I have used most of all those 3Com cards but never did a speed test on them. iirc my fav is the Dec 21041 of that time.

I vaguely remember using a few cards with the string 3C905 but not the specific ones.
Funny how I can remember that and not my Mom's birthday...

The user manual shows the "C" as "3C905C For Complete PC Management"
The square chip in the upper right is some kind of Desktop Management chip.
If I ever had one I wasn't using it's functions.

This https://www.manualslib.com/manual/634205/3com-3c905cx.html
.. covers 3C905C, 3C905B, 3C900B.

I've run my own benchmarks on these cards and also added the Intel Pro 100 S. There is a clear winner from the tests. I'll try to get the results posted sometime this weekend. I ran LAN Speed Test v2 in Win95 and NT4.

Horun wrote:

I have used most of all those 3Com cards but never did a speed test on them. iirc my fav is the Dec 21041 of that time.

The Tulip was nice, although the 21041 was only a 10baseT lan chip. The 2114x chips supported fast ethernet, so I choose them.

Attached are the results for the one Intel and three 3Com cards. The oldest 100mbit 3Com card (3c905) was the clear winner in these tests, followed by the newer 3c905B and Intel Pro100S, with the 3c905C taking the tail. 8 tests were run per network card to formulate an average. Multiple operating systems were used with two points initiating the queries.

I tried added CheckSum Offload for the 3c905c card, but this resulted in an average of 8.3% slower transfers.

On the data table you will also see two non-NIC benchmarks being run: GLQuake and CPUMark99. I ran these with an idle system to see if the mere presence of different NIC's were causing the computer to be slower. GLQuake shows a consistent decrease in benchmark results when the 3c905B or 3c905C were inserted.

In the past, I've noticed that CPUMark99 is rather sensitive to USB and LAN cards on the bus slowing down benchmark results. The 3c905c showed the most performance hit of about 9% compared to the card with the least performance penalty, which was the original 3c905.

It is possible [likely] that the difference between the fastest card, the 3c905, and the slowest card, the 3c905c, would shrink as the test system becomes noticeable faster.

Interesting, I have around 10 of those 3c905b cards that have just accumulated over the years, and one thing I notice Is there are bunch of revisions. I wonder if revisions of 3c905b have difference in speed. In adition i have a couple cards that are 3c905b-txnm without wake on lan, both are differn't board revisions.

Interesting. I'd expect newer cards to be be better. But it all depends on the use pattern.

As for the HW checsum ( Scatter/gather and TX checksums) is not meant to be improving throughput. It is a kind of processing speed/throughput tradeoff. It can only improve TX and only if using specific kinds of system calls (depending on if the data needs to be copied to the user space or not). This should be working also on 905B (Cyclone ASIC) too.

The original 905 hardware (Boomerang ASIC) preserved a legacy PIO/DMA transfer modes which was meant to be a workaround for old flawed PCI chipsets with broken PCI bus master support.The "B" and subsequent revision eliminated that and use bus-master transfers exclusively. In some cases (like large transfers) the simple DMA mode might actually deliver a better throughput (although that's inferior in any "normal" use case). Could that be somehow enabled in your 905?

The DP83840 chip on the original (Boomerang) card is the MII transceiver. It is built-in in more recent designs (Cyclone and Tornado).

Warlord wrote:

Interesting, I have around 10 of those 3c905b cards that have just accumulated over the years, and one thing I notice Is there are bunch of revisions. I wonder if revisions of 3c905b have difference in speed. In adition i have a couple cards that are 3c905b-txnm without wake on lan, both are differn't board revisions.

I only have one 3c905b card, so I cannot see if there is any speed difference between the revisions of the 905b. Perhaps stick it on the list?

mpe wrote:

Interesting. I'd expect newer cards to be be better. But it all depends on the use pattern.

For my purposes, which mostly consist of benchmarking, I find the drop in CPUMark99 scores most disturbing. Because of this, I will be avoiding the use of 3c905c and 3c905b cards and stick with 3c905 and Intel Pro 100S cards in these slower systems.

mpe wrote:

As for the HW checsum ( Scatter/gather and TX checksums) is not meant to be improving throughput. It is a kind of processing speed/throughput tradeoff. It can only improve TX and only if using specific kinds of system calls (depending on if the data needs to be copied to the user space or not). This should be working also on 905B (Cyclone ASIC) too.

Huh, I always thought it would improve performance on slow systems, like 486 and Cyrix 6x86 systems, but I never tested it until now. I will be leaving it disabled from now on.

mpe wrote:

The original 905 hardware (Boomerang ASIC) preserved a legacy PIO/DMA transfer modes which was meant to be a workaround for old flawed PCI chipsets with broken PCI bus master support.

Mostly for PCI 2.0 I presume?

mpe wrote:

The "B" and subsequent revision eliminated that and use bus-master transfers exclusively. In some cases (like large transfers) the simple DMA mode might actually deliver a better throughput (although that's inferior in any "normal" use case). Could that be somehow enabled in your 905?

I'm not sure. I don't see anything in the DOS config or Windows utility to explicitly indicate DMA/PIO modes. I have network optimisation set to Normal on all cards and media type to "auto select (N-Way)". Screenshot provided.

Hard to tell what's wrong.

However there is actually a long list of 900b/905b features that should make it clearly superior to 900/905. Such as:

- 2 KB transmit FIFO and 2 KB receive FIFO
- dual-channel DMA engine
- improved scatter-gather engine that reduces the number of I/O operations required to support data transfers.
- hw checksums
- bigger multicast filter
- support for VLANs
- much improved boot ROM support
- improved PCI commands incl. ability to do bursts that are double in size compared to 905
- and several new commands that could be used by driver

There are three different version of Cyclone ASIC (40-0502-00x, 40-0476-00x and 40-0483-00x) randomly appearing in 905b series cards and each of these can was further released in about 4 or 5 revisions with different errata.

You can never rule out some driver/mb/nic incompatibility in your specific combination...

How does the 3Com 3c905 and Intel Pro 100 compare with the Linksys lne100tx and the Netgear Fa310tx?

Stretch wrote:

How does the 3Com 3c905 and Intel Pro 100 compare with the Linksys lne100tx and the Netgear Fa310tx?

Any particular reason you want to know this information? My objective of this thread was just for selecting a network card for my K5 build rather than undertaking a full-on PCI NIC comparison for older systems. I may have a few other brands of PCI network cards if these results are something of importance to you.

I went digging through my pile of network cards. For standard 32-bit PCI cards, I found these:

Netgear FA311 Rev-B1 (DP83815DVNG chip)
Netgear FA311 Rev-D2 (DP83816AVNG chip)

Linksys LNE100TX Rev 5.1

D-Link DGE-530T Rev-A1 (88E8003-LKJ chip) - this is the only NIC of the group with a 1085 VRM onboard.
D-Link DFE-530TX Rev-C1 (DL10030B chip)
D-Link DFE-538TX Rev-F2 (DL10030C chip)
D-Link DFE-538TX Rev-C1 (DL10038 chip)
D-Link DFE-538TX Rev-E1 (DL10038D chip)

Intel Pro/100S (82550EY)
Intel Pro/100S (82550GY)
Intel Pro/100S (GD82559)

Intel Pro/1000MT (no working Win95 drivers. edit: just tested this card and it is dead)
NE-320G-TX1\2B (Realtek RTL8169S-32 1000mbit card)

D-Link certainly went through a lot of revisions. I wonder what happened there? I'm guessing that the Linksys and D-link cards are lousy. They certainly look cheap.

EDIT: I have more than a dozen of the D-LINK and Intel cards. I went looking through the Intel cards and nearly all were 82550EY, but I found a 82550GY and a GD82559. Any idea what the differences are?

This would be a fairly large undertaking as there are an additional 11 card revisions here, each with 8 tests.

No need to benchmark. Around 2000-2002, When I had my original Athlon 550 MHz and Duron 800 MHz computers I had the Netgear FA310tx cards and Linksys lne100tx.

If I am not mistaken, they were budget NICs with broad OS support, so I was just curious feature-wise at the hardware level why they were so cheap compared to 3com and Intel.

EDIT: I will not de-rail the thread any further, I will research these NICs on google

The 3c905C is keyed for both 5V PCI and 3.3V PCI slots - there are two notches in the card edge connector instead of just one. This is relevant on server boards and certain embedded boards with 3.3V PCI slots, but essentially every desktop motherboard has 5V PCI slots.

Stretch wrote:

No need to benchmark. Around 2000-2002, When I had my original Athlon 550 MHz and Duron 800 MHz computers I had the Netgear FA310tx cards and Linksys lne100tx.

If I am not mistaken, they were budget NICs with broad OS support, so I was just curious feature-wise at the hardware level why they were so cheap compared to 3com and Intel.

EDIT: I will not de-rail the thread any further, I will research these NICs on google

Let us know what you uncover about the differences. On a purely visible level, the Netgear FA311 looks like well made card in comparison to the Linksys and D-Link.