Nvidias Jen-Hsun Talks NV4x

[Robin B]

Hentet fra Beyond3d.

 

 

 

The Morgan Stanley Semiconductor and Systems Conference was held earlier today, and among the speakers was NVDIA’s CEO, Jen-Hsun Huang there to discuss NVIDIA’s business and its prospects. Very early on in the conference Jen-Hsun was asked if he’d like to talk about the NV4x generation of parts to which, in a possible reference to the performance of the NV3x generation, his quick response was "Nothing would give me more pleasure to talk about NV4x", and so he did.

 

Jen-Hsun noted that that NV4x series was a new architectural generation that has been designed to offer specific goals: give more programmability, more performance, take advantage of the PCI-Express platform, have higher yields and to be very scalable.

 

When looking at the performance element of NV4x, Jen-Hsun expects the performance increment from the previous generation to be dramatically higher than any previous architectural transition they have previously been through. Indeed, presumably speaking about NV40 specifically, NVIDIA’s CEO states that "if we’re not a lot more than 2 times faster I’m going to be very disappointed". Upon discussing where such performance increases could come from he made note that due to the programmable nature of the graphics pipeline and that now applications are making use of this, more and more elements can be brought over from the CPU world to enhance the instruction execution performance, and its expected that NV4x will adopt a lot of these techniques.

 

In terms of scalability, as with previous generations, NV4x will span a top to bottom line of graphics processors for the PC market space, however there has been a greater emphasis in NV4x’s design to implement this goal. Its expected that by the end of the year there will be an entire family of NV4x processors spanning the very high end performance space, right down to the entry-level market; while this strategy sounds similar to their previous generations what marks it as different this time is the number of processors that may be available. In the NV1x and NV2x generation more or less one two distinct processors were introduced and produced at the same time, with the NV3x architecture this increased to 3 processors, however Jen-Hsun made note that potentially as many as 5 distinct NV4x graphics processors may be in production at any one time.

 

The NV4x generation is also designed to be fully PCI-Express compliant, and take full advantage of the benefits this bus architecture brings. The indication here is that the parts produced on the NV4x platform will be introduced with a native PCI-Express interface – rumours suggest that NV40 will be AGP compliant, but Jen-Hsun’s comments raises the possibility that NV40 may be a PCI-Express chip but the board may utilise their bridge chip to enable it to operate in an AGP system. It was noted that their strategy of not porting any of their current line over to PCI-Express does mean they will need to utilise their bridge chip for the low end initially as its they are not expecting to see the entry level NV4x part available until the end of Quarter 3 ’04.

 

Yields on the NV3x line of chips appears to have been a bugbear of NVIDIA over the past year, and that is something they are hoping to address with NV4x. Jen-Hsun spoke of "heavily patented technology" utilised in the design of NV4x in order to bring the yields up, but there was no expansion on just what this technology was. They noted that due to the fast cycling nature of the market they are not able to get the similar types of benefits as CPU vendors by refining a design to bring costs down – it was noted that for these reasons and due to the issues of yield which are not able to be resolved in the timescales available for one platform NVIDIA actually has the largest "scrap" (wasted die) of any semiconductor in the industry, which is probably a good reason for their margin performance over the past year. Jen-Hsun made note of his desire to see a "100% yield coming out of TSMC" – while NVIDIA talked up IBM last year, its clear that TSMC is rapidly becoming NVIDIA’s primary foundry partner in a more vocal sense, as the number of processors produced from IBM for NVIDIA are still likely to be dwarfed by those that have continued to be produced by TSMC.

 

Jen-Hsun commented that the rumours are suggesting the announcement of NV4x in "a couple of short months" and that these are "likely to be correct".

 

 

http://www.beyond3d.com/#news10574