Introduction

However you look at things, it can’t be said that the NV3x / GeForce FX series was a resounding success for NVIDIA. Although their financial reports have still been displaying healthy revenues and profits, looking at those reports over the past two years shows very little actual financial growth for NVIDIA. With an ever widening product range and a growing PC market, “treading water” cannot be something that Jen-Hsun Huang, NVIDIA’s CEO, relishes – not least because ATI, their primary rival, has capitalised on these opportunities, slightly jostling NVIDIA into third place in PC display market share. This is not a year that NVIDIA wishes to repeat, and NV4x, or the GeForce 6 series, is what NVIDIA have been busily producing during this period with the intention of stemming the tide and, they hope, reversing it.

NV30 suffered from being both late and having issues with shader performance. The configuration of NV30 appeared to be very similar to NV25 (or XBox’s NV2A, in fact) and it wasn’t geared toward the latest shader generation. NVIDIA tell us that during the design phase of NV30 certain elements of the DirectX9 specification had not been settled in time so they had to effectively guess them by shooting beyond the specifications. We do, however, get a different picture from ATI though, which when added up possibly suggests that there was some friction between NVIDIA and Microsoft early in the DirectX9 specification setting period. Yet, NVIDIA state that this left them with a part that went beyond some of the DirectX9 specifications, which impacted on the outright performance of NV30 and left a number of optional features out. Indeed the outright performance of NV30 in full precision Shader 2.0 rendering was very overshadowed by the competition, and even the DirectX8 rendering performance was lagging slightly in many cases.

One of the reasons cited for such a performance difference was NVIDIA’s choice to utilise FP32 precision in the shaders, while ATI opted for FP24 – which has a lesser transistor footprint. At a recent NVIDIA Editors Day Jen-Hsun stood up and stated that the move to FP32 has cost NVIDIA several hundred thousand dollars; however, given the performance of FP32 rending in NV30 the more prudent choice may have been to adopt lower precision first. Yet when Microsoft settled DirectX9 they also introduced the Shader Model 3.0, not supported by any hardware in DirectX9’s introduction, and since it's been available for such a long time NV4x could be designed to meet its specification exactly, without much overshoot and certainly no undershoot.

Jen-Hsun also made note that NV30 was far too reliant on process. It's widely believed that NV30 was late due to NVIDIA initially trying to adopt 130nm low-k before it was truly ready and emphasizing and clock speed, neglecting large scale parallelism. He stated that NV4x has been designed not so much with pure speed in mind, but a wider, more parallel architecture – effectively doffing his hat to ATI.

And so, now we have it – NV40, or GeForce 6800, designed to be a very parallel architecture, and hence very scalable as well, with full Shader Model 3.0 and a number of other improvements over previous NVIDIA graphics parts. In this preview we’ll take a look at the features made available in the NV4x line, look at the pipeline overview of NV40, with details on the composition of the major functional units, and finally look at the performance of a reference GeForce 6800 Ultra.