Computex 2001 - Day 2

More on nForce

We met with NVIDIA today at Computex to address some of the questions we had about nForce after the official launch. 

The big news actually seems to be the real reason that NVIDIA is not producing a Pentium III version of the nForce is because Intel won’t give them a license for the GTL+ bus used on the Pentium III. Of course our next question was how they were able to produce the chipset for the XBOX, which uses a Pentium III processor, without a GTL+ license from Intel. Well it turns out that Microsoft is the one with the GTL+ license for the XBOX and by virtue of being picked by Microsoft, NVIDIA was able to develop a GTL+ chipset.  That license, however, is only for the XBOX and NVIDIA will not be able to take advantage of it for a desktop chipset.  Why wouldn’t Intel want such an advanced chipset powering their processors? Remember that the chipset market is a big part of Intel’s business model – maybe they’re just as scared of NVIDIA taking over the whole market as the rest of the chipset makers out there.

As far as future chipset plans go, NVIDIA wouldn’t give us much info, but did state that they’d follow the same model they took with graphics, taking a current platform and extend it to higher and lower end market segments.  Based on their success in the graphics market, it looks like they could very well be the top dog in the chipset market in a relatively short period of time. Of course, they got there by producing high quality products and staying a step ahead of the competition.  If they can do the same in the chipset market, they very well could dominate.

We also learned a bit more about DASP, but not too much because much of the technology is still patent pending.  Once the patent is granted, they will release a lot more detail.  What they did tell us, however, is that it is basically a cache on the north bridge that does function much like a processor’s hardware prefetch. The nForce IGP uses an 8-way set associative cache and can perform 8-way prefetching. Unfortunately, size of the TLB and cache are not being disclosed at this time.

NVIDIA confirmed that the sound in the standard MCP is indeed identical that of the MCP-D, with the exception of Dolby Digital Encoding. It turns out that the two chips are probably identical in design; it’s just a matter of paying extra for the Dolby license to gain the AC-3 encoding functionality. This is much like the situation with the GeForce and Quadro GPU lines, which are also identical internally with some features simply disabled for the GeForce line. Information has been on the web for quite some time about how to convert a GeForce into a Quadro, and it’s very possible that something similar will be possible with the MCP and MCP-D.

We originally speculated that the TwinBank architecture would probably be able to support a total of 4 DIMM slots if carefully designed. NVIDIA confirmed that in our meeting today, although it’s apparently quite difficult to design a 4 DDR DIMM slot board according to the motherboard manufacturers due to loading issues.

NVIDIA remains tight lipped on pricing since they try to stay away from that side of things, instead leaving it up to the actual board manufacturers. Fortunately, a couple of motherboard manufacturers were being a bit more forthcoming. One suggested 30% less than the cost of a current DDR motherboard plus a GeForce2 MX card.  Another seemed to be indicated a price of under $200, probably in the $150-175 range, exactly the price range that many AMD 760 based motherboards are selling in right now.  Of course, by the time nForce is released in September, it’s very possible that prices on current products will have dropped noticeably.

As we hinted out with our initial coverage, NVIDIA’s StreamThru technology is really a feature of Hyper-Transport where a virtual channel is created along the bus to guarantee the availability of bandwidth for isochronous data. It turns out that StreamThru is a bit more in that it includes an isochronous-capable Ethernet controller. Incidentally, it turns out that the MCP Ethernet controller was not handled in house, but they wouldn’t disclose exactly who did design it. It also turns out that the dual master Ultra ATA 100 controller in the MCP may also be capable of taking advantage of a StreamThru-like virtual channel when transferring isochronous data from the hard drive.