First NV31 Announcement Rumors...

[MuFu]

The clockspeeds of the two NV31 boards in development right now are 350MHz and 450MHz (~Ultra version). If that isn't true I'll eat the 12" subwoofer I just bought.

That is pretty useless information without knowing the core functionality, of course. Sorry I don't know much more. I would be VERY suprised if the ASIC is based on the NV28 though.

MuFu.

 

[Joe DeFuria]

Hi MuFu...

Where are those ATI daughtercards, anyway?

In any case, if what you're saying is true, we'd have:

NV30 "Ultra" at 500 Mhz.
NV30 "Light" at ??
NV31 "Ultra" at 450 Mhz
NV31 "Light" at 350 Mhz.

Strange...doesn't seem to be room for the NV30 "Light" in there, unless:

1) NV31 is 4 pipe board, not 8 pipe. (Or maybe 6 pipe?)
or
2) NV30 Ultra is more like 550-600 Mhz.

I assume number 1 is the case. Don't see a real reason for a different chip production if it's the same number of pipes as the NV30.

Further assuming that the memory for each config is essentially synchronous, and 128 bit DDR(I or II), we'd end up with the following in performance by my estimation...from lowest to highest in the $150-$250 segment:

1) Radeon 9500 non-pro
2) NV31 350 Mhz
3) Radeon 9500 Pro / NV31 450 Mhz (Both win and lose depending on scenario)
4) Radeon 9700 non-pro

 

[Ante P]

I'm just not sure that I see room for three separate NV3x chips, all with similar levels of feature (API) support, and selling within $200-$300 of one another. Not sure nVidia's OEMs would appreciate that.

Right now we have
GeForce 4 Ti 4200 64 MB
GeForce 4 Ti 4200 64 MB with AGP 8x
GeForce 4 Ti 4200 128 MB
GeForce 4 Ti 4200 128 MB with AGP 8x
GeForce 4 Ti 4400
GeForce 4 Ti 4800SE/Ti 4400 with AGP 8x
GeForce 4 Ti 4600
GeForce 4 Ti 4800/Ti 4600 with AGP 8x

Some of the more exlusive GeForce 4 MX 460 and 440 with AGP 8x (and even one or two 440 and 440SE) also come barging in at the same pricerange as the cheaper Ti4200.

Oversaturated... over 20 cards in the GeForce 4 series, that's WAY too much, especially since they don't differentiate enough to validate their existance at all.

I sure hope we don't see a similair situation with the nv3x as with nv17, 18, 25 and 28

 

[CMKRNL]

My info indicates that it will be a 4x1 128bit DX9 part with clock speeds not yet finalized. MuFu's predictions may well be correct though, they sound in-line with expectations.

 

[Hellbinder]

NV34 = cut-down NV30 wich will compete against RV350.

Give me a freaking break..

Nvidia is going to announce a card/chip that is faster and better than their Nv30 Ultra part before/within weeks it actually hits the street.

 

[KimB]

In the past, nVidia has always used two chips, with multiple speed grades / memory configs on each, to cover the three markets.

Not always. The number of chips that nVidia is working on at the same time has been continually increasing.

I'm just not sure that I see room for three separate NV3x chips, all with similar levels of feature (API) support, and selling within $200-$300 of one another. Not sure nVidia's OEMs would appreciate that.

It shouldn't matter. From the OEM's perspective, it's no different from purchasing different speed-binned chips, provided they are pin-compatible, which is highly likely.

 

[KimB]

]

NV34 = cut-down NV30 wich will compete against RV350.

Give me a freaking break..

Nvidia is going to announce a card/chip that is faster and better than their Nv30 Ultra part before/within weeks it actually hits the street.

A chip termed "RV350" would be aimed at the mainstream market. It would probably be close in performance to the Radeon 9500.

 

[MuFu]

Hi MuFu...

Where are those ATI daughtercards, anyway?

Here mate, here:

http://www.compd.co.kr/korean/products/ati.htm

Right... so we're getting somewhere now. CMKRNL always seems to know what's going down in the ghetto. DX9, 128-bit, 4x1 sounds suspiciously like NV34 to me but I guess that if it were NV31, with a high clockspeed and enough clever chutzpah it'd meet the performance estimates currently flying around. Maybe the boards use DDR-II (?).

As for the NV30 "standard" - surely the "missing" clockspeed is 400MHz, Joe? Assuming it is architecturally superior to the NV31 then that would place it in between the NV30 Ultra and the NV31 Ultra in terms of performance (?). Generally speaking of course... hard to account for all circumstances.

To be honest I wouldn't be suprised if the faster of the two NV31 parts is competitive with the 9700 (amateur). Presume NV30 is as fast as nVidia claims it is and there is a neccessity for the R350 - I know things won't all line-up nicely, but I'd assume that a 400MHz NV30 would be competetive with the 9700 Pro, the 450MHz NV31 with the 9700, the 350MHz NV31 with the 9500 Pro etc etc... below that, the NV28 and the stockpile of NV25's will continue to sell. The NV35 is too far off to fit into the puzzle at this stage (although obviously it is the next high-end part).

As for RV350 - I can see the fastest version of this treading on the toes of the already-redundant, rather costly Radeon 9500 (amateur).

MuFu.

 

[snk]

Here is an interesting image:

It seems that the NV34 is the 4x1 pipelined chip, and NV31 maybe nothing more than an NV30 with lower clockspeeds? Maybe there will be no NV30 "standard" at all. At least we can be sure they're both DX9 chips.

 

[Ailuros]

The NV31 is expected to utilise numerous technologies introduced in the GeForce FX VPU and also offer comparable performance to the GeForce4 Ti4600 with antialiasing and anisotropic filtering enabled. Generally speaking, the NV31 code-named products will compete with the RADEON 9500 PRO and 9700-based solutions, so, it is very logically for them to be cheap enough.

Unless there are some suprises we haven't seen so far, Intellisample will include only minor modifications to the AA/aniso algorithms already present in NV25 (see colour compression for MSAA).

I don't know if xbitlabs has any reliable sources or not or makes lucky guestimates, but it depends on the viewpoint you take on things like that.

I'd guestimate that core logic related to aniso (see fillrate impact with bilinear and trilinear the same) present in NV25 remains the same in future products, and we've seen so far the hybrid MS/SS antialiasing modes, that are fully operative on former products too.

A 4x1 400MHz NV3x, has a theoretical Texel fillrate of 1600MTps; with colour compression and some other occlusion refinements Multisampling performance will be most likely quite a bit higher, while the hybrid MS/SS modes don't stand that much of a chance to see any performance improvements.

If s.o. would tell me that a NV3x (1 or 4) performs almost the same as a Ti4600 with 8xS, I wouldn't be that much surprised.

 

[Joe DeFuria]

Nice find.

So, assuming the following:

* NV30 = 8 pipe, DX9, 400 and 500 Mhz clock variants
* NV31 = 4 pipe, DX9, 350 and 450 Mhz clock variants

Those assumptions make sense given the road-map. What's very strange to me is having the Ti-4200 in there in the low end of the performance segment. A DX8 card sandwiched bewteen two DX9 cards?

The only way that's a viable product is if the NV34 performs notably worse than the Ti-4200, since the NV34 obviously beats it on features.

Which means that we can make some guesses as to what the NV34 is, I'll throw out these two:

1) 4 pipe DX9, clocked notably slower than NV31 and NV28.

2) 2 pipe (perhaps 2x2, like GeForce4 MX) DX9, clocked about the same as Ti-4200.

I'd lean toward option 2. I would guess that NV34 would probably only have a DDR-I memory interface, wheras the NV31 would likely support both DDR-I and DDR-II.

 

[Dave Baumann]

These roadmaps are a little screwed up because of the late arrival of NV30.

Consider previous launches of their low end products. The new architectures are usually launched on their own and then the rest of the line is fleshed out with the refresh. This is not the case with NV31/34 because NV30 is late, and NV31/34 is coming out relative to when they initially had NV35 slated for.

So, it could be the case that you shouldn't be thinking of what NV31/34 is in relation to NV30, but what you think NV35 may be like.

It also rasies the question - does this run NV35 off the roadmap entirely?

 

[antlers]

So the NV34 is not an integrated part? Or is it the integrated NV31, cheaper because it is part of the chipset and slower because it is tied to slower memory?

 

[MuFu]

Um, I don't think any of them are integrated parts. I believe this confusion first arose when Anand mentioned something about the next nForce chipset integrating the NV34 graphics core. I think I recall the article saying something about it being based on NV30 technology - obviously that contributes a little to this discussion. Maybe is was worded wrongly or was essentially incorrect but from then on many people started thinking that the NV34 was an integrated chipset. Past chipset codenames haven't been numeric anyway - Crush, Crush18 etc, so I don't why that should suddenly change. I think we'll see "Crush 34" (blah!) with the NV34 forming the basis for the video side of things.

MuFu.

P.S. I may have some consolidating information to add very soon...

 

[Ailuros]

Dave,

I sure hope that reality in 2003 will not include NV18. In any other case they're out of their mind.

 

[MuFu]

*w00t!*

NV30 - Definitely targeted at R300 and R350.

NV31 - DX9, bastardised version of NV30 core, brand new DDR-I interface. Apparently, a 400/350MHz version could be competitive with the Radeon 9700 (amateur). I imagine the 450MHz clockspeed I have mentioned in the past is prohibitive when it comes to complimentary memory expenditure (it'd make no sense to equip such a card with anything less than 400MHz DDR-I). It seems the clockspeeds aren't fixed yet after all CMKRNL, so I will be eating subwoofer.

NV34 - DX9 "compatible" (urgh!), mega-bastardised NV30 core (some functional blocks removed and compensated for in software), NV18 memory interface, targeted at RV250/RV350.

~~<Edit - got my wires crossed here initially, sorry - now corrected>~~

The NV31 is a POS from what I gather - there are many software shader hacks to compensate for the loss of native functionality and reduce execution speeds on CPU time. In addition, parts of the ASIC itself are inherently buggy so the card is a bit of a "problem child" at this stage. :?

OAO,

MuFu.

 

[Doomtrooper]

 

[Joe DeFuria]

NV34 - DX9 "compatible" (urgh!), mega-bastardised NV30 core (some functional blocks removed and compensated for in software), NV18 memory interface, targeted at RV250/RV350.

If that is true...I really, REALLY hope that at least all DX8 shader functionality is done in hardware.

If RV350 is a direct compeitor to NV34 (cost wise), and RV-350 is a more "pure" DX9 part...nVidia could be in for some trouble. These two chips will likely be the basis for a very large chunk of revenue and volume for each of these companies for the next year to 18 months, and if one of the chips turns out to be perceived as notably inferior, there's going to be some major profit hits...

 

[KimB]

NV34 - DX9 "compatible" (urgh!), mega-bastardised NV30 core (some functional blocks removed and compensated for in software), NV18 memory interface, targeted at RV250/RV350.

Strange...nVidia hasn't done this yet. ATI has, and so have other companies for their low-end parts. Nobody has seemed to like the idea. It would be odd if nVidia suddenly thought it was a good idea.

Hopefully, at its worst, it's just a weaker version of the vertex shader engine, not a castrated version (i.e. fewer execution units, as opposed to none).

 

[RussSchultz]

I'd rather they remove it entirely if it saves the pixel shaders from getting castrated.

I don't believe that the vertex shaders in software have sucked too terribly bad in comparison to the hardware versions. (But I may be wrong).