NV40 clock speed

[Josiah]

I don't think this has been posted yet (cross fingers)

http://www.nvnews.net/

NV40 Exclusive Tidbit - 6/22/03 4:58 pm - By: volt - Source: Uttar
This was meant to be posted a few days back, but because I went on a short vacation and some things weren't ironed out, we decided to do it today. nV News has gotten exclusive information regarding the NV40 project. Not much to report but it came from a very reliable source (as usual) - thanks to Uttar.
As we all know, NVIDIA's next generation chip is the NV40 (NV35 was a refresh). What we don't know is what will it bring to the table and when will it actually appear on the market.

To cut to the chase, the NV40 is currently being targeted for the following:

550MHz-600MHz Core
700MHz-800MHz Memory

So how accurate are those numbers? VERY. We were not given the exact clock speeds because those do not exist yet. Stay tuned for more

 

[Dave H]

To cut to the chase, the NV40 is currently being targeted for the following:

550MHz-600MHz Core
700MHz-800MHz Memory

So how accurate are those numbers? VERY. We were not given the exact clock speeds because those do not exist yet. Stay tuned for more

At first I was a bit skeptical about the DRAM clock rate, but this is actually very much in line with the use of GDDR3. Obviously it bears noting that ATI cards in the same timeframe are very likely to be using similar speed DRAM, especially as it was ATI who first indicated interest in GDDR3.

The disparity in core and clock rates would seem to indicate either that NV40 is moving back to a 128-bit bus (which seems very unlikely IMO), or that it has a very beefy core, along the lines of an 8x2. 45-51 GB/s would make that thing a multisampling monster; let's hope Nvidia's added the sample patterns and gamma-correction to take advantage of it.

As for the core clocks, they are neither surprising nor revealing, as we don't know what it's capable of per-clock.

 

[overclocked]

Nothing of a suprise really, GDDR3 was/is expected to hit volume already in Q3 acording to that link. The timeframe of NV40 is in late Q4 or early next year.
R420 should/is(?) taped out now and going for a final respin(has??) so it will have a headstart of a couple of months(IMO).

 

[Evildeus]

Doesn't say much on PS2/3 improvements, AA improvements, etc.

But it seems quite reasonnable, it's just a 100-150MHz bump from the NV35 and the memory, well GDDR3 or GDDR2 should be @ those figures at that time frame.

 

[Arun]

Okay, so because I'm a lazy lazy man, let me quote myself:

Cotita: Hmm, it IS 700Mhz-800Mhz RAW.

That means 44.8-48GB/s considering a 256-bit memory bus ( I must admit I got no info that it's a 256-bit memory bus, but it'd make no sense to use a 128-bit one, since overall, it'd cost even more than the NV35 and be less effective! )

And frankly, no problem if you don't trust me - but I certainly know I trust my source Never has been wrong before, if you know what I mean.

Now, AFAIK, and this comes from another source, also very reliable ( but the info is also less recent, so it *might* be outdated, doubt it though ) - the NV40 uses GDDR2.
I was very surprised by this myself - seemed like insane for GDDR2 figures.
But considering there's already 600Mhz GDDR2 IIRC, having 700Mhz in Q4 is not insane - and 800Mhz would then kinda be the best-case scenario.
As I said, the fact it uses GDDR2 might be outdated - or heck, the memory controller could support both ( GDDR2 for the $399 part and GDDR3 for the $499 one? Who knows - I don't ) - but it's still the most likely.

Also, the NV40 is *not* a NV35 refresh. It's a completely new core - just like the NV10, NV20 or NV30. The biggest feature additions over the NV30 are texture lookups in the Vertex Shader and branching in the Pixel Shader, making it PS3.0. / VS3.0. compliant.
A very cool feature is also the addition of a Programmable Primitive Processor. Few details on the implementation as of yet, though


So, to summarize, the targets are, I believe:
- 550+Mhz core clock on IBM's 0.13µ process ( no idea if it uses SOI or not )
- 44.8GB/s+ memory bandwidth using most likely GDDR2, maybe GDDR3

What's a given, too, is that Loki will also feature that type of memory - details are obviously hard to tell, because it's all targets right now ( and I don't even have Loki's targets, eh... )

Finally, a disclaimer I believe volt should have added, but oh well:

Those numbers are targets. Problems might arise, such as an overoptimistic look at IBM capabilities or too aggressive RAM decisions. In conclusion, the final clock speeds might be either higher or lower than those targets.
However, the info is still very accurate - I firmly believe those are nVidia's current targets.

BTW, for avaibility - the idea there is still "hopefully before year end". Speculation: Don't hope for more than limited avaibility though - more like reviews in early December and a few cards in stores several days before christmas, just to make sure Loki doesn't steal the spotlight.

Anyway, I hope you appreciated the info. Remember though that the most reliable piece are those targets - the rest is also very reliable, but still not as much


Uttar

Need any more precisions?


Uttar

 

[Dave H]

Nothing of a suprise really, GDDR3 was/is expected to hit volume already in Q3 acording to that link. The timeframe of NV40 is in late Q4 or early next year.
R420 should/is(?) taped out now and going for a final respin(has??) so it will have a headstart of a couple of months(IMO).

No, it said general availability of samples in Q3, not production. That means that in Q3, if you call up Micron and say "I'm thinking of designing a product using GDDR3," they'll send you some, whereas now (limited samples) they only have enough to send to important people like Nvidia or ATI. Samples in Q3 would tend to mean production in Q104, which would mean availability of cards using the stuff in Q2. Of course that's very general; it could easily be off by a quarter, although it'd be pretty difficult to get an OEM product out in Q4 using a new type of DRAM that didn't sample until Q3.

But Infineon, who's also making GDDR3, might be ahead of Micron (who traditionally concentrates on low-cost production more than meeting cutting-edge standards), although I can't find any mention of the stuff on their website.

 

[andypski]

Need any more precisions?
Uttar

No - they have too many precisions already.

(Couldn't let a setup like that go past...)

 

[DegustatoR]

No - they have too many precisions already.

BTW, i have e feeling, that there'll be only FP32 in NV4x...

 

[overclocked]

No, it said general availability of samples in Q3, not production. That means that in Q3, if you call up Micron and say "I'm thinking of designing a product using GDDR3," they'll send you some, whereas now (limited samples) they only have enough to send to important people like Nvidia or ATI. Samples in Q3 would tend to mean production in Q104, which would mean availability of cards using the stuff in Q2. Of course that's very general; it could easily be off by a quarter, although it'd be pretty difficult to get an OEM product out in Q4 using a new type of DRAM that didn't sample until Q3.

But Infineon, who's also making GDDR3, might be ahead of Micron (who traditionally concentrates on low-cost production more than meeting cutting-edge standards), although I can't find any mention of the stuff on their website

And i thought i was conservative

 

[overclocked]

Btw, damn Uttar.. It´s not that it´s easy to chose what to buy now!

 

[Luminescent]

Will NV40 offer the unified-shading-model architecture?

 

[KimB]

BTW, i have e feeling, that there'll be only FP32 in NV4x...

It would make sense for there to only be FP32 pipelines (as anything less wouldn't be useful in a vertex shader, given the rumors of a unified shader architecture). But that doesn't mean that lower precisions won't be supported. There may still be performance benefits from using FP16 from register usage, though hopefully the hardware will have a much easier time handling large numbers of registers.

 

[Hellbinder]

Also, the NV40 is *not* a NV35 refresh. It's a completely new core - just like the NV10, NV20 or NV30. The biggest feature additions over the NV30 are texture lookups in the Vertex Shader and branching in the Pixel Shader, making it PS3.0. / VS3.0. compliant.
A very cool feature is also the addition of a Programmable Primitive Processor. Few details on the implementation as of yet, though

Well that statement right there throws in some doubt for me. As Nvidia does not even have one single *Completely new core*.. Since the origional Geforce. They rehash the same Design with additions and slight modifications over and over and over..

Then try to cover it up with cleaverly worded PR statements about *execution units* and *pipelines are not pipelines anymore*..

The Nv30 is the closest thing to a new completely new core they have.

 

[Hellbinder]

One more thing about the Nv40...

It was origionally slated for late Q1 2004 at the earliest. I have seen several comments about this. In fact a while back when it was even suggested Nv40 would be shown this year many people said that was just impossible.

Which means (to me) that They have scaled back the Design of the Nv40 a considerable ammount. So i dont buy for one min that its a *Completely new core*. My bet is that it is exactly a tweaked Nv35. With Full 8 pixel output and 8 FP32 units instead of 4. Some added PS3.0 functionality.

That is hardly a *new core* Its a Tweaked Nv35 with 2x the shader performance and Higher Clockspeeds.

 

[Evildeus]

Well, could be. But with the NV3* fiasco, i think it is wise to get out the NV40 before the previous scheddule.

I don't think 3 month got anything to do with scale back or there would be big issues on how the chip are created, but more with less NV3* period.

 

[BoardBonobo]

One more thing about the Nv40...

It was origionally slated for late Q1 2004 at the earliest. I have seen several comments about this. In fact a while back when it was even suggested Nv40 would be shown this year many people said that was just impossible.

Which means (to me) that They have scaled back the Design of the Nv40 a considerable ammount. So i dont buy for one min that its a *Completely new core*. My bet is that it is exactly a tweaked Nv35. With Full 8 pixel output and 8 FP32 units instead of 4. Some added PS3.0 functionality.

That is hardly a *new core* Its a Tweaked Nv35 with 2x the shader performance and Higher Clockspeeds.

You're right!! That wouldn't be a new core.

Now I wonder what the NV40 will be...

 

[nonamer]

Also, the NV40 is *not* a NV35 refresh. It's a completely new core - just like the NV10, NV20 or NV30. The biggest feature additions over the NV30 are texture lookups in the Vertex Shader and branching in the Pixel Shader, making it PS3.0. / VS3.0. compliant.
A very cool feature is also the addition of a Programmable Primitive Processor. Few details on the implementation as of yet, though

Well that statement right there throws in some doubt for me. As Nvidia does not even have one single *Completely new core*.. Since the origional Geforce. They rehash the same Design with additions and slight modifications over and over and over..

Then try to cover it up with cleaverly worded PR statements about *execution units* and *pipelines are not pipelines anymore*..

The Nv30 is the closest thing to a new completely new core they have.

I don't many companies have every produce a "Completely new core." Almostly any sane company would put good technology from their old chips to their new one. I think everyone is trying get at is that the NV40 has enough major changes to be consider a totally new core.

 

[overclocked]

Well that statement right there throws in some doubt for me. As Nvidia does not even have one single *Completely new core*.. Since the origional Geforce. They rehash the same Design with additions and slight modifications over and over and over..

Then try to cover it up with cleaverly worded PR statements about *execution units* and *pipelines are not pipelines anymore*..

The Nv30 is the closest thing to a new completely new core they have.

Depends of how you define a "new" core. With cpu´s very long development time it´s a different story and "easier" to see.

With gpu´s past and what seem´s to be current timescale it´s trickier to get what a "new gpu-core" is.
This is my own opinion on wich cores what new and why.

First NV10 because of the FF-T&L engine. / NV15 a refresh.
Then NV20 because of the implention of Vertex and pixelshader´s / NV25 a NV20 on steroid´s.
The NV30 because of the DX9 shaders and floatingpoint through the pipeline/ NV35 a NV30 on stereoids.

The number of transistors also IMO can define these "past" Nvidia-cores.
I don´t remember but NV10~20m, NV15~25m, NV20~57m, NV25~63m, NV30~125m and finally NV35~130m

 

[Luminescent]

Nvidia did choose to stick with register combiners for quite a while.

 

[Arun]

Hellbinder:
Well, you could argue even the GeForce isn't a new core, it's based on the TNT2, eh...

More importantly though, I don't think nVidia reuses much of their past code for new cores. Often they heavily base themsleves on their past generation ideas, but I'd guess they'd always reimplement it or put time on optimizing it, in order to save transistors. Just speculation of course, but it makes sense IMO. But yeah, they aren't really "completely new cores" - what is one, though?

Also, in relation to avaibility.
Christmas is definitively for "limited" avaibility, maximum. For "real" avaibility, it's still Q1 2004 I guess. As I said, my avaibility data is not very precise - but some type of avaibility before year end definitively is nVdia's target right now.

How come it even seems to go better than they expected before?
I believe a good reason might have been the NV40 team probably had ZERO bugdet compromises. nVidia knows that if they don't deliver this time, they're toast - so letting them use more money than truly necessary in order to have finished very slightly earlier and have less bugs / more optimizations wouldn't bother them I guess.
Heck, in one of their latest conference call, nVidia told their R&D was going to increase this quarter because they were going to buy emulators for their next-generation part. Err, okay, that's expensive, but up to the point for it to be really noticable financially? Eh, you can give me one of those emulators too, hehe

Another reason, as you said youself, might be that they've cutted corners. I must admit to doubt it, however. I think that the problem comes more from some people ( read: mostly me ) being overoptimistic at the NV40 features, such as dynamic allocation - those things quite simply aren't in the NV40 I now believe ( NV50, though... j/k )

As for your "simply full 8 pixel and 8 FP32 units + PS3.0." - well, you're maybe nearly describing the original NV40 design there, eh.
The NV40 is what the NV30 *should* have been ( that is, the very initial design plans of the NV30, which I even think included a PPP - not for long though ) + full shaders 3.0. compliancy and loads of optimizations/fixes.

Now, I believe one of the biggest remaining NV40 mystery lies in its pipeline organization. I've heard 8x1 like the R300 - I've also heard people saying "likely to be 8x2" - but that's probably speculation.

Heck, I believe we've got one excellent indication from good ole CMKRNL:

I had stated earlier that the programmable tessellation unit was going to be in NV35. It's not clear to me anymore that this is the case. Apparently this will be in a 4Q'03 part, which leads me to believe that it will be NV40, not NV35. This part will also contain a completely revamped unified shading model. This means that both vertex and pixel shaders will share the exact same ISA and constructs.

Bolding is mine.

The NV3x uses a very specific VS construct - very strange, too. I'd guess it's a good design choice for branching. So since the pixel shaders will also have access to branching, it *might* make sense to include such a model in the pixel pipelines too. Although it could not be compatible, I don't know - just enumerating possibilities here

We'll discover it soon enough, I hope...


Uttar