VOGONS


First post, by Sphere478

User metadata
Rank l33t++
Rank
l33t++

Drop your opinion.

Regular pci cards only (pci-x is acceptable if it is backwards compatible with 32-Bit pci) no pci express cards.

usb is fine since they can be plugged into a usb to pci card.

I noticed googling that there wasn’t really a good round up. While I don’t have a very good selection to test and say myself a few others have recommended a intel 1000 gt so I ordered and will try that. I thought for the sake of people googling in the future we could have a thread where people talked about this and gave the card/driver/OS combo they found to be the best for them.

My personal experience on a pci only computer (old dual pentium 1 233 running xp sp3) has been so far:
A 10/100 realtek pci nic is slightly faster than a usb 3.0 realtek gigabit nic on a usb3.0 to pci card and my gigabit realtek pci card doesn’t work in this motherboard. So can’t say there. I’ll soon test a intel 1000 gt

3MB/sec at 80% cpus usage with usb 3.0 gigabit nic
4.5 MB a second 70% cpus usage with 10/100 pci nic
these results are really only comparable to each other though as these cards/dongles are going to perform faster in a faster computer or different operating system. So for comparison I included those details.

On my windows 10 computer I can get about 40 MB/sec out of the usb 3.0 to gigabit dongle

Not a large test group I know so for those searching in the future I figure others can drop their opinions here on combos they have tried 😀 because when I set out to figure this out I was kind of lost as to which direction to go. And finding a thread like this would have helped a lot.

Last edited by Sphere478 on 2021-02-27, 17:19. Edited 10 times in total.

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 1 of 30, by dionb

User metadata
Rank l33t++
Rank
l33t++

Make sure you define which OS you're talking about, drivers make a MASSIVE difference in CPU load, as minimal ones don't offload any CPU cycles but advanced ones can take advantage of chip functionality. So what's fastest in Win2k/XP might be slowest in Win9x and vice versa. Also make sure to get people to post the exact driver version, particularly in case of DOS (packet driver vs ODI/NDS makes big difference too).

Reply 2 of 30, by Oetker

User metadata
Rank Oldbie
Rank
Oldbie

I don't get it, are you soliciting opinions on what people think has the lowest CPU usage, or are you going to benchmark all of them? Everyone will tell you to use the Intel card, like before. Though it would be interesting to test a 100mbit adapter, might have lower CPU usage at 100mbit due to older drivers?

Intel vs 3com vs Killer would be interesting to test.

Reply 3 of 30, by darry

User metadata
Rank l33t++
Rank
l33t++

This is an interesting idea, but the scope is too wide, IMHO . Other than the difference the OS can make, there are these concerns :

- CPU usage results would need to be run on identical CPUs to be comparable
- different motherboards/chipsets can influence maximum NIC bandwidth performance and, indirectly, likely CPU usage even with identical CPUs
- a PCI-X card running in a PCI-X slot will not be constrained by a 32-bit 33MHz PCI's 133 Megabytes/second theoretical shared bus bandwidth limit . At Gigabit link speed, this can be a significant advantage
- NIC driver versions can be a variable as well .
- A test methodology would need to be defined (i.e. Iperf)

In the meantime, there are these results to consider : Why are 3Com NICs in such high regard? They were run on a level playing field and give a good idea of relative performance .

Reply 4 of 30, by Oetker

User metadata
Rank Oldbie
Rank
Oldbie
darry wrote on 2021-02-27, 14:33:

In the meantime, there are these results to consider : Why are 3Com NICs in such high regard? They were run on a level playing field and give a good idea of relative performance .

Interesting, thanks for finding that (and dionb for running those tests). Seems it isn't as clear-cut as 'Realtek bad'.

Reply 5 of 30, by Sphere478

User metadata
Rank l33t++
Rank
l33t++
Oetker wrote on 2021-02-27, 14:05:

I don't get it, are you soliciting opinions on what people think has the lowest CPU usage, or are you going to benchmark all of them? Everyone will tell you to use the Intel card, like before. Though it would be interesting to test a 100mbit adapter, might have lower CPU usage at 100mbit due to older drivers?

Intel vs 3com vs Killer would be interesting to test.

Yeah, I ordered a intel card on your guys advice this is a thread to help others decide. 😀 I noticed there wasn’t one I could find when searching. I’m trying to help! When I get the intel card I’m going to post results here so others can see 😀 would be cool to so a round up but I don’t have near a large enough selection for that.

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 6 of 30, by Sphere478

User metadata
Rank l33t++
Rank
l33t++
Warlord wrote on 2021-02-28, 08:54:
The main thing with the intel nics is the interupt moderation rate. The lower that is the more CPU it uses but its faster on fa […]
Show full quote

The main thing with the intel nics is the interupt moderation rate. The lower that is the more CPU it uses but its faster on fast cpus, the higher it is the less. There is a Extreme setting at least on the MT adapters which is max. It does lower CPU utilization in a measurable way.

Just like 3 com there is offloading of receiving and sending tcp/ip checksum those should be on by default.

last there is the send and receive transmit discripters. Lower the value it uses less ram but transmit speed is also slower, higher the opposite.

for a 586 Id probably put it on extreme setting, memory usage you will have to play with. Guessing on minimal which is 256, but it goes to 2048 max.

Probably I wouldn't put it on 1000 mbs setting and Id lower it to 100Full duplex even on the gb adater on a 586

Tryin to keep info in one place 😀

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 7 of 30, by megatron-uk

User metadata
Rank Oldbie
Rank
Oldbie

Intel will win.

Based on all of my experience, and all things being equal, a modern Intel PCI 10/100 NIC will outperform the others. The RTL8139 is nowhere near in terms of performance - if you look at the feature matrix of the Intel cards, even the lowliest ones have offloading of a whole load of features.

If you talk about gigabit, and get into PCI-X or PCIe, then it's not so clear cut; but Intel poured a huge amount of effort into their chips... and remember that they virtually dominated the integrated NIC sector on motherboards. It was therefore easy for them to lift those designs and put them on a seperate PCI card.

The problem you'll have is tracking down details of the actual part number used; their PCI card 'brand', the 'Pro 100' actually covers many, many chips and many years... with the later designs having many more features than their initial cards.

My collection database and technical wiki:
https://www.target-earth.net

Reply 8 of 30, by Sphere478

User metadata
Rank l33t++
Rank
l33t++
megatron-uk wrote on 2021-02-28, 11:46:
Intel will win. […]
Show full quote

Intel will win.

Based on all of my experience, and all things being equal, a modern Intel PCI 10/100 NIC will outperform the others. The RTL8139 is nowhere near in terms of performance - if you look at the feature matrix of the Intel cards, even the lowliest ones have offloading of a whole load of features.

If you talk about gigabit, and get into PCI-X or PCIe, then it's not so clear cut; but Intel poured a huge amount of effort into their chips... and remember that they virtually dominated the integrated NIC sector on motherboards. It was therefore easy for them to lift those designs and put them on a seperate PCI card.

The problem you'll have is tracking down details of the actual part number used; their PCI card 'brand', the 'Pro 100' actually covers many, many chips and many years... with the later designs having many more features than their initial cards.

Can you upload pics or chip numbers of the best chip to get?

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 9 of 30, by megatron-uk

User metadata
Rank Oldbie
Rank
Oldbie

Best I can do is point you to the Intel product comparison table for Fast Ethernet and Server adapters:

https://ark.intel.com/content/www/us/en/ark/p … onnections.html

https://ark.intel.com/content/www/us/en/ark/p … ter-series.html

That will get you part numbers, but links to the datasheets are gone.

Most of the time you'll find almost all of the cards listed on Ebay and other places as either just "Intel Pro 100" or "Pro 100/S"... they almost never differentiate between part numbers.

My collection database and technical wiki:
https://www.target-earth.net

Reply 10 of 30, by megatron-uk

User metadata
Rank Oldbie
Rank
Oldbie

E.g in the Fast Ethernet list you'll find the "Intel® 82559ER" .. but the link to the datasheet is gone.

Fortunately, dropping that part number into Google gets a hit:

https://www.intel.sg/content/dam/doc/datashee … i-datasheet.pdf

Your mileage may vary, but I'd expect the latest chips and/or the server implementations to probably offer the most functionality.... but we're probably talking about single percentage point improvements over other equivalent part numbers. I would expect any of them to offer substantial cpu improvements over something like the RTL8139. 3Com seemed to shrink back from the end-user add-in card market during the fast-ethernet generation, so there aren't many models to compare against. You might find the odd alternative like a DEC, Netgear or SMC, but they're going to be few and far between.

My collection database and technical wiki:
https://www.target-earth.net

Reply 11 of 30, by Sphere478

User metadata
Rank l33t++
Rank
l33t++

pci killer nic tests coming

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 12 of 30, by Sphere478

User metadata
Rank l33t++
Rank
l33t++

sitting around 4 mb/sec with intel pro gt and killer nic. the killer seems to be a few tenths of a mb/sec faster though

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 14 of 30, by Woody72

User metadata
Rank Newbie
Rank
Newbie

I was wondering that, too. MB or Mb? I'm using a Realtek PCI on my Win 98 machine and typically get 0.8-1.2 MB/s connected to my Win 10 machine. Might try an older Intel card if they're that much better.

Last edited by Woody72 on 2021-04-16, 08:21. Edited 1 time in total.

Modern PC: i7-9700KF, 16GB memory, RTX 3060. Proper PC: Pentium 200 MMX, 128MB EDO memory, GeForce2 MX(200).

Reply 15 of 30, by Doornkaat

User metadata
Rank l33t
Rank
l33t

I'm getting more confused.
Asking those who are more knowledgeable if my understanding of the terminology is correct:
kilobyte = kB (or kByte)
kilobit = kb (or kbit)
megabyte = MB (or MByte)
megabit = Mb (or Mbit)

1 megabyte = 1000 kilobyte = 8 megabit = 8000 kilobit
1 megabit = 1000 kilobit = 0,125 megabyte = 125 kilobyte

Reply 16 of 30, by Oetker

User metadata
Rank Oldbie
Rank
Oldbie
Sphere478 wrote on 2021-04-16, 06:16:

sitting around 4 mb/sec with intel pro gt and killer nic. the killer seems to be a few tenths of a mb/sec faster though

So the realtek was faster for you?

Reply 17 of 30, by Woody72

User metadata
Rank Newbie
Rank
Newbie

Sorry, I've amended my post - I claim lack of coffee. It's the capitalisation of the b that is particularly significant. 'B' for bytes and 'b' for bits.

Modern PC: i7-9700KF, 16GB memory, RTX 3060. Proper PC: Pentium 200 MMX, 128MB EDO memory, GeForce2 MX(200).

Reply 18 of 30, by Doornkaat

User metadata
Rank l33t
Rank
l33t
Woody72 wrote on 2021-04-16, 08:22:

Sorry, I've amended my post - I claim lack of coffee. It's the capitalisation of the b that is particularly significant. 'B' for bytes and 'b' for bits.

Thanks for the quick reply!👍 I'm never too certain about those things. 😅
Especially with lower case k for kilo but capital letters for all the others this never became intuitive for me.

Reply 19 of 30, by megatron-uk

User metadata
Rank Oldbie
Rank
Oldbie
Doornkaat wrote on 2021-04-16, 08:00:
I'm getting more confused. Asking those who are more knowledgeable if my understanding of the terminology is correct: kilobyte = […]
Show full quote

I'm getting more confused.
Asking those who are more knowledgeable if my understanding of the terminology is correct:
kilobyte = kB (or kByte)
kilobit = kb (or kbit)
megabyte = MB (or MByte)
megabit = Mb (or Mbit)

1 megabyte = 1000 kilobyte = 8 megabit = 8000 kilobit
1 megabit = 1000 kilobit = 0,125 megabyte = 125 kilobyte

If you consider the relatively modern terminology of using base-10, yes, but the traditional computing science units that I was taught some 30 years ago are:

1 Megabyte = 1024 Kilobytes

The use of base-10, rather than base-8 numbers is a relatively new concept - you'll see some references to SI-sounding unit names like kibibit and kibibyte and mebibit and mebibyte. These terms didn't exist when most of us oldies were going through school/college/university.

It's why your 2TB hard drive is actually 1.85TB in reality - disk drive manufacturers always use the base-10 version to make it seem larger.

That's why gigabit (1000 megabit) ethernet has a maximum transfer rate of ~125 megabytes/sec (1000 / 8 = 125).

My collection database and technical wiki:
https://www.target-earth.net