Since the nForce4’s release, the chipset’s ActiveArmor networking capabilities have struggled to live up to their promised potential for reducing CPU utilization by handling TCP checksumming in hardware rather than on the host CPU. Over time, we’ve watched ActiveArmor’s CPU utilization oscillate between impressively low and alarmingly high. Every new driver release seemed to change ActiveArmor’s CPU utilization, and to make matters worse, many found that using ActiveArmor at all could lead to data corruption or instability.
NVIDIA claims the latest ForceWare drivers6.70 for the nForce4 for AMD and 6.85 for the nForce SLI X16have resolved ActiveArmor’s data corruption issues for good, but the fix comes at the expense of CPU utilization. NVIDIA has today confirmed to TR that it’s been forced to scale back ActiveArmor’s TCP offload engine in order to avoid data corruption.
We’ve done some preliminary testing with NVIDIA’s latest ForceWare drivers, and we can confirm that, although ActiveArmor’s impact hasn’t been completely marginalized, it’s less effective than we’ve seen with some earlier driver revisions. The scaled back offload engine yields lower CPU utilization than with ActiveArmor disabled entirely (24% to 42%.) CPU utilization is also lower than what we’ve seen from the popular Marvell PCI Express-based Gigabit Ethernet controllers (35% for the Marvell to 24% for ActiveArmor.)
We’d rather swallow a few extra CPU cycles than deal with corrupt downloads, of course. However, we have not yet tested enough with these new drivers to state confidently whether the data corruption problems have been resolved.
These problems have persisted too long into the life of the nForce4 lineup, and the purported fix may be disappointing to those who bought into the initial ActiveArmor hype. NVIDIA says ActiveArmor’s troubles will be fully resolved in its next-gen products. Rest assured we’ll be busting out our testing tools to double-check once those products become available.