So - 6800u - Low Volume Review Board...

[AlphaWolf]

Intel, VIA and SIS have PCIE chipsets too ... or were you referring to something completely different?

Something completely different, I don't think many consumers are ready to jump to adopt PCI-e yet.

Most consumers just buy whatever comes in the package that Dell (or whoever) sells them. If Dell sells them a pci-e system, they will make the jump, likely blissfully ignorant of it until they attempt to wedge an agp card into that slot.

 

[MeltedRabbit]

If nV is indeed delaying for NV45, how easy is it to produce two different versions of that chip (one with the HSI on-package, one without)?

Only 50-100 6800Us doesn't square with more 6800GTs, though. You'd think nV would want more 6800Us out the door for their fatter profit margins, and 350 to 400MHz doesn't seem like a huge leap for the core to make. Maybe this is a hint of a severe 1200MHz GDDR3 shortage?

nv45 was discovered to just be the nv40 with the bridge chip on package instead of a diffrent chip. If anything that may hurt yields or slow the chip down

Is the HSI on the nv45 on die or on package? If the HSI is on die, or on the same chip of silicon, it would require a new revision of the nv4x chip and a lager die, not a cheap option. However, if the HSI is on the same package, ie two chips of silicon on the same organic package, this may be even worse. Multichip packages like this are expensive to make and IIRC a have a high rate of defects. The third option, a separate chip for the HSI would also be expensive. The cheapest option would be just to get rid of the HSI and AGP bus interface entirely.

 

[jvd]

If nV is indeed delaying for NV45, how easy is it to produce two different versions of that chip (one with the HSI on-package, one without)?

Only 50-100 6800Us doesn't square with more 6800GTs, though. You'd think nV would want more 6800Us out the door for their fatter profit margins, and 350 to 400MHz doesn't seem like a huge leap for the core to make. Maybe this is a hint of a severe 1200MHz GDDR3 shortage?

nv45 was discovered to just be the nv40 with the bridge chip on package instead of a diffrent chip. If anything that may hurt yields or slow the chip down

Is the HSI on the nv45 on die or on package? If the HSI is on die, or on the same chip of silicon, it would require a new revision of the nv4x chip and a lager die, not a cheap option. However, if the HSI is on the same package, ie two chips of silicon on the same organic package, this may be even worse. Multichip packages like this are expensive to make and IIRC a have a high rate of defects. The third option, a separate chip for the HSI would also be expensive. The cheapest option would be just to get rid of the HSI and AGP bus interface entirely.

not sure it was discussed some where on the boards . will try and find it

 

[Lezmaka]

Is the HSI on the nv45 on die or on package? If the HSI is on die, or on the same chip of silicon, it would require a new revision of the nv4x chip and a lager die, not a cheap option. However, if the HSI is on the same package, ie two chips of silicon on the same organic package, this may be even worse. Multichip packages like this are expensive to make and IIRC a have a high rate of defects. The third option, a separate chip for the HSI would also be expensive. The cheapest option would be just to get rid of the HSI and AGP bus interface entirely.

There's a picture of it at Anand's

And here's a quote

Underneath the heatsink you'll see the NV45 GPU, as well as a separate die on the same package.

The second, smaller die is NVIDIA's High Speed Interconnect (HSI) which is an AGP-to-PCI Express bridge, meaning that NV45 is still an AGP part.

 

[incurable]

What I think is more likely is that the native PCIe version of nV40 doesn't need a bridge chip because its pcb support is exclusively PCIe, like ATi's bridge-chipless PCIe product announcements. Later, I would expect to see both companies marketing a hybrid native PCIe pcb with a PCIe x16 to AGP x8 bridge chip onboard so that they could sell the same product into both markets. It's cheaper of course to sell native versions for each standard right now (as they already have the AGP tech in place), and as of now there's about ZERO demand for PCIe products aside from system OEM demand--which can be served by native PCIe pcbs. I can't imagine nVidia taking the nV40, putting it on a PCIe pcb, and calling it "nV45"--unless, of course, they are somehow desperate for new product PR. There'd be no difference between nV40 and nV45 in that case, except for the pcb host-bus interface circuitry the nV40 would be sitting on.

Walt, I think you should try to get a better grip on the vocabulary you're using.

IIRC, this is not the first time that you abused the term PCB (saying "PCIe pcb" suggests that the PCB is an active component, which it clearly isn't), and I think I even commented on that before.

As for your last comment, you seem to be corret, there's no difference between the GPU die used for NV40 and NV45. That also explains the 'closeness of tapeout'-rumors we've been hearing about both. I think alot of confusion stems from the fact that up until the PCX line of products, it seemed like nVidia codenamed GPUs, now they either changed or clarified that and codenames apparently are for designs solutions. (NV18_GPU+HSI = NV19,NV36_GPU+HSI = NV39 and now NV40_GPU+HSI = NV45)

cu

incurable

 

[WaltC]

Walt, I think you should try to get a better grip on the vocabulary you're using.

IIRC, this is not the first time that you abused the term PCB (saying "PCIe pcb" suggests that the PCB is an active component, which it clearly isn't), and I think I even commented on that before.

You're suggesting that the Printed Circuit Board for PCIe bus interface is not an "active" component?.... I'm not sure what you think I'm saying, but what I'm saying is that there's a distinction between a gpu and the host-bus circuitry interface the gpu sits on--which is usually mounted on the same pcb the gpu is mounted on. By "pcb" I mean the circuitry extraneous to the cpu, but local to the same pcb, which defines the host-bus interface for the card (slot pins, etc., included.)

As for your last comment, you seem to be corret, there's no difference between the GPU die used for NV40 and NV45. That also explains the 'closeness of tapeout'-rumors we've been hearing about both. I think alot of confusion stems from the fact that up until the PCX line of products, it seemed like nVidia codenamed GPUs, now they either changed or clarified that and codenames apparently are for designs solutions. (NV18_GPU+HSI = NV19,NV36_GPU+HSI = NV39 and now NV40_GPU+HSI = NV45)

cu

incurable

nV10, nV15, nV20, nV25, nV30, and now nV40 have all denoted some type of architectural if not generational gpu advance in succession. For nVidia to name an nV40 sitting on a PCIe pcb (as described above) "nV45" seems completely incongruent to me, as these numbers have always denoted gpus, not reference designs. Certainly, I have no objection to nVidia calling an nV40 gpu mounted on a PCIe pcb "nV45"--but then I wonder what nVidia will start naming its gpus afterwards...

In case you still don't quite follow my vernacular here, when ATi shipped the R300 the company made it plain that the R300 vpu itself would have been perfectly at home sitting on any host bus from PCI 66-AGP x8, but that the limitation of the 4x-8x AGP reference-design cards actually sold was an electrical limitation imposed strictly by the 4x-8x host bus interface circuitry of the reference design chosen for R300 deployment. IE, the same R300 would have done just fine mounted on a pcb congruent with PCI66 electrical connectivity as it would on a pcb congruent with AGP x8 electrical connectivity.

So, it's generally not a good idea to confuse the gpu/vpu with the local bus circuitry the pcb it sits on uses to electrically connect with the host bus. So, I say "PCIe pcb" only to cut down on wordy descriptions, and to denote the distinction between a gpu/vpu and the host-bus interface circuitry of the pcb it happens to be mounted on. Just trying to illustrate that there's technically no such thing as a "PCIe gpu/vpu" as far as I can see, just as R300 was no more PCI66-native than it was AGP x8 native, etc.

 

[incurable]

You're suggesting that the Printed Circuit Board for PCIe bus interface is not an "active" component?....

Yes, exactly. PCBs are just a bunch of wires between multiple layers of plastics.

I'm not sure what you think I'm saying, but what I'm saying is that there's a distinction between a gpu and the host-bus circuitry interface the gpu sits on--which is usually mounted on the same pcb the gpu is mounted on.

The bus interface unit is part of the GPU, it's on-die.

By "pcb" I mean the circuitry extraneous to the cpu, but local to the same pcb, which defines the host-bus interface for the card (slot pins, etc., included.)

I'm still not sure which part you're referring to, are you referring to the flip chip carrier or substrate which the GPU die is affixed to?

nV10, nV15, nV20, nV25, nV30, and now nV40 have all denoted some type of architectural if not generational gpu advance in succession. For nVidia to name an nV40 sitting on a PCIe pcb (as described above) "nV45" seems completely incongruent to me, as these numbers have always denoted gpus, not reference designs. Certainly, I have no objection to nVidia calling an nV40 gpu mounted on a PCIe pcb "nV45"--but then I wonder what nVidia will start naming its gpus afterwards...

I'd expect them to have different codenames for every different pin-layout of the substrate, regardless of GPU and HSI combination used.

In case you still don't quite follow my vernacular here, when ATi shipped the R300 the company made it plain that the R300 vpu itself would have been perfectly at home sitting on any host bus from PCI 66-AGP x8, but that the limitation of the 4x-8x AGP reference-design cards actually sold was an electrical limitation imposed strictly by the 4x-8x host bus interface circuitry of the reference design chosen for R300 deployment. IE, the same R300 would have done just fine mounted on a pcb congruent with PCI66 electrical connectivity as it would on a pcb congruent with AGP x8 electrical connectivity

Quite frankly, I've never heard/read such comments from ATi, but it's conceivable that you could support all these modes with a single bus interface unit, as they're rather simliar. (PCI66 in that case probably equals AGP 1x without AGP texturing support)

So, it's generally not a good idea to confuse the gpu/vpu with the local bus circuitry the pcb it sits on uses to electrically connect with the host bus. So, I say "PCIe pcb" only to cut down on wordy descriptions, and to denote the distinction between a gpu/vpu and the host-bus interface circuitry of the pcb it happens to be mounted on. Just trying to illustrate that there's technically no such thing as a "PCIe gpu/vpu" as far as I can see, just as R300 was no more PCI66-native than it was AGP x8 native, etc.

As I've written above, the bus interface unit is part of the GPU, it's on the GPU die, and the only function of the different pieces of PCB the die is affixed to is the channel the signals coming from/going to the die. R300 'talks' AGP to the outside world (at different speeds though, as you mentioned), as do R420 and NV40. For these designs, the outside world is the AGP interface of the chipsets northbridge.

Now for the NV45 design it's a little more complicated, nVidia implemented the HSI bridge on the same carrier as the NV40_GPU, so the NV40_GPU 'talks' AGP (at an accelerated rate) to this bridge and the bridge then 'talks' PEG to the PEG interface in the northbridge of the chipset.

Native PEG designs like RV370, RV380 and RV423 'talk' PEG to the outside world, no amount of PCB work can change that and the only way for them to communicate with an older chipset having an AGP interface would be through a translator, or bridge, somewhere in the signal path.

cu

incurable

 

[Rys]

To this very day, some 6800 Ultra board vendors still don't know the clock speeds of their own cards. Some are shipping out review samples at clocks that in no way reflect what a customer will get, and one 6800U partner board looks to have a 'floating' clock speed, so two boards from that partner could very well come at different clocks.

And if there's only 100 6800 Ultras in the world, I have over 1% of them in my office

They need to get them to retail NOW, screw the new driver and stop pissing about with the clocks, before it gets much worse.

Rys

 

[digitalwanderer]

"Anticipation, anticipay-ay-shun...."

 

[fallguy]

Didnt NV criticise ATi for differently clocked boards sent to reviewers? Kinda hypocritical eh.

 

[jimmyjames123]

Didnt NV criticise ATi for differently clocked boards sent to reviewers? Kinda hypocritical eh.

Not really, these situations are not entirely comparable. All of the 6800 boards sent for review had the same core and mem clocks (6800UE = 450Mhz core/550Mhz mem, 6800U = 400Mhz core/550Mhz mem, 6800GT = 350Mhz core/500Mhz mem). Also, the actual cards that the board makers ship out should not have lower core clock and mem speeds than what was tested in the reviews, meaning that the consumer really only stands to gain. On the other hand, ATI sent out several X800XT PE cards to reviewers, many of which had different core clocks and mem clocks, so if the reviewers did not manually change the clocks then the review boards would not have perfectly comparable clocks. Now, it's not that big a deal because the core and mem clocks on the X800XT PE were generally only different by a couple percent at most, but it does make you wonder because it is generally not standard practice to send cards out to reviewers with varying core and mem clocks. Most likely, ATI felt the need to display the X800XT as soon as possible in order to respond to the 6800U, so that's why you saw them launch the card along with the X800Pro even though it is generally not their policy to launch until the product is ready to ship in quantity.

 

[digitalwanderer]

My buddy Skuzzy over at Elite Bastards got his Gateway X800 XT PE, they're real.

 

[whql]

Not really, these situations are not entirely comparable.

Not really? Please, lets not have this fanboy defense.

Go and listen to what your buddy JHH actually said - he told investors that they "decide on a clock speed and ship at that clockspeed". Everything that happened since then has just ptoven that he's either outright lying to investors or doesn't know his own business - neither of which should look good to investors!

 

[Bouncing Zabaglione Bros.]

Didnt NV criticise ATi for differently clocked boards sent to reviewers? Kinda hypocritical eh.

Not really, these situations are not entirely comparable.

How would we compare it to the "Ultra Extreme" that was sent out as a spoiler on the R420 launch? You know, that hand picked, highly overclocked card that it turns out no company is making or shipping, and isn't actually a product at all, except as a spoiler engineering sample from Nvidia?

 

[Rys]

Didnt NV criticise ATi for differently clocked boards sent to reviewers? Kinda hypocritical eh.

Not really, these situations are not entirely comparable.

How would we compare it to the "Ultra Extreme" that was sent out as a spoiler on the R420 launch? You know, that hand picked, highly overclocked card that it turns out no company is making or shipping, and isn't actually a product at all, except as a spoiler engineering sample from Nvidia?

That's not quite the case. At least one board vendor will market and ship the Extreme, using the Extreme name, at 450MHz core.

Rys

 

[Bouncing Zabaglione Bros.]

That's not quite the case. At least one board vendor will market and ship the Extreme, using the Extreme name, at 450MHz core.

Rys

Link?

 

[kipper67]

E-VGA are doing an Ultra Extreme , that's the one they are currently raffling the chance to BUY one each week

http://www.evga.com/articles/public.asp?AID=188

One of the posters over at NVnews got one and the box is as shown on the site with "Extreme edition" on it.

Mark

 

[Bouncing Zabaglione Bros.]

E-VGA are doing an Ultra Extreme , that's the one they are currently raffling the chance to BUY one each week

http://www.evga.com/articles/public.asp?AID=188

One of the posters over at NVnews got one and the box is as shown on the site with "Extreme edition" on it.

Mark

http://www.evga.com/articles/public.asp?AID=192 425mhz, *maybe* 440 maximum if your setup is perfect. That's still slower than the Ultra Extreme Nvidia sent to reviewers.

 

[kipper67]

Yeah , I know it's a bit "ifs and buts" but both the page I linked to and the link you have to follow to the dynamic overclocking info page both state 450Mhz for the UE.

I'll see if I can dig up the link to the thread where the guy has posted photos of the card and box if i can remember which forum it was on.

Mark

 

[fallguy]

Another example of NV being hypocrits.

http://www.nvnews.net/vbulletin/showthread.php?p=339351#post339351

Supposed to be 325Mhz core he says, yet the sample they got was 335Mhz on the core. So NV complains about ATi's different speeds with preview cards, yet they do the same thing.