Ethernet performance
We evaluated Ethernet performance using the NTttcp tool from the Microsoft Windows DDK. We used the following command line options on the server machine:

ntttcps -m 4,0,192.168.1.25 -a

..and the same basic thing on each of our test systems acting as clients:

ntttcpr -m 4,0,192.168.1.25 -a

We used an Abit IC7-G-based system as the server for these tests. It has an Intel NIC in it that's attached to the north bridge via Intel's CSA connection, and it's proven very fast in prior testing. The server and client talked to each other over a Cat 6 crossover cable.

We tested the nForce4 several different ways, with and without NVIDIA's Firewall enabled, and with and without ActiveArmor acceleration. We also tested the nForce4 with the firewall included in Microsoft's Windows XP Service Pack 2, just to see how it compared.

The Radeon Xpress 200 system here is using a PCI Express Gigabit Ethernet card from SysKonnect (based on a Marvell chip) that ATI supplied to go along with its reference board. This is the first PCI Express X1 card I've ever seen. Check it out:

Innit cute?

Since ATI has included no Ethernet controller in its chipset, the use of an external PCI-E Ethernet controller chip should be standard practice with the Radeon Xpress 200. The K8T800 Pro system, meanwhile, is using a Marvell Ethernet chip embedded on the Asus A8V Deluxe motherboard and connected to the PCI bus. This, too, is a typical GigE solution for this chipset. Our Pentium 4 system, meanwhile, uses the unfortunate combination of a PCI Express chipset and a PCI network chip, unnecessarily hampering its performance.

NVIDIA's ActiveArmor works more or less as promised, delivering much lower CPU utilization at the price of about 20Mbps of total throughput. Not a bad tradeoff, all things considered. That gives the nForce4 Ultra the best combination of Ethernet throughput and CPU utilization of the lot, even with NVIDIA's firewall active. The Radeon Xpress 200, with the Marvel PCI-E chip, achieves even higher throughput, but it requires about another 20% of CPU time in order to do so.

I was curious to see what would happen to NTttcp throughput and CPU use if I enabled jumbo frames, a provision in Gigabit Ethernet implementations that could potentially improve throughput and lower CPU utilization. I turned on frame sizes up to 9K bytes for the server, the Radeon Xpress 200 system, and the nForce4 Ultra. (None of them had jumbo frames enabled by default.) Here's what I found.

Jumbo frames can indeed boost throughput and lower CPU utilization, but NVIDIA's ActiveArmor acceleration has a throughput problem with jumbo frames. NVIDIA says that ActiveArmor should be able to handle jumbo frames just fine, and they're looking into this problem.