NVIDIA shows signs ... [2008 - 2017]

[INKster]

But I thought Nvidia was getting out of the chipset business - Charlie and "analysts" said so so it must be true. Now Apple's entire line is using Nvidia chipsets. One of these two does not make sense.

They're both right.
Nvidia did exit the "chip-set" business and now has just... a chip, for both AMD and Intel-based CPU platforms.

That's the advantage of integration, and the company itself had proclaimed the death of their traditional Northbridge/Southbridge chipsets quite some time ago (around the nForce 790i SLI launch, if i'm not mistaken).

It's funny though. Apple chose to use expensive 1.5v DDR3-1066 RAM IC's across the board with MCP79/Geforce 9400M, yet that chip also supports the much cheaper DDR2-667/800.
Frankly, i think a 0.3v difference can't make that much of a difference on battery life in the real world. The delta between DDR's 2.6v and DDR2's 1.8v was much larger.

 

[silent_guy]

Frankly, i think a 0.3v difference can't make that much of a difference on battery life in the real world. The delta between DDR's 2.6v and DDR2's 1.8v was much larger.

I haven't really gone through the raw numbers and there'll obviously be a chunk of marketing speak in there, but according to this page:

DDR3 not only allows for higher system performance, but it also uses only 1.5V of power compared to DDR2's 1.8V, thereby extending battery life by approximately 20 minutes.

Even if it's only 10 minutes, that's still significant enough to go for DDR3, especially if your customers typically have a known tendency to pay more than for your than for other brands...

Edit: So I did look it up... The highest current (IDD7 in the data sheet) in a 2GB DIMM for a Samsung 1066 DDR3 = 2640mA. For a similar 667 DDR2, it's 2120mA. Multiplied by 1.5V and 1.8V respectively, gives 3.96W vs 3.82W... So DDR2 is doing better here, barely.

I thought maybe some other numbers would give the advantage to DDR3, but that doesn't seem to be the case.

Obviously, corrected for performance, DDR3 should win in terms of perf/W...

 

[Bouncing Zabaglione Bros.]

But I thought Nvidia was getting out of the chipset business - Charlie and "analysts" said so so it must be true. Now Apple's entire line is using Nvidia chipsets. One of these two does not make sense.

The actual (current) rumour is that Nvidia is going to sell it's chipset business lock, stock and barrel to Apple who will continue to run it for it's own PC business.

 

[INKster]

The actual (current) rumour is that Nvidia is going to sell it's chipset business lock, stock and barrel to Apple who will continue to run it for it's own PC business.

Frankly, i find that proposition completely ridiculous.
During yesterday's presentation, a particular slide showed MCP79's die, and Apple itself said 70% of it was pure IGP, while the remaining 30% shared Northbridge and Southbridge duties.

Do you see Nvidia handing out key DX10/OpenGL 3.0/CUDA hardware intellectual property to Apple just like that ? Not me...

 

[Arun]

There is uncertainty regarding how successful they can be because of QuickPath, DMI and Fusion, but NVIDIA definitely has a roadmap. And I think MCP7C is still aimed at 1H09, probably 1Q09 (Socket 775, 4 TMUs, 8 SPs, 64-bit DDR2, 4xSATA/2xPATA, 10xUSB). If you look at the MCP7A die shot released by Apple, it's very clear they were planning on a half-MCP7A from the start (see: live.gizmodo.com/page/6) - BTW, MCP7A is apparently 8 TMUs+16 SPs while G98 & MCP78 are 4 TMUs + 16 SPs. No idea about the MUL in either case, heh.

MCP79 will also be the first NVIDIA MCP to support VIA CPUs, and will be manufactured on 55nm. Beyond that, I suspect they don't have anything pre-40nm. Either way, any and every rumour you heard about NVIDIA's chipset business is bullshit.

 

[INKster]

But isn't DMI just basically the same as PCI-Express x4 ?
What's so troubling about using the same MCP7A chip on a mainstream Core i7 motherboard in LGA-1160 guise next year ?
Theoretically, it would be even better, as then they would be able to turn off the memory controller ("Lynnfield" retains the integrated memory controller, even though it's limited to dual-channel, instead of triple-channel like "Nehalem"/"Bloomfield") and, with it, improve yields/lower costs.

 

[Arun]

But isn't DMI just basically the same as PCI-Express x4 ?
What's so troubling about using the same MCP7A chip on a mainstream Core i7 motherboard in LGA-1160 guise next year ?
Theoretically, it would be even better, as then they would be able to turn off the memory controller ("Lynnfield" retains the integrated memory controller, even though it's limited to dual-channel, instead of triple-channel like "Nehalem"/"Bloomfield") and, with it, improve yields/lower costs.

DMI is a proprietary form of PCI-Express x4, I presume the I/O is the same, the PHY is slightly different, and the MAC is extremely different. So I think it might be possible to reuse the same silicon, but it does require designing it with that in mind.

Secondly, there's the "little" problem of bandwidth. The LGA-1160 chip will communicate to the integrated GPU not through DMI, but through the PCI Express x16 slot. However, that risks being extremely bandwidth limited; so the logical answer would be to use that memory controller as a "sideport" to memory soldered on the motherboard. This is starting to become quite complex, and it's very unlikely MCP7A was designed with any of those goals in mind either...

Plus, a 65nm IGP seems pretty pointless to me in that timeframe!

 

[Jawed]

Rambus at it again

Wants an exclusion order, seemingly for the patent dispute that came up back in July.

http://www.rambus.com/us/news/press_releases/2008/081106.html

The complaint seeks an exclusion order barring the importation, sale for importation, or sale after importation of products that infringe nine Rambus patents from the Ware and Barth families of patents. The accused products include NVIDIA products that incorporate DDR, DDR2, DDR3, LPDDR, GDDR, GDDR2, and GDDR3 memory controllers, including graphics processors, and media and communications processors.

Jawed

 

[fellix]

GDDR5 -- a safe harbor, then?

 

[Kaotik]

GDDR5 -- a safe harbor, then?

While perhaps unlikely, do we actually know that nV HAS a GDDR5 memory controller to begin with?

 

[hoom]

Surprised nobody mentioned it but Charlie is still at it, got a new Macbook pro & gave it an investigation with an electron microscope showing High Lead bumps rather than Eutectic.

Got some official rebuttal saying that they kept high lead bumps & only changed the underfill.

So he started looking at the underfill & may have found a problem there.

In summary NV is saying 'We had a too soft underfill on some chips & changed to a harder one which fixed the issue. Other chips we have switched to eutectic bumps & those are fine too'
While Charlie is saying 'I think they made the switch in underfill too hastily & without properly engineering the change, so the underfill change seems to have only moved the problem because people with the new chips are still having similar issues'
I think he's also insinuating that the eutectic shift they did on some chips may not have been done right too since it should require a redesign of at least some of the rest of the chip?

I dunno, it seems quite possible that they really did just have a too soft underfill & the new harder one better matched the other aspects of the NV design which fixed the issue without going too stiff thus producing the new problems Charlie is insinuating.
On the other hand, NV has a track record of lying & cheating when they made engineering fuckups in the past so I wouldn't be surprised if they were doing the same thing again :neutral:

 

[Rufus]

Because clearly Charlie has a PhD in materials engineering, and single-handedly knows more than NV's and TSMC's packaging teams.

 

[Arun]

I can't see where he said he tested the underfill to see whether it really was massively higher Tg; so I assume he simply didn't. I'm willing to bet he's making a major mistake in assuming that's what they did and it's the only factor there. The NV quote says the underfill is more "robust"; anything beyond that is mere interpretation.

Michael Hara made clear in a CC (although it's not clear whether he was supposed to say that publicly, good ole Hara!) that the problem was only present at a specific temperature RANGE; so if you went above that range, it'd actually be okay. This doesn't fit my very limited understanding of a straightforward "too low Tg" underfill problem.

Seems likely to me Charlie is even more blinded by his own arrogance than NVIDIA! Of course, if he was right -- OUCH. But that's not the most likely explanation here I think... (although perhaps NVIDIA decided to find a sweetspot between different kinds of failures, and the failure rate is likely to still be higher than normal, just 'Good Enough' for OEMs not to give a shit - who knows really, there are plenty of possibilities beyond Charlie's Black & White)

 

[Bouncing Zabaglione Bros.]

Because clearly Charlie has a PhD in materials engineering, and single-handedly knows more than NV's and TSMC's packaging teams.

Charlie obviously went to a lot of people who know better and asked them about it. He doesn't claim to be a materials engineer, but he claims to know where to find them and their electron microscopes. He's obviously been fed a lot of this info, the same as when he scooped the story of the high failure rates and the underfill/bump issues.

TSMC will fab your design, they won't make your design for you, as shown by the original set of problems with underfill and lead bumps. It doesn't look like either TSMC or Nvidia were too clever there. Another screwup in an attempt to put a patch on top of the first screwup isn't beyond the bounds of possibility. If it's a stopgap measure until a next gen redesign, then it becomes more understandable why Nvidia would do it.

Nvidia are still selling the old chips that they know will fail alongside the supposedly fixed ones, so it's obvious that Nvidia is willing to let the customer take the hit in the short term so they can keep making money and produce a fixed product some time down the line. A quick patch that's cheap and just moves the problem elsewhere in a problematic design to tide them over for a year is just their sort of response.

 

[Sxotty]

Actually Charlie had errors that I could spot in his explanation. He was completely incorrect in some of his descriptions and it was obvious to anyone that has done dynamic mechanical testing in a temperature controlled environment.

I do not have a great deal of knowledge about many of his assertions, but the fact that he is wrong or grossly oversimplifies things (so he is wrong) on the issues I know something about tends to diminish his credibility in my eyes.

 

[ShaidarHaran]

Actually Charlie had errors that I could spot in his explanation. He was completely incorrect in some of his descriptions and it was obvious to anyone that has done dynamic mechanical testing in a temperature controlled environment.

I do not have a great deal of knowledge about many of his assertions, but the fact that he is wrong or grossly oversimplifies things (so he is wrong) on the issues I know something about tends to diminish his credibility in my eyes.

You make the same mistake Charlie often makes then, in that you make an assertion and fail to back it up with evidence.

IOW: explain.

 

[Sxotty]

You make the same mistake Charlie often makes then, in that you make an assertion and fail to back it up with evidence.

IOW: explain.

I did? Where did I write and publish an article? I fail to see it, but when I direct you to something I publish I will be sure to back up assertions I make.

 

[v_rr]

Nvidia warns about weak fourth-quarter revenue

NVDA 7.81, +0.20, +2.6%) on Tuesday warned that its fourth-quarter revenue will come in lower than it previously expected. The Santa Clara, Calif.-based chip maker now sees revenue declining 40% to 50% sequentially, blaming weak end-user demand and inventory reductions by its channel partners. On average, analysts surveyed by FactSet Research were looking for revenue of $805.3 million for the quarter ending Jan. 25.

http://www.marketwatch.com/News/Story/Story.aspx?guid={0DCA49D2-694F-4CC0-91A9-4C9F5932B844}&siteid=yhoof2

 

[dizietsma]

Nvidia warns about weak fourth-quarter revenue


http://www.marketwatch.com/News/Story/Story.aspx?guid={0DCA49D2-694F-4CC0-91A9-4C9F5932B844}&siteid=yhoof2

I just saw that, it is really shocking. Wonder how AMD/ATI faired ?

 

[Arun]

Weak? This is not just weak; this is a mind-blowingly massive miss. They'll make a HUGE loss. Q4 is normally their best quarter of the year, too... Here's hoping at least part of the problem is an inventory correction, although I suspect the Quadro business just fell off a cliff given the current capital investment dynamics worldwide.