crappy 3dmark score on NV31/34?

Exactly. One of the advantages of the GF3 compared to an NV34 is that the GF3 is 4x2 while the NV34 is perhaps 4x1 but more likely 2x2 with 4 zixels/clock.

 

Probably something like:

"NEW GEFORCE FX 5200 ULTRA!!!

• FULLY DX9.1+ COMPLIANT WITH CINEFX TO CREATE BREATHTAKING NEW WORLDS OF NEVER-BEFORE-DREAMED-OF REALISM!!!
• "4"* "PIXEL"* PIPELINES AND FLEXIBLE NEW "PUDDLE-OF-SHADERS" PIXEL ENGINE USHER IN THE DAWN OF CINEMATIC COMPUTING BY LIMITING ALL YOUR FAVORITE GAMES TO 24 FPS FOR ENHANCED CINEMATIC REALISM!!!
• ENHANCE SCENE GEOMETRY WITH ALL-NEW HYPER-PIPELINED, HYPER-THREADED VERTEX SHADERS!!!**
• INCLUDES BONUS DVD OF AMAZING NVIDIA CINEFX DEMOS IN GORGEOUS FULL-SCREEN, FULL-MOTION MPEG2 TO DEMONSTRATE THE AWESOME POWER OF YOUR NEW GEFORCE FX 5200 ULTRA!!!"

*This statement is false.
**With participating CPU.

 

Bwhaha, that first * was classic!

 

MuFu.

 

LOL, I didn't laugh that much in *weeks*!


I've got a better idea for the demos... I suggest to put them in black & white

MuFu: You already said the NV31 had 6 Pixel Shading units and the NV34 had 4 Pixel Shading Units. But what are zixels & pixels/clock numbers? Are they 6 & 3 on the NV31, and 4 & 2 on the NV34? So, the NV31 could be called a 3x2?

Dave H: Yes, but you've got to remember the NV34 has a small clockspeed advantage. The original GF3 was launched at 200Mhz. The NV34 will be launched at either 250Mhz or 300Mhz.
So, the NV34 would be 47.5% slower than the original GF3.
That is, if you don't count the slightly higher memory bandwidth, better compression techniques and probably better efficiency.
What you can be pretty much certain of, however, is that in DX8 games, the NV34 will be slower than the original GF3!


Uttar

 

Not sure to be honest - 3 coloured pixel output per clock just seems plain *wrong* to me, but we'll see. My own personal guess is that NV31 will be 6/4 (not quite sure how efficient that would be) and NV31 will be 4/2. 2 pixels/clk sounds awfully low for a 250Mhz part though, doesn't it? :?

Bah - I dunno... Can't say I fully understand the organisation/associativity of NV3x rendering functionality at all.

MuFu.

 

Yeah, this was all in response to my post comparing mr's GF3 @ 230/230 to an NV34 (I was using Chaintech's clocks of 250/200). The clockspeed difference is mentioned..

 

and you caught me Dave H, i compleaty missread what you were saying there.

 

I retested my GF3Ti200 with 240/400 and a XP2400+ with forced Software Vertex Shaders and scored ~855 in 3DMark03.
I'm trying to overclock my CPU and see how much it effects the score. GT1 seems very slow with Software Vertex Shaders and should scale very well with CPU clock.

I hope this NV34 score is bogus because even if it runs about the speed of my GF3 in GT1, GT2, and GT3 and crawls through GT4 with 5fps, it should get 200 points more than my card.

Please note that I'm not judging any card on it's 3DMark score alone (and on an unconfirmed score too), but I think these are interesting observations (and speculation is fun ).

 

Well, FWIW, 250/200 would make it essentially:

NV34 = (original) 128MB GF4MX 440 - 20MHz + double-rate zixels + pixel shaders + AGP 8x

Or to put it another way:

NV34 = 2*(TNT2 + 50MHz DRAM clocks) + crossbar memory controller + 96MB + DDR-Z + PS + AGP 8x + hardware T&L (presumably)

So, to answer your question...yeah, a bit.

But it all depends on how serious Nvidia is about really moving NV3x across most of their product line. The $99 price target seems a bit high for street price of a 250/200 NV34; after all, that original 128MB MX 440 is in the mid-$50's at pricewatch, and the new 8x version (mid-$60's) is a substantially bumped 275/250 (albeit 64MB). If a 250/200 NV34 goes for $80 street, it could fit in well.

It's a very interesting question which will be more popular with the OEMs: such an NV34 or a 9200 with better performance, less CPU dependance, and probably a lower price tag...but no DX9 or Nvidia brand.

 

Dave,

probably 350/350 NV34 is another actual freq

so whats about pipes

is here 4х1 for 31 and 2х2 for 34 or

2х2 for both but with unsacrifaced 31 and sacrifaced 34 (no somewhat compression, no vertex shaders)?

 

Just wanted to comment that running with software pipeline processing is not exactly the same as the driver implementing software shaders. Even when the shaders are done in software in the driver, clipping and viewport transformation are done on the card, while the DX software vertex processing does them in software. There is some speed benefit in this, although more for simple shaders than complex ones.

Of course, that just means that the NV34 is even slower by comparison.

 

true 4x1 (not NV30 '8x1! )

 

NV31 and NV34?!

Welcome news in any case!

 

NV34 can run the Game Test 4 of 3Dmark04, it support PS2.0 in hardware.

 

No one should care if it's chugging along at single-digit framerates. Seems like nV may go from one technical extreme to the other (GF"4"MX -> NV30"MX"), with the same amount of marketing. :\

 

See that tallies with what CMKRNL said last year around the time of tape out. NV34 is obviously significantly more bastardised than NV31 though (which retains much of the Z-trickery, colour compression etc of NV30).

MuFu.

P.S. "John Smith" @ The Inquirer. LOL!

 

lol @ DDR-Z...

gotta love the various ways we need to come up with to describe the way things are designed in the NV30.

I can see the next marketing paper...

features:
...
DDR Z rendering with 3200MZixels/sec!
...

 

What does it say in the DX9 SDK regarding VS modes? Is a pure software mode even permitted (as in DX8.1)?

MuFu.

 

The SDK says:

"Software shaders have been implemented to facilitate shader development without hardware support. The following software versions have been added: vs_2_sw, vs_3_sw, ps_2_sw, ps_3_sw."

Edit: Others will certainly correct me on this but I guess it's still the same as with DX8 though for software vertex processing - use your application to send pre-TLC'd vertices to a vertex buffer. To my (poor) eyes, it would appear to be the same for all vertex shader revisions.