NV48 is NV40 refresh?

[DSC]

http://www.anandtech.com/showdoc.html?i=2068

We've heard two different things from manufacturers. Some manufacturers have told us that NV45 will be available in August, while others have basically said that NV45 won't be out this year, instead we will see NV48 arrive as an AGP refresh to NV40 by Christmas.

What will most likely happen is that NVIDIA will determine the fate of NV45 by the acceptance (or lack thereof) of PCI Express this year. If the acceptance is low enough, NV45 will be pushed out to next year, otherwise we may see it in Q3/Q4. Regardless of what happens, it does seem like NV48 will be NVIDIA's fall refresh product, and that will be an AGP solution

 

[Dave Baumann]

This [the topic title] is what I had always been expecting.

 

[Lezmaka]

Well thanks for letting us know earlier Dave...

 

[Evildeus]

Well, i think for some time people talked of NV45 as a PCI-E NV40...

 

[Dave Baumann]

OK, I have to admit, though, I was expecting NV45 to be native PCIe and not have an on-package bridge chip. Innovative, but it will still have the fundamental limitations that HSI may have.

 

[Hanners]

Are there any real benefits from moving the HSI bridge chip onto the package rather than leaving it on the PCB?

 

[Dave Baumann]

I don't know if NV45 is, but I guess that potentially it would make it easier to run the chip at higher speeds, if the AGP bus could cope with it.

The obvious advantage is board space and configuration. If you look at the other HSI parts, with the descrete chip, the chip and memory configuration is either moved round or pushed back, something that would be difficult with NV40's memory arrangement.

 

[Aivansama]

I don't know if NV45 is, but I guess that potentially it would make it easier to run the chip at higher speeds, if the AGP bus could cope with it.

The obvious advantage is board space and configuration. If you look at the other HSI parts, with the descrete chip, the chip and memory configuration is either moved round or pushed back, something that would be difficult with NV40's memory arrangement.

Reduced wiring capacitances would enable lower power consumption at same speeds or faster clocks at same power - between the HSI and chip. The wiring from the HSI -> PCIE wouldn't change much. I doubt if changing the HSI - chip bus speed would gain anything, so most likely this is a way to A) reduce the power consumption of the HSI and B) decrease the packaging costs (and board costs, also).

 

[Evildeus]

According to HFR, it's not really a scoop:

S’agit-il d’un changement des plans de NVIDIA vis-à-vis du PCI Express ? Pas vraiment. En effet, en février dernier nous avions posé cette question à David Kirk de NVIDIA :


A propos du PCI Express, vous avez dit qu’au moins certains membres de la famille NV4x auraient un support natif de cette technologie. Si ce que nous savons est vrai, ce ne serait pas exactement le cas puisqu’il s’agirait en fait d’une implémentation du HIS sur le packaging ou au mieux d’une intégration du HIS au sein du GPU sans un redesign du GPU afin d’exploiter au mieux le PCI Express. Qu’en est-il réellement ?

Ceci est une façon de faire, et quelques produits seront packagés de la sorte. Toutefois, la plupart des membres de la famille NV4x sont vraiment ce que j’ai dis, c'est-à-dire nativement PCI Express avec une intégration sur le die. Seule la première version du NV40 sera AGP, et il y’aura donc une version de cette puce interconnectée avec un HIS. Toutes les GPU milieux et bas de gamme de la famille NV40 auront une interface PCI Express intégrée. Pour les versions AGP de ces puces, nous utiliserons un HSI.

http://www.hardware.fr/news/lire/01-06-2004/

 

[Tridam]

According to HFR, it's not really a scoop:

S’agit-il d’un changement des plans de NVIDIA vis-à-vis du PCI Express ? Pas vraiment. En effet, en février dernier nous avions posé cette question à David Kirk de NVIDIA :


A propos du PCI Express, vous avez dit qu’au moins certains membres de la famille NV4x auraient un support natif de cette technologie. Si ce que nous savons est vrai, ce ne serait pas exactement le cas puisqu’il s’agirait en fait d’une implémentation du HIS sur le packaging ou au mieux d’une intégration du HIS au sein du GPU sans un redesign du GPU afin d’exploiter au mieux le PCI Express. Qu’en est-il réellement ?

Ceci est une façon de faire, et quelques produits seront packagés de la sorte. Toutefois, la plupart des membres de la famille NV4x sont vraiment ce que j’ai dis, c'est-à-dire nativement PCI Express avec une intégration sur le die. Seule la première version du NV40 sera AGP, et il y’aura donc une version de cette puce interconnectée avec un HIS. Toutes les GPU milieux et bas de gamme de la famille NV40 auront une interface PCI Express intégrée. Pour les versions AGP de ces puces, nous utiliserons un HSI.

http://www.hardware.fr/news/lire/01-06-2004/

Actually the original version is in English

About PCI Express it was said that at least some members of the NV4x family will have native support. If what I know is true it won't be the case. Isn't this "native" support got by putting the HSI in the packaging or at the best by adding the HSI on the die without a GPU redesign for a better use of PCI-Express ?

That is one way to do it, and some products will be packaged that way.
However, most members of the NV4X family are really what I said - native PCI-Express - it's on the die. Only the first version of NV40 will be AGP16X, so there will be a HSI-connected version of that chip. All of the mid-range, mainstream, and value chips in the NV40 family will have only a PCI-Express interface on the chip. For AGP versions of those chips, we will use HSI.

(Asked in february)

 

[Evildeus]

Actually the original version is in English

I'm disapointed that Mister Kirk doesn't speak french

 

[DegustatoR]

Since it's an nv40 chip + hsi chip in one packaging, why call it nv45? There gotta be something more in nv45 chip itself...

 

[bloodbob]

Since it's an nv40 chip + hsi chip in one packaging, why call it nv45? There gotta be something more in nv45 chip itself...

Probably just another stepping fixing a few minor bugs ;P

 

[Joe DeFuria]

Since it's an nv40 chip + hsi chip in one packaging, why call it nv45? There gotta be something more in nv45 chip itself...

Well, if it's fabbed at TSMC, it's a new chip...

 

[Tridam]

nv18 + hsi = nv19
nv34 + hsi = nv37
nv36 + hsi = nv39
nv40 + hsi = nv45 ?

Why do they need another refresh for NV45 ?

 

[Joe DeFuria]

Why do they need another refresh for NV45 ?

It can simply be to change fabs from IBM to TSMC, due to either production or capacity issues at IBM (or better costs in general from TSMC).

 

[Tridam]

Why do they need another refresh for NV45 ?

It can simply be to change fabs from IBM to TSMC, due to either production or capacity issues at IBM (or better costs in general from TSMC).

It doesn't require a name change.

 

[Joe DeFuria]

It doesn't require a name change.

Never said it did.

 

[Geeforcer]

It's not a name, it's an internal codename for a different design. They can call it NV145.67 if they want to.

 

[Pete]

OK, I have to admit, though, I was expecting NV45 to be native PCIe and not have an on-package bridge chip. Innovative, but it will still have the fundamental limitations that HSI may have.

What are HSI's fundamental limitations? Just perhaps higher latencies and greater power draw, right? Or is there something about NV40's memory controller/bus interface that is specific to AGP and thus handicapped with PEG? Is it possible the NV45 isn't just a NV40 + HIS, but rather a tweaked GPU geared to fully exploit PEG's greater bandwidth?