Nvidia, ATI battle it out over fast Intel chipsets

WHEN NVIDIA unveiled its low-mid range Nforce chipsets for the Intel platforms earlier this month, it used the opportunity to create yet another PR spin showing how its high-end Nforce 4 SLI 2x16 chipset beats the Intel 975X in various benchmarks. That's from Sandra memory benchmarks to various system & game tests. This the one used in the phantom Dell Dimension Quad SLI gaming rig, it's claimed, delivers up to 25% improvement in memory runs, and up to 11% gain in game runs.

All that on a single GPU, of course - separate from the "benefits" of SLI which Nvidia steadfastly refuses to license to Intel. Interestingly, when benchmarking the Nforce4 before, the Sandra test differences I was getting from it versus the most optimised Intel settings - at that time it was 955X - were more like one per cent, which is not exactly a stellar increase.

But this raises an interesting question: are Intel's own chipsets really the best showcase for the "Plentiums" - as in Plenty of profit margin? The answer may come from the "other side" - ATI.

I had a look at a Sapphire Intel Crossfire mainboard, the PC-I7RD400, with ATI's own CrossFire chipset, LGA775 socket and, guess what, DDR-1 memory. I compared the results using the best OCZ Platinum DDR-1 DIMMs running at CL 2-2-2-5 at 446 MHz using the Pentium4 3.2 GHz EE CPU overclocked to 3.55 GHz with 892 MHz FDB, vs similarly overclocked Intel 875X chipset, and of course the brand new 975X on Asus P5WD2-E using Corsair's DDR2-533 CL 3-2-2-6 memory with a 3.73 GHz Pentium 4 XE CPU on a 1,066 MHz FSB.

The results were interesting, with 5,639 MB/s int memory bandwidth on the ATI, it was quite a bit better than 5,583 MB/s on the Intel 875X, and very, very near to the 5,748 MB/s on the 975X with much faster FSB and the world's fastest DDR2-533 memory in there.

Come to think of it, this is no surprise - both ATI and Nvidia have a tough fight in the 3D GPU arena, where the performance specs, including fancy multi-channel DDR controllers, advance far faster than the Intel/AMD CPU specs.

So, either of these should be able do bring that design to very efficient North Bridge memory controllers, too.

I'm curious to see what happens with the next-generation chipsets for the Conroe platform and its 1,333 MHz - 1,600 MHz FSB range. Remember, in the meantime ATI will have a new North Bridge chipset with all 44+ PCI-E lanes without having to do link hops to another bridge, like Nvidia has to suffer with now in its 2x16 SLI chipset - so I believe Nvidia will follow suit on that.

Add to this the need for hardware-accelerated SATA-2 RAID, hardware-accelerated HD sound (like mini Creative X-Fi), same for Gigabit & Wireless TCP/IP stack and associated firewall? These, if implemented right, can up your systems' real speed by an equivalent of two or three CPU speed bumps, by removing constant I/O interrupts and CPU handling of those mundane tasks - yeah, some cheap skate mainboard vendors will argue how those high-end users who really want this would simply buy the add-on cards, but my point is: those high-end users who really want the highest-end board will surely choose one that gives them all that fully integrated, without paying extra - I mean, they already pay for expensive dual-core CPUs, SLI/Crossfire capability etc.

Parts of such things were or may be seen in the graphics vendors' PC chipsets - but not in those crafted by the hand of the almighty Intel. Could it be that the highest-end chipsets for Intel desktop CPUs might not be coming from Intel anymore, but from ATI and Nvidia?

This is not to mention the Dempsey/Blackford platform - good for servers, but those high-latency FB-DIMMs are sure not much cop for the 3D workstation jobs, where lower latency DDR2 / DDR 3 RAM would be more welcome.

Socket F Opterons don't have a problem there, as they can simply use the North Bridges from the desktop side, like ATI RD580c, if needed. No official comments come our way, but at least one of the two big graphics vendors is considering a dedicated workstation chipset for the Dual-FSB Dempsey and Conroe CPUs for workstation/HPC server use: how does quad-channel DDR2-667 memory look, combined with dual FSB 1333 and dual PCI-E x16 GPU slots?

After all, the Intel 3D multimedia workstation market is still too large a beast to give it a miss. µ