Review: NVIDIA's nForce4

by Ryszard Sommefeldt on 19 October 2004, 00:00

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qa3z

Add to My Vault: x

Features Overview - PCI Express

Bye bye AGP, hello PCI Express. A single CK8-04 bridge (the base nForce4 package) defines 20 lanes of PCI Express. Those twenty lanes are then split 16/4 for graphics/peripherals. Then on an nForce4 SLI (more on the differing product variants shortly) the 16 lanes for a full PEG16X are fed into an SLI switch on the bridge silicon that does bus and data arbitration between the two physical and electrical PEG16X slots on the mainboard. Install two graphics cards on nForce4 SLI and you get a 8+8 split over the switch, with other control signals sent over the smaller SLI connector between boards. Install only card in an nForce4 SLI board and the switch delivers a full PEG16X implementation to the primary slot.

It's pretty simple. The four peripheral lanes are entirely devoted to PCI Express 1X slots on consumer, single-CPU mainboards.

PCI Express on the dual Opteron version of nForce4, nForce4 Pro, is something different. There's a pair of CK8-04 bridges in that setup, giving a full 40 lanes of PCI Express in the system. That's 32 for a pair of complete PEG16X graphics links without the SLI switch (but still data arbitration between the bridges, presumably using a HyperTransport link) and 8 lanes to be used elsewhere. I forsee a setup much like Intel's Tumwater Xeon chipset where four lanes are bridged to a pair of PCI-X segment bridges and the other four used for PCI Express 1X slots. With Pro not announced today, I could be wide of the mark, but some initial speculation can't hurt.

More on nForce4 Pro when that's announced and more on nForce4 SLI soon.

It's clear that nForce4 is split into a number of variants, let's check those out.