Background:
Standard network cards first arrived on the scene as ISA cards.  Later, the PCI version became the de facto standard that we all know and love.  We now have USB network adapters available from several companies.

Let's talk numbers.  Network cards generally come in two flavors: 10 Mbps and 100 Mbps.  This refers to how fast data can be transmitted through the network card across the network.  Theoretically, this translates into 1.25 Megabytes per second for 10 Mbps cards and 12.5 Megabytes per second for a 100 Mbps network card.  The PCI bus on your motherboard can easily handle these speeds.

All the USB network adapters are labeled "10/100 Ethernet Adapter" or "10/100 Fast Ethernet Adapter".  USB or Universal Serial Bus is currently at standard 1.1 which runs at two different speeds:
 

If you're like me, you're now talking to yourself - "Wait a minute!  Umm if the maximum speed of a USB connection is 12 Mbps, why is the USB network adapter labeled "10/100"."  Something doesn't smell right.  What's going on?  Are the USB network adapters comparable to 10/100 PCI NIC's?  At first glance, it would seem so.  I mean, if I walked into a store and looked at two different network adapters that both have "10/100" on the box, I would expect both to perform comparable even though they used different interfaces right?  But since we know the max speed of USB 1.1 is 12 Mbps, we know there is going to be a problem.

Another thing to consider is that all of these USB network adapters require an operating system that supports USB.  This means some Macs, Windows 98, Me, 2000, or even 95 with heavy patches.  If you use Linux, you might have a really tough time finding driver support.

The contestants:

I picked the network cards that are currently most popular and also widely accessible to mainstream consumers, so I went to Best Buy, CompUSA, Fry's, etc.

Click to Enlarge

The price that I paid is for reference only.  For the most part, I paid too much.  If you are planning to drive down to the Best Buy and pick up a card, you'll probably pay around what I paid.  However, if you do a bit of searching on the Internet or look for sales, you should be able to get a better price.  In the end, it depends on the convenience of getting the card right away versus having to wait for a package to ship if you bought it on the Internet.

Here are the cards tested:
 

Methodology:

The first thing you should know about network cards is that they are rarely the limiting factor.  The biggest hurdle in your system is your hard disk.  Even though network cards can be saturated by  newer high-end hard disks, most regular hard disks can't, especially when transferring several small files.  The seek and read times of the hard disks are just too slow.  In order to get around this hurdle, I moved to the other white meat: RAM.  RAM is thousands of times faster than hard disks, so this is how I configured my test systems.

Server:  AMD K6-III 333 MHz

This was the system that had the network card that would never change.  This server was running Windows 2000 Advanced Server with IIS 5.0 FTP server.  I created a 200 megabyte RAM disk on the server using RAMdiskNT.  Let me tell you, this is a cool product.  It's shareware and is limited to 100 uses without registering.  If you choose to keep using the product, the cost is $35.  This program allows you to create a disk drive from RAM.  You can specify the size of the RAMdrive using the included applet.  I didn't use the RAMdrive that comes with Windows because that program only allows up to 32 megs of RAM to be used as a RAMdrive.  By using RAMdiskNT, I could be sure that the hard drive would not be the bottleneck.  The server did not have any other application running on it and was disconnected from the Internet.

The network card in the server is a Linksys Etherfast 10/100 v.1.0 using the drivers that came native with Windows 2000.  You may wonder if this card might be a limiting factor in my testing.  Let me tell you, I'm confident it was not.  During the course of testing, I found that when a network card  is receiving data at full speed, there was often 100% cpu utilization.  However, the network card that is sending the data usually uses much less cpu time.  While sending files, the Linksys card in the server never exceeded 40% cpu utilization which tells me that there was still plenty of juice left in the card and that it wasn't the limiting reagent.

Client: Intel Celeron 366@550 Mhz

CPU: Intel Celeron 366@550 Mhz
Motherboard: Tyan Tiger 100
RAM: 256 Kingston PC100 CAS 2

Yes, finally I got my computer stable at 550 Mhz.  Previously, it was running at 412 Mhz.  Actually, this machine is a dual cpu machine, but since I ran Windows 98 SE on it, only one of the cpu's was actually being used.  I chose to use Windows 98 SE on the client machine because it's the operating system most people are using these days.  Windows 2000 is very popular with the computer savvy group of techies, but most consumers and gamers still use Windows 98 SE.

This client also had a RAMdrive of 200 megabytes using the Windows 98 version of the aforementioned software: RAMdisk98 (Registration is $25 versus RAMdiskNT's price of $35). 

The only software running on the client computer was WS FTP, a file transfer program.  After I had configured the client computer without any network adapter or network drivers, I made a clone of the hard disk using Norton Ghost.  After each network card was finished testing, I restored the fresh image of the hard disk to the client .  This was to insure that the network drivers of one network card would not interfere with subsequent network cards.  Time consuming, but necessary.

The server and client were connected to a Allied Telesyn LANedge 8 port 10/100 Switch which was isolated from the rest of the network and disconnected from the Internet.

The test consisted of transferring a 200 megabyte file from the server to the client - from the RAMdrive of the server to the RAMdrive of the client.  The 200 megabyte file is made of MP3's of Britney Spears songs zipped into one large file.  (Oops!.. I Did It Again)  This was to make a 200 megabyte file that is nearly incompressible.

For each network card, I installed the card using the drivers included in the box, regardless if Windows 98 SE already came with native drivers.  I then ran 3 practice runs to get the server and clients ready.  The 200 megabyte test file was then transferred 10 in each direction (server to client, client to server).  WS FTP keeps track of how long it takes to transfer the file which I wrote down.  The fastest two 2 times and slowest 2 times were thrown out and the rest were then averaged.

Results:

I calculated the speed of each network card by dividing the size of the files by the time it took to transfer the file in seconds then multiplied the whole thing by 8 (to convert bytes to bits).  The speeds you see here are the actual speeds the file was transferred, minus the overhead and extraneous network jabber.  Simply stated, the speeds represent actual data transfer and does not include network overhead.

Here are the speeds in Mbits/second.
 

There first thing you notice is that the numbers, for all intensive purposes, are identical.  The next thing you will notice is that the numbers are identically SLOW.  (Or was that the first thing you noticed?)  These numbers are not even close to what a 10/100 PCI network card can do.  Another interesting point is that the speeds shown here aren't even close to the upper limit of the 12 Mbps that USB 1.1 can handle.

To get a better idea of how these numbers translate into working numbers, here is the same data presented as Megabytes/second.
 

This numbers look even worse as Megabytes per second.

The other aspect of these USB network adapters that I examined was the CPU utilization when downloading at full speed.  These numbers are very crude and have an error of probably more than 5% which begs the question, why did I even put them here?  The reason is because that's what my best guess was. 
 

These number are good in the sense that the CPU isn't being stressed too much.   Most 10/100 PCI networks cards reach 100% cpu utilization when downloading at full speed.  The reason that the USB network adapters don't reach 100% is because the USB network adapter isn't able to work fast enough to stress the cpu.

Interpretation:

Needless to say, I was disappointed with the performance of these USB network adapters.  What bothers me the most about these adapters is not the performance, but instead the labeling and marketing of these USB adapters.  All of the USB adapters are labeled "10/100" or "Fast Ethernet" which in my opinion is very deceptive advertising.  The manufacturers may argue that they mean the adapters are compatible with 10/100 networks, but this is like saying a Pinto is compatible with the Autobahn.  It's ridiculous.  The worst part is that these adapters aren't even capable of 10 Mbps, let alone 100 Mbps.

To illustrate this point, I pulled out a eight year old 3COM Etherlink III PCI 10 Mbps card and a current Linksys 10/100 Mbps card for comparison:
 

All the USB adapters get smoked again.  They can't even hang with the 10 Mbps card.

Conclusions:

I previously stated that the network card is rarely the bottleneck in a network.  In this case, I was wrong. The USB network adapter along with the USB port both are bottlenecks to network speed.  The upcoming USB 2.0 standard should open things up with a theoretical max speed of 360-480 Mbps.  However, the current batch of USB network adapters won't be able to take advantage of the new speed.  You would need to buy new USB network adapters that work with the USB 2.0 standard.  If you have a network with heavy traffic, a USB network adapter is not the way to go.  If all you do is surf the web from a DSL or cable modem, you probably won't see a difference but then again, why limit your LAN speed.

I said this before, but I'm gonna say it again.  If I waked into a store and saw two boxes labeled with "10/100" or "Fast Ethernet" and the only difference was USB or PCI, I would expect the performance to be the same, just the interface to be different.  Here we have shown that the performance is vastly different.  The seasoned computer geek might not fall for this trick (or did you?!), but for the general public, I think it's extremely deceptive and a poor way to dupe people into buying products.

Because of their high cost (2-3x more) and low performance (12x worse than 10/100 PCI) , I cannot recommend purchasing a USB network adapters unless you really need to have one.  Also, you must have an operating system that supports USB which pretty limits you to Macs, a few versions of Linux, Windows 98, Me, 2000 or even 95 with heavy patches.  Here are a few reasons that you might use a USB adapters.

• Your computer doesn't have or ran out of PCI slots
• You have an "Internet Appliance"
• Your computer ran out of IRQ's
• You will void your warranty by opening your computer
• You aren't technically comfortable enough to open your computer
• Your laptop ran out of PC Card slots

Okay, since this is a roundup to compare the different USB adapters, we're suppose to pick a winner here.  Well, the result is that there is a 4 way tie among the different USB network adapters.  They are all priced the same, have the same performance, and have about the same cpu utilization.  How do you decide which one to buy if you indeed plan on buying one?  Buy the one that is the cheapest and has your favorite color or shape because there isn't much else to differentiate one from another.  Really.
 

Take Me Home!