heat dissipation to be less on DDR II since they expect it to eventually clock higher.

I mean it is possible to wage war on circumstantial evidence, why not wage a victory for common sense? Answer - common sense is not so common.

 

[martrox]

By the way here's an interesting quote from Samsung - makers of the GDDR-II used on the GFFX:

Samsung GDDR roadmap revealed the availability of 800/1000Mbps GDDR-II (128MB) now which NVIDIA GeForce FX Ultra 5800/Ultra will be using and in second half of this year, we will expect to see 1200 to 1400Mbps GDDR-II (256MB) which most probably NV35 will be using. The VDD of the GDDR-II will be reduced from 2.5V to 1.8V somewhere along the roadmap so the memories to run much cooler as a result. The next generation GDDR-III is also plotted in the roadmap and we can expect to see them in second half of 2004 with speed of 1.5 to 2.xGbps. The GPU frequency is expected to reach a stunning 900Mhz next year so GDDR-III will be required to provide enough memory bandwidth.

And so, memory heat is not a problem?
GDDRII, as it is right now, costs approx. 50% more(as per CMKRNL), runs too hot(as per Samsung - hence lower voltage will run much cooler, has higher latiency(from many sources) and requires a more complex PCB(from many sources).
Come on, guys, admit it. DDRII, as we now know it, is not ready for prime time, not too much different than .13 was just a few months ago. Does that mean it won't get better? NO! But as it stands right now, it's a technology that needs a bit of work. It will take a bit longer - just as the .13 die reduction did - to get right. And then it will take awhile for the price to come down so that it can be used in budget cards.

 

[PiNkY]

Why is everyone doubting the possibility of 8 texel sampling units per pixel pipeline for R350. I doubt many people still prefer biliniar over triliniar filtering. Also, I don't see that as being too bandwidth consuming as you propably have seperate texture caches for at least two mipmap levels on modern graphics chips. Thus even with 4 sample units you have to have (I am no native english speaker 8) the data in your on-chip caches to render efficiently (or do you really think your R300 flushes its texture cache every cycle to load a different mipmap?) so this basically will not increase your bandwith requirements, but will just make trilinear free and give everyone a serious increase in anisotrophic f. performance. Also, refering to these as "TMUs" nowadays seems a bit anachronistical to me.

 

[RussSchultz]

Hmmm, I was under the impression that Vdd of GDDRII was already 1.5V.

If its still at 3.3, then its not too suprising it runs hotter. (Though I was also under the impression that the memory arrays inside ran at 1/2 whatever DDR-I did, and were simply addressed out half at a time, but twice as fast)

 

[Joe DeFuria]

Why is everyone doubting the possibility of 8 texel sampling units per pixel pipeline for R350.

That is a good question. The reason why most people doubt "2 TMUs per pipe", is because it is believed to be a significant architectural change from the R-300, which most tend to assume won't happen, givin both the constraints of time and transistor budget on 0.15u.

That being said, it would not surprise me at all if R-300 turns out to be more or less what you are proposing: Double the texel sampling units per pipe, though expressly limited to filtering capability. Especially if we consider the R-300 might already have the transistors there for it, but it might not be enabled or completely functional.

Consider the difference between GeForce256 and GeForce2 GTS. AFAIC, the GeForce256 had a "broken" or just not fully implemented 2nd TMU per pipe. The GeForce256 could do 8 samples per pipeline, "with only 1 TMU per pipe." The GeForce2 could also do 8 samples per pipeline, even though it had "2 TMUs" per pipe. Seems to me that the GeForce256 essentially had 2 TMUs per pipe also, but that 2nd TMU was not completely functional (intentional or not), and was limited to just reading in texels for Trilinear, as opposed to being a fully flexible and independent unit.

This might apply to R-300 similarly: The extra texel reading units might be there already, just not "functional." I still do NOT expect to see a "2nd TMU" per pipe, meaning that each TMU is fully flexible and in theory would (for example) double the multitexture fill rate test on 3D Mark 2002.

I do accept as possible, however, that the R-350 might be able to do exactly as you propose: enable double the texel reads for increased trilinear / anisotropic performance. That would very nicely explain how there are two camps that claim seemingly contradictory specs: 1 TMU or 2 TMUs per pipe.

Exactly because the definition of "TMU" is not clear....

Also, refering to these as "TMUs" nowadays seems a bit anachronistical to me.

Right.

 

[Fuz]

Just a quick question. What advantages would 2 TMU's per pipe bring to the consumer. I am talking real world benefits.

I know that adding an extra TMU per pipe would increase fill rate, but what else? How would AA and AF be effected by the extra TMU?

 

[Ante P]

Just a quick question. What advantages would 2 TMU's per pipe bring to the consumer. I am talking real world benefits.

I know that adding an extra TMU per pipe would increase fill rate, but what else? How would AA and AF be effected by the extra TMU?

aniso performance would get a boost for an example
it should also prove at least somewhat valuable when using SSAA