Audio codec selection balances local CPU usage, network load and acceptable audio quality. The only H.323 mandatory codec is G.711, which uses 64 Kbps in each direction but is fairly light on the endpoint's CPU. Optional codecs are G.723.1 (5.3 or 6.4 Kbps), G.722 (48, 56 or 64 Kbps), G.728 (16 Kbps), G.729/ G.729a (8 Kbps), or MPEG-1 audio. Endpoints negotiate codecs at call setup, and gateways can provide transcoding between codecs if necessary.

In our tests, calls made with Microsoft's G.711 codec kept the local processor hovering at 30 percent, but its G.723.1 codec consistently held the CPU up to 60 percent utilization. At these loads, even a fairly current PC can quickly become swamped.

But there may be hope for desktop-based VoIP yet. Using Quicknet Technologies' Internet LineJACK, CPU utilization for our G.723.1 calls dropped from 60 percent to just 12 percent on the same machine. At the same time, we enjoyed dramatic improvements in audio quality. The Internet LineJACK works by off-loading tasks from the main CPU and performing several operations--digitization, echo cancellation, compression and packetization--in parallel that are normally done in series.

The Internet LineJACK also lets you gradually transition to PC-based VoIP. It offers jacks for a standard analog telephone as well as the PSTN (public switched telephone network), allowing the phone to be used for PSTN calling even if the PC is turned off. The card can also turn the host PC into a single-line gateway between the IP network and the rest of the world.

Unfortunately, the LineJACK offers no great benefit for NetMeeting, though it can serve as a NetMeeting I/O device. The adapter operates most efficiently with Quicknet's own H.323 application (which could call NetMeeting desktops without any problem).

If you are considering VoIP to the desktop using NT, think again. Windows95/98 is more stable and offers better quality because 95/98 sound card drivers tend to be more consistent than their NT versions--a situation acknowledged by H.323 application vendors.

I Hear Your Pain Make no mistake, IP is one of the least efficient methods of sending voice on a data network. Our RADCom PrismLite analyzer showed that a single VoIP call using the G.711 (64 Kbps) codec consistently burned up 168 Kbps--60 packets per second, at about 300 bytes per packet. Remember that audio streams work in both directions, so users on half-duplex LANs will notice that traffic consumes twice the rated amount for the codec, plus a substantial amount of overhead.

In addition to the traffic load, overhead is extremely high with VoIP. H.323 specifies RTP (Real Time Protocol) and RTCP (Real Time Control Protocol) for IP calling. While RTCP traffic is light on the network, the RTP traffic containing audio streams is obese. Considering all framing (Ethernet, IP, UDP and RTP), uncompressed packet overhead is at least 52 bytes. With 296-byte packets, that's 17.5 percent, far higher than voice over ATM's much-reviled "cell tax." Running this traffic over an IPSec (IP Security) tunnel pushes overhead to 23 percent. And if you're running your data over ATM, you'll still owe the cell tax.

Despite the overhead, we found that VoIP traffic can exist on heavily loaded subnets without too much degradation, especially with the G.723.1 codec. On a single 10BASE-T segment hovering at 80 percent utilization, G.723.1 calls could still be completed without a noticeable drop in fidelity.