[speculation] NV35

[Uttar]

Hey everyone,

Considering all the rumors, the stuff, the thingies - you get the point - about the NV35, maybe some of you wonder what it is, really.

Considering I've been following all of this a lot, here is what I expect for the NV35:

Please note that this if for one model. I'd expect this model to retail at $399, but I wouldn't be surprised if there was a $299 or $499 model too.

- 500Mhz core, manufactured on a more mature TSMC's 0.13u copper process.
- 400Mhz DDR-I, for lower heat and cost.
- 256-bit memory bus, for doubled bandwidth.
- Pretty much equal per-clock VS power, possibly slightly increased, if so, probably mainly for things like Branching ( although I think drivers will do most of the job there, though )
- Exact same AA IQ *sigh*
- Not too modified AF implementation, some driver tweaks for the whole NV3x series coming with the Det 50 according to some not-particularly-reliable rumors on forums, but it's still quite likely it'll be mostly a software modification.
- Shadow Volume Accelerator, for slightly improved shadow volumes ( such as in Doom 3 ) speed, don't hope for 300% improvements though :P This is what was asked for by Carmack and that got in the NV35.


As for the PS... This is slightly more complicated.
As you may know, the NV30 got native Integer ( 8/clock ) and FP ( 4/clock ) units who can work in parallel. Textures, however, cannot be run in parallel to float operations - on the R300, they can.
There are two possibilities for the NV35:
1. There's two FP/Texture units, so you potentially got up to 16 textures/clock or 8 FP ops/clock ( unlikely, practically impossible considering a 130M transistor part )
2. The FP & Texture units are decoupled - that means FP & Tex ops can run in parallel, but you've still got less FP power than the R300/R350 who got 8 ops/clock. Performance, compared to the NV30, should still be well improved though ( very likely, hinted by a few people in some forums )


Uttar

[Sazar]

Originally posted by Uttar
Hey everyone,

Considering all the rumors, the stuff, the thingies - you get the point - about the NV35, maybe some of you wonder what it is, really.

Considering I've been following all of this a lot, here is what I expect for the NV35:

Please note that this if for one model. I'd expect this model to retail at $399, but I wouldn't be surprised if there was a $299 or $499 model too.

- 500Mhz core, manufactured on a more mature TSMC's 0.13u copper process.
- 400Mhz DDR-I, for lower heat and cost.
- 256-bit memory bus, for doubled bandwidth.
- Pretty much equal per-clock VS power, possibly slightly increased, if so, probably mainly for things like Branching ( although I think drivers will do most of the job there, though )
- Exact same AA IQ *sigh*
- Not too modified AF implementation, some driver tweaks for the whole NV3x series coming with the Det 50 according to some not-particularly-reliable rumors on forums, but it's still quite likely it'll be mostly a software modification.
- Shadow Volume Accelerator, for slightly improved shadow volumes ( such as in Doom 3 ) speed, don't hope for 300% improvements though :P This is what was asked for by Carmack and that got in the NV35.


As for the PS... This is slightly more complicated.
As you may know, the NV30 got native Integer ( 8/clock ) and FP ( 4/clock ) units who can work in parallel. Textures, however, cannot be run in parallel to float operations - on the R300, they can.
There are two possibilities for the NV35:
1. There's two FP/Texture units, so you potentially got up to 16 textures/clock or 8 FP ops/clock ( unlikely, practically impossible considering a 130M transistor part )
2. The FP & Texture units are decoupled - that means FP & Tex ops can run in parallel, but you've still got less FP power than the R300/R350 who got 8 ops/clock. Performance, compared to the NV30, should still be well improved though ( very likely, hinted by a few people in some forums )


Uttar

you wouldn't happen to know what samsung is able to run their ddr II memory to now would you ?

I do know hynix had developed some high clock memory a few months back... didn't know if samsung had gone with higher clocks for its ddr II memory...

[Fotis]

What are your thoughts on crossbar Uttar?
4x64 or 8x32?

[Lezmaka]

I know I'm not Uttar, but I'm almost sure it'll be 4x64. 8x32 would just be really complex.

[Lezmaka]

Originally posted by Sazar
I do know hynix had developed some high clock memory a few months back... didn't know if samsung had gone with higher clocks for its ddr II memory...

Samsung has GDDR 2 in mass production at 500 (1000), 450 (900) and 400 (800) mhz.

The website also says "Engineering Sample(Q2.'03)" for 256megabit chips. Speeds are 1.4, 1.6, 1.8, and 2.0ns. That translates to 714, 625, 555, and 500mhz. That'll probably mean 700 (1400), 625 (1250), 550 (1100), and 500 (1000) mhz.

[Uttar]

As Lezmaka said, it's 64x4. nVidia has relatively good experience with 64-bit bus lengths - it's what they use in the NV31, NV34, GF4MX, and a lot of other products. Okay, so it's 64x4 and not 64x2, so it's slightly more complex, but it's still pretty much the same idea.

Also, the expected product name is "GeForce FX 5900" - no real proof of that, though.


Uttar

[Lezmaka]

Hmmm, if 5900 for nv35, the maybe 5700 for nv36?

I wonder how nv36 will compare to nv30 non-ultra performance-wise. They could be quite similar if some nv35 rumors are true and nv36 ends up being 1/2 a nv35. 128bit buses, maybe 400mhz for both (just considering nv31 is 350mhz), maybe similar # FP/cycle though that part i'm not sure about.

[Uttar]

First, forgot to note that I expect the NV30's register usage problems to be reduced on the NV35, but it wouldn't be 100% free yet - heck, it's pretty much a design choice, I think. The effect was just exagerated.

My guess for the NV36 is: ( and this is really much more of a guess - it's based no nearly no serious info at all )

That would be the high-end model, BTW.

- 450Mhz+ core ( remember, the NV31 original target was 450Mhz - I'd be surprised if they missed their target *again* - although the original NV35 target was 625Mhz too... Ah well... )
- 350Mhz memory on a 128-bit memory bus ( 64x2 ) - maybe slightly more. I'd guess we might see a GDDR-2 for notebooks version too, since the NV35 heat problems are supposed to be fixed and GDDR-2 got lower power costs IIRC.
- On the PS POV, it'd be the half of the NV35 - so it's kinda like the NV31, but with decoupled textures/FP units and maybe slightly higher register usage efficiency.
- On the VS POV, I'd guess half the per-clock power of the NV35, maybe slightly more.
- Shadow Volume Accelerator

I'm wondering if nVidia could do some other minor tweaks to it - heck, it's to tape-out nearly 5 months after the NV35! Most likely not though - but who knows!

The Inquirer seems to believe the NV36 got identical per-clock VS power - I'd be surprised by that, but who knows...


Uttar

[mikechai]

Originally posted by Uttar
Also, the expected product name is "GeForce FX 5900" - no real proof of that, though.

Uttar

Could it be GeForce FX2 5900 ultra ?

[mikechai]

Also, the 4200 3DMark03 score is not a good sign ....

I expect it to score > 7500 in 3Dmark03 for the ultra part.
If the NV36 performance is half of the NV35 then NV36 will score
only about 3800, still not a good score....

[Lezmaka]

Originally posted by mikechai
Also, the 4200 3DMark03 score is not a good sign ....

I expect it to score > 7500 in 3Dmark03 for the ultra part.
If the NV36 performance is half of the NV35 then NV36 will score
only about 3800, still not a good score....

How about we get someone to underclock their GFFX to 300 core and since we don't know how fast the ram on that one was running, maybe clock the memory to 300, 400 and 500 and see what they get in 3dmark03.

[mikechai]

Originally posted by Lezmaka
How about we get someone to underclock their GFFX to 300 core and since we don't know how fast the ram on that one was running, maybe clock the memory to 300, 400 and 500 and see what they get in 3dmark03.

This is a good idea. Too bad I haven't got the FX5800.
Please, anyone ?

[volt]

:eek:

[GlowStick]

Originally posted by mikechai
Also, the 4200 3DMark03 score is not a good sign ....

I expect it to score > 7500 in 3Dmark03 for the ultra part.
If the NV36 performance is half of the NV35 then NV36 will score
only about 3800, still not a good score....

How so, my 9800pro at 400/420ish gets 5500.

Then the NV30 Ultras get 6000+, id expect close to 7000

[Dazz]

Originally posted by Lezmaka
How about we get someone to underclock their GFFX to 300 core and since we don't know how fast the ram on that one was running, maybe clock the memory to 300, 400 and 500 and see what they get in 3dmark03. Won't work as it is more then likly it will have half the pipelines then the FX.

[Phyre]

So, what's the dealio with the NV35? Will this thing be faster than the NV30? Just curious. I'm wanting to get a new card, but just want to know if it's worth waiting for the NV35.

Phyre

[Uttar]

Originally posted by Phyre
So, what's the dealio with the NV35? Will this thing be faster than the NV30? Just curious. I'm wanting to get a new card, but just want to know if it's worth waiting for the NV35.

Phyre

If the NV35 isn't faster than the NV30, then I'm gonna have to seriously question WTF is happening! Everything point towards a 25%-80% performance boost, depending on situation.


Uttar

[volt]

Originally posted by Phyre
So, what's the dealio with the NV35? Will this thing be faster than the NV30? Just curious. I'm wanting to get a new card, but just want to know if it's worth waiting for the NV35.

Phyre

I'm pretty confident it's worth it.

[borntosoul]

Originally posted by mikechai
Could it be GeForce FX2 5900 ultra ?


it will be called geforce FiX :)

[GlowStick]

Originally posted by Phyre
So, what's the dealio with the NV35? Will this thing be faster than the NV30? Just curious. I'm wanting to get a new card, but just want to know if it's worth waiting for the NV35.

Phyre

If you dont have an urgent need to upgrade now waiting only has perks, lower prices and/or better hardware. Truthfully i should of waited, when i fire up StarCraft or Half Life, it plays/looks exactly like my old card : P

[Lezmaka]

The all-knowing INQ has posted some details of NV35...
http://www.theinquirer.net/?article=9174

Says Ultra will be 450/900 DDR1 and non ultra will be 425/850 and that the name *could* be GeForce FX 6800.

Don't know about anyone else, but a 25/50 mhz difference doesn't really seem like a lot. I mean there's about a 10% difference between Ti 4400 and 4600. With the supposed non-ultra speeds, it's not even a 6% difference, and 6% sure isn't noticeable. I would expect at least 400/400 or maybe even 375/375 to differentiate the two more and get more usable chips out of it as well.

[GlowStick]

If they ship at 425core, i see alot of Overclocking possibilitys, Maybe a FlowFX would fit on there. And gainward already said that the CoolFX will fit on it : P probly get it to 550ish or more with that!

[Uttar]

Well, the prices are probably $399 and $299.

The 25/50 difference is probably correct. Remember the Ultra got 256MB of memory while the Regular got only 128MB
What type of RAM would they use for the Regular, though? The same 2.2? Considering the amount The Inquirer claim they bought, it would make sense.

Although using 910Mhz capable RAM on 850Mhz sounds like waste - unless you'd get volume discounts by buying a LOT of the same RAM - and nVidia sure did, it seems.

I'd guess nVidia would also do some minor feature differences on the two models, though.

The 450/425 speed is probably because above 450Mhz, they might have needed Flow FX or something. Still not too bad, though. I'd guess the Per-clock efficiency is well improved ::)

What I don't understand, though, is the real motivation behind the 450Mhz memory ( = 28.8GB/s ) if there wasn't any AA quality / maximum number of samples with pure MSAA improvement...

Frankly, my 19 inch LCD cannot do more than 1280x1024 - and having ridiculously high speeds with 4x AA at those res is nice, but having 6x AA or 8x AA with pure MSAA for acceptable performance would be even better, though.
Heck, nVidia could have upgraded AA number of samples but not quality - although upgrading the number of samples easily allow you to upgrade quality. So my guess is that they'd have done some PR job on it - but nothing about that has been leaked, even in my PR spec list.

So, really, it's surprising. Either something hasn't leaked yet, or those clockspeeds are wrong ( they do seem right, though - maybe a minor mistake on the memory clockspeed of the Regular, but certainly not more ) , or maybe nVidia is expecting us to use 22 inch monitors to play at 2048x1536 @ 4x AA...


Uttar

[GlowStick]

Originally posted by Uttar
or maybe nVidia is expecting us to use 22 inch monitors to play at 2048x1536 @ 4x AA...


Uttar

Im all for that.

[Alphy]

Im really looking forward to seeing what kind of benches that thing will put out. Uttar, what kind of performance increase do you think we will see in 3DMark03 and other standard benches? 10-20%?