R350 - 2 TMU's per pipe

[druga runda]

Since noone posted it's here

http://www.theinquirer.net/?article=7589




IN THE HIGH PERFORMANCE graphics industry, performance must be combined with quality. It is this neverending race which will lead to initial availability of Nvidia's new Geforce FX cards this month.
However, we have learned that ATI's next generation R350 chip which will compete with the Geforce FX is now in mass production, and will definitely be in retail stores in April. It will be officially announced at CeBIT in Hannover, Germany on March 12, 2003.

Specifications of the new chip are being kept under extremely tight scrutiny because of the tight competition between the two rivals. It would give too much lead-time to overclock one chip just to paper launch a product and claim the performance lead.

Many facts are already known. The R300 will be built on a 150 nanometer process. is basically an overclocked R300 with eight rendering pipelines and two texture units per pipeline.


Be aware that there are numerous enhancements to its design and manufacturing process, which will increase performance and yields thus lowering the price of the chip. If yields of the R300 are any indication, ATI could have yields in excess of 90% using TSMC's proven copper process, although without a low k dielectric.

We expect the chip to launch at a core clock of 400MHz. It will incorporate DirectX 9.1 features, which is great if you're going to keep the card for at least two years, but don't expect to see any games written for it released for at least a year.

ATI will brand the chip the Radeon 9900 Pro, and the card will be equipped with 128 MB memory while supporting 256MB configurations for its partners. They will use 10 layer PCBs for the graphics cards. The 128MB variations will be priced around the ?400 mark. Lower bin parts will be sold through ATI's partners at a lower speed for less. A 350W power supply would be recommended if you are using a hyperthreading P4 system. µ





Coming in April as expected and will be introduced at CeBit 2 TMU's per pipeline, 400 mhz and 800 mem.

I can swear that the article has changed somewhat since I have seen it first 1/2 hour ago, it was bold and the guy was claiming that R350 will use DDR2. I have checked my cache but all was overwritten with this never version :mad: anyway, perhaps this was too much of a speculation so the guy removed it. Anyway 800 mhz DDRI or DDRII... pretty much the same except DDRII could have bit better OC potential. But well we will see in a month.

The first iteration of the article had that "Many facts are alrady known" paragraph much longer with more facts listed when I skimmed trough it.

Anyway... can that 2 TMU's per pipeline be true?

[Steppy]

Originally posted by druga runda
Since noone posted it's here

http://www.theinquirer.net/?article=7589



Coming in April as expected and will be introduced at CeBit 2 TMU's per pipeline, 400 mhz and 800 mem.

I can swear that the article has changed somewhat since I have seen it first 1/2 hour ago, it was bold and the guy was claiming that R350 will use DDR2. I have checked my cache but all was overwritten with this never version :mad: anyway, perhaps this was too much of a speculation so the guy removed it. Anyway 800 mhz DDRI or DDRII... pretty much the same except DDRII could have bit better OC potential. But well we will see in a month.

The first iteration of the article had that "Many facts are alrady known" paragraph much longer with more facts listed when I skimmed trough it.

Anyway... can that 2 TMU's per pipeline be true? I highly doubt the 2 TMU's. The use of loopbacks and the programmable T&L made extra TMU's a LOT less valuable. I doubt you'll ever see a multi TMU card again.

[druga runda]

he he...

and now we get another Inq article that says that the card will use DDRI... I guess that was the reason for the change in the first article.

http://www.theinquirer.net/?article=7582
and apparently only 325 mhz for memory.

It seems that FX is making ATI keep the specs down a bit :(

anyway...

here is another quote

I don't need to say DDR I is less expensive than DDR II since DDR 2 is still in development and will not be introduced until early next year in motherboards. I don't need to say it but I have.

We heard that ATI is full of optimism after the first "reviews" for Geforce FX since it expected that Nvidia would deliver more than they have. But what would we know. We don't sign stinky non disclosure agreements (NDAs).

The latest skittleblatt we hear is that the R350 will be between 10 and 30 percent faster than the mighty Geforce FX and definitely less noisy, since it will use a normal cooler which appears absurd at first glance, as the chip is built using a 0.15 micron process and has 107+ million transistors.

This is the reason why Geforce FX with 500MHz DDR II memory will cost around €650. The R350 will stay at the magic limit of €400 for the 128MB version.

Hey! You get what you pay for in this world... Hopefully µ



Still 400 mhz clock should help the card and those 2 TMU's should help too if this is really true. As I hope this R350 is at least a little more than an overclocked R300. :hmmm: --- edit: well I would think that there is at least some value in multiple TMU's as there is a double increase in texel fill rate.

Still $400 RRP price is a good indication nevertheless.

[ZoinKs!]

My guess of the most likely specs for r350: 400 Mhz core clock, perhaps higher
350-400 MHz ddr1 memory
128-256 megs
beyond dx 9, perhaps support for longer shader instructions in one pass
probably 1 tmu per pipe
probably 96 bit floating point color precision, just like r300
$399 for pro version, $299 for non-pro

While 2 tmu's per pipe is possible and would probably be useful, it seems unlikely. I doubt much can be added to the chip without increasing the die size too much, which is why I don't expect much as far as new features.

The only features of note which the nv30 has but the r300 lacks are longer shaders in one pass and 128 bit floating point color precision. Neither will be vital for games during the useful lifetime of either the nv30 or r350. That's another reason I don't expect much for new features.

New features would be nice for "bragging rights" but they aren't needed for the time being. All ati needs for now is a card which will outperfom the geforce fx by a noticable margin and win the price/performance/image quality ratio.

(A card able to run 6x aa/16x af at the same fps as the 9700pro runs 4x/16x would be a nice performance target for ati to aim at. Based on the rumored specs, can anyone figure out what would be the 6x performance of the r350?)

[Hellbinder]

If I just LAY it out for you guys in plain english, will you believe me over those hokey web sites??? well.. anyway..

THE R350 IS AN 8x1

Trust me on this one..ok?

[druga runda]

Well OK... than we trust you ;)

So it is only a slightly overclocked R300 (or is it more programmable as well), sounds lame, but well not that R300 is bad :D... hopefully R300 prices will drop too a bit.

Furthermore they were speculating it will only be 375/750 card. Now that is even lamer, but well... R300 is still the leader so getting a little better than that is sufficient for them until NV35 comes out. :(

[Bigus Dickus]

I'd really expect 400/800 or perhaps even 425/850 for the retail version, and 375/750 for OEM.

With the additional time to work on tweaking the core, I also expect there to be about as much overclocking headroom in this card as is in the R300. I fully expect to see routine overclocks into the 450/900 region with the R350.

If there was one "power adder" I would like added to the R350 core, it would be the ability to do one trilinearly filtered anisotropically filtered pixel per clock per pipe. I think right now enabling trilinear (quality AF) results in two clocks per pixel... correct?

[digitalwanderer]

What is the big deal with 2TMU's per pipe? I'm not trying to be sarcastic, I just ain't knowledgable enough to understand the importance of 1 or 2 TMU's per pipe on real world gaming performance.

Since this seems like such a hotly debated feature (or not) of the upcoming R350, could one of the bigbrains please take a second and give me a thumbnail sketch of what the big deal is with it?

I was under the impression that 2 TMU's would only help out with the really advanced DX9 stuff...but I realize that I'm actually pretty clueless as to what you're all talking about. :(

[Hellbinder]

So it is only a slightly overclocked R300 (or is it more programmable as well), sounds lame, but well not that R300 is bad ... hopefully R300 prices will drop too a bit.

Furthermore they were speculating it will only be 375/750 card. Now that is even lamer, but well... R300 is still the leader so getting a little better than that is sufficient for them until NV35 comes out.


I can simply not comment on exact clock rates. But let me just say... you wont be dissapointed by the results.

What you will be getting with the R350 is a card that is enough faster than the Nv30 to warrant chosing it over the Nv30. That uses much less power, with much less heat, a normal cooling solution, only takes up one slot. Some improved/new features over the R300. And the real awesome benefit will be usable 6x FSAA in all games (i use 6x FSAA in most games now except a few like UT 2k3 etc. and Simply SMOKIN 4x FSAA.

I will go this far.. Overclockers are going to have a hayday with this sucker depending on where they have finally settled the core speed At. I know it is *capable* of a lot more than what people are expecting, to the point of being shocked i think. But, I honestly dont know what direction Ati has decided to go with the final clock. They dont want it getting leaked at all for obvious reasons. I only know what it was running at in their lab over a month ago.

[Steppy]

Originally posted by digitalwanderer
What is the big deal with 2TMU's per pipe? I'm not trying to be sarcastic, I just ain't knowledgable enough to understand the importance of 1 or 2 TMU's per pipe on real world gaming performance.

Since this seems like such a hotly debated feature (or not) of the upcoming R350, could one of the bigbrains please take a second and give me a thumbnail sketch of what the big deal is with it?

I was under the impression that 2 TMU's would only help out with the really advanced DX9 stuff...but I realize that I'm actually pretty clueless as to what you're all talking about. :( Ok, this isn't to scale but it's an illustration of the benefits of TMU's.

As a scene makes a "pass" through the pixel pipeline, each TMU can apply a texture at the same time...for easy math say that a pass takes 100 clock cycles and the TMU slapping the texture on takes 30 of those clock cycles.

In a scene with 2 texture layers...

Card A with 1 TMU will apply a texture, send that back through pipeline again, slap the 2nd texture on it, then send it out the framebuffer to your screen this would be two "passes" to render the scene or 200 clock cycles.

Card B with 2 TMU's each will apply a texture at the same time and then send the scene out. 1 pass 100 cycles.



Now cards nowdays can do "loopbacks" meaning a 1 tmu card can apply one texture(one pass 100 clock cycles), then the same TMU can apply another texture by keeping the scene at that point in the rendering pipeline without sending the scene back to the beginning of the pipeline. It just adds the TMU part of it on(30 cycles). so this card renderes the scene in 130 clock cycles. While not as fast as a true 2TMU card it's much faster than a 1TMU card that can't do loopbacks. The slight speed loss is made up by the die space saved that can be used for other things.

[digitalwanderer]

It'll take me a bit to digest that all, but I think I get the gist. Thank you! :)

[Lezmaka]

I remember reading somwhere the other day that the R350 would use copper interconnects. If its true, would that switch (along with hand tweaking) be part of the reason why it will be able to hit high clocks? Anyone know if R300 was copper, or aluminum (or whatever R200/nv2x used)

[Gar]

Originally posted by Lezmaka
I remember reading somwhere the other day that the R350 would use copper interconnects. If its true, would that switch (along with hand tweaking) be part of the reason why it will be able to hit high clocks? Anyone know if R300 was copper, or aluminum (or whatever R200/nv2x used)

R300 was Aluminum. R350 will be copper (AFAIK).

[LORD-eX-Bu]

Originally posted by Hellbinder
If I just LAY it out for you guys in plain english, will you believe me over those hokey web sites??? well.. anyway..

THE R350 IS AN 8x1

Trust me on this one..ok?

Eh, I don't beleive you, or those websites. So no thanks, no trust ;)

[-=DVS=-]

Well it doesn't make alot of a difference if R350 will have 1 TMU or 2 , moust importent thing is it will be faster then R300 and NV30 :D with bugs fixed/tweaked/ better AA speed and perhapse IQ ;)

[AngelGraves13]

Originally posted by Hellbinder
If I just LAY it out for you guys in plain english, will you believe me over those hokey web sites??? well.. anyway..

THE R350 IS AN 8x1

Trust me on this one..ok?

Sorry for asking stupid questions....but doesn't this make it the same specs as the R300 mem?? I don't know too much about what 8X1 means, but if I remember correctly it's the same as the R300??

[Uttar]

Originally posted by Gar
R300 was Aluminum. R350 will be copper (AFAIK).

Hmm, that's probably where *most* of the R350 clock rate increase will come from. Remember, nVidia said most of its high clockrate for the GFFX came from copper too.


Uttar