Cardbus Gigabit adapter only gives a moderate boost

THERE IS no reason not to use Gigabit Ethernet on the desktop. But I decided to test the Edimax 4203 cards to see what kind of performance edge a portable can get by such an upgrade.

Thousands of laptops out there - even plenty of new systems on retail shelves - still have 100mbit "fast ethernet" ports. In my test lab, I have one Gateway 7422GX, and an eMachines M6810. They are almost identical systems, based on the Arima K7 system board, both feature an AMD Athlon 64 CPU, ATI Radeon 96xx video, and both have VIA Rhine II fast ethernet ports in the back.

Since I recently upgraded the Gigabit Ethernet switch at my lab to a Jumbo frames aware SMC 8508T, I could finally test a pair of Edimax EP-3203DL cards, that Edimax kindly sent my way some time ago, at full performance. Find the SMC switch review here and pay special attention to the section explaining Jumbo Frames.

Two Edimax EP-4203DL cards

There's not much to say about a cardbus network card on the external side, other than describing it: it features two LEDs, one indicating the speed, and the other the link/activity. Intelligence and common sense has prevailed and the era of "dongles" is hopefully in the past, as the RJ45 connector in this one -like on every modern cardbus network card- is an integral part of the card's enclosure.

Internally, things are more interesting, Edimax uses the ubiquitous Realtek 8169/8110 family chipset on these cards, which means that cross-platform support is good, as Realtek is one of the best companies when it comes to non-Windows OS support, with Broadcom probably being among the worst and most secretive ones I know of. As such, Realtek offers drivers for the chipset used in this card even for Windows NT 4.0, and all Windows flavours going up to WinXP 64bit, and alternative OSs like FreeBSD, Unixware, Linux, Novell 6.x, 5.x, MacOS-X, even MacOS-9 and DOS. Truly impressive.

Link speed doesn't mean real performance.

The drivers are also updated often, with the latest builds dated only a couple weeks ago - June 2006. On a test boot, the card was apparently even recognised by Linux, which identified two NICs, the system's built-in VIA ethernet port as ETH0, and the cardbus card as ETH1, but I didn't proceed with any further testing because I planned to run the benchmarks on a common OS.

Testing begins

Both notebooks ran WinXP SP2 with the latest patches, no antivirus or firewall to remove all possible software bottlenecks, and the SMC switch was disconnected from the Interweb, leaving just the two machines talking to one another. Since the greatest advantage of Gigabit Ethernet is seen when moving big files around -like CD images or multimedia collections- from one system to another I decided to mount a shared folder from the Gateway to the eMachines, and then run Passmark Performance Test on one system, reading and writing data on the shared drive on the other notebook.

The test bench: eMachines M6810 and Gateway 7422GX both Athlon 64 system with 1GB RAM.
The switch is a 8 Port SMC8508T, which supports Jumbo Frames

Reading and writing on a shared drive accurately simulates file reads and writes like those that happen when copying files around from one system to another on a home or corporate LAN - in fact that's what Passmark PT does one the shared drive, create a big temporary file, and test sequential read, sequential writes, and then random seek and read/write operations. To ensure both systems were equal, the drivers were also updated to the latest ones available from Realtek, the chipset maker.

The Realtek drivers support Jumbo frames up to 7k
However, disk sharing performance shows only a ~28% performance boost
over the notebook's built-in 100Mbit VIA Rhine II fast ethernet port.

The results were surprisingly bad, to be honest, I expected a lot better. In one instance, and with "Jumbo Frames" disabled, write performance was marginally slower over the cardbus adapter when compared to the same operation run through the notebooks' bundled VIA Rhine II ports -reads were always faster, it should be nothing, albeit the difference was negligible.

Only after going to the networking drivers in WinXP and manually enabling "Jumbo Frames", was I able to obtain significantly higher scores, and that was with Jumbo Frames set at the maximum possible size supported by the Realtek drivers: 7k - the SMC switch supports Jumbo Frame sizes of up to 9k, the manufacturer claims. Even in that case, the results were: 28.8% faster for sequential read, and 21.5% faster for writes. I repeated the operations several time and obtained similar results. I concluded that the slow bus at 33 MHz -Cardbus- is a big bottleneck for GigE performance. In the same way that Intel PRO/1000MT cards suffer from the 33 MHz PCI bottleneck on desktops.

The Good, the Bad and the Ugly

The Good Inexpensive cards, good support of almost every operating system out there, even legacy/obscure ones.

The card was recognised in BLAG (Fedora Core 3) Linux on the 7422GX
thanks to the use of the ubiquitous Realtek 8169/8110 family chipset.
Benchmarking was performed on WinXP to have a common OS on both systems

The Bad Edimax shipped the drivers on a 3.5 inch diskette. Someone should tell Edimax that most laptops these days no longer have a diskette drive. I had to use an external USB diskette drive to initially load the drivers, then after the chipset was identified through Control Panel, I went to the Realtek web site and updated the drivers.

Finally, the real speed you get, -and I suspect neither Edimax nor Realtek are to blame here- is only marginally faster than 100Mbit fast ethernet. In the Passmark tests, PT showed only a 28% speed increase in reads, and 10% in writes over windows file sharing. The slow speed of the cardbus bus and its associated bottleneck certainly can be blamed on this.

The Ugly The drivers available on the Edimax downloads page are not old, they're ancient, dated mid 2004. I originally tested it with the bundled drivers but quickly switched to the latest ones from Realtek.

The Verdict
I give these cards 3 Fernandos in my one-to-five personal rating scale. If you're a speed freak, someone who enjoys every tiny speed increase you'll certainly enjoy accessing your wired gigabit network from your laptop, even if it just means a ~28% speed increase over fast ethernet. But then, inserting one of these cards blocks the only cardbus port which for instance I use to plug a SCSI controller and large scanner, so this is also a disadvantage, particularly now that single slot cardbus is the norm in the current pool of notebooks in use, rather than the exception.

I believe Realtek makes great chipsets, and Edimax constructs and sells affordable quality devices -often based on Realtek chipset designs- but this kind of kit reminds me of those "USB to Fast ethernet adapters" that used to be sold in the era of USB 1.x ports: it was half an illusion.... as the devices lighted a "100Mbit link" LED, but in reality you never really achieved the full 100mbit ethernet performance due to the bottleneck that was present in the form of the slow USB 1.x bus. Well, with GigE cardbus cards a similar illusion happens, but the slow Cardbus slot -instead of USB 1.x- is the one to be blamed.

Even on desktop systems Gigabit ethernet can't be fully achieved unless you bypass the PCI bus completely, which is why Intel developed the CSA scheme -Communication Streaming Marchitecture- for integration into its motherboards and with the only purpose to bypass the slow PCI bus bottleneck for Gigabit Ethernet. The new expresscard slot promises to bury forever every trace of the old "PC Card" aka Cardbus slot for good, since it's faster, it's serial instead of parallel, and has a compact 26 pin connector to boot.

But guess what, that's of no use for those of us with relatively still new notebooks like my Gateway 7422 which is one and a half years old. Moral of the story: get a EP-4203DL only if you really feel you need to squeeze every possible drop of performance out of your old notebook and your GigE network.?

See Also
SMC 8508T is a Jumbo step for Gigabit Switches
The time to go Gigabit Ethernet is now
Intel's CSA marchitecture: bypassing the PCI bus bottleneck
ExpressCard technology web site from the PCMCIA Association