Apple Ethernet NB - does it deliver?

[IIfx]

There are a bunch of Apple Ethernet NB nubus cards on eBay right now, some of them sealed in box. The card has it's own on board 68000 CPU which runs a dedicated networking OS to offload network traffic processing off of the host Macintosh. Apple claims this architecture results in overall increased system performance when performing networking tasks.

Other than the 68000, some RAM, and the OS ROM the board seems to use the same Ethernet controller as found in a typical Quadra, the "SONIC" chip.

If the card actually delivers some performance boost due to it's accelerated nature I will probably go ahead and buy one to put in my IIfx. My goal is to have as many co-processors as possible in that machine and it's looking like I may be able to have a DSP or co-processor in every NuBus slot.

Is it possible to win an award for "most overly accelerated 68030 based Macintosh"? [8D]

 

[NJRoadfan]

It'll give that A/ROSE extension in the system folder something to do

https://en.wikipedia.org/wiki/A/ROSE

The card was total overkill. Ironically hardware developers avoided driver development by just cloning Apple's "dumb" SONIC-based Ethernet card and using the built-in driver.

 

[olePigeon]

Now I absolutely have to have one.

 

[Unknown_K]

I have SE II SCSI cards with a full 68000 on them as well, perfect for 020/030 systems.

 

[Gorgonops]

I'm curious what if any level of higher-level protocol acceleration these cards handle. If all they do is act like glorified DMA controllers to take some interrupt load off the CPU they do seem like somewhat egregious overkill.

 

[joethezombie]

IIRC, running the A/ROSE extension will disable the internal data cache on the processor, slowing things down "considerably".  Whether the tradeoff is worthwhile would require... testing!

 

[olePigeon]

So, could someone theoretically make a USB NuBUS card using A/ROSE to hand off all the CPU dependent crap that USB requires?  Might make for a decent USB card for a 68k NuBUS Mac.

 

[olePigeon]

By the way, I ended up buying one for $20.  Overpaid for an ethernet card, but I was immensely curious.  I can test it against my "unaccelerated" generic brand ethernet card.

IIRC, running the A/ROSE extension will disable the internal data cache on the processor, slowing things down "considerably".  Whether the tradeoff is worthwhile would require... testing!

But does the ethernet card itself have a cache it can use?  Maybe it disables the CPU's cache because it uses onboard cache.  I have no clue, just wondering.

 

[Gorgonops]

IIRC, running the A/ROSE extension will disable the internal data cache on the processor, slowing things down "considerably".

Do you have a citation on this, by any chance? Trying to google this turns up an Apple KB that contains some arcania about a need to mark pages that are used for DMA transfers by an MCP card as uncachable, but the MMUs used by the 68030 and up contain registers to denote if a given page is cacheable or not. Assuming the A/ROSE extension is aware of this it shouldn't be necessary to turn off caching altogether. (That would be absolutely devastating to the performance of a 68040-equipped machine in particular.)

 

[IIfx]

I went ahead and bought a new in box Apple Ethernet NB for $22. Anyone have any suggestions for how to scientifically analyze the potential performance boost?

Perhaps I will film the unboxing because it's a new Apple thing from 1991. It's just an ethernet card but it's a <i>cool</i> Ethernet card!

A/ROSE by any other name is A/ROSE. Supposedly Microsoft was mad at the original name for the project - MR-DOS (multitasking real time distributed operating system).

 

[joethezombie]

Do you have a citation on this, by any chance?

I'm reasonably sure I read that in the troubleshooting section of a Turbo 040 manual or in the last minute release notes for PowerCentral.  I'm not at home right now, so I can't definitely say, but those are the materials I've really been delving for the past several months.

 

[Gorgonops]

If it appeared in an accelerator's manual it wouldn't surprise me if it was a problem either with the particular card or it might be a situational thing involving certain hardware/OS combinations. I do remember that when the Quadra's first came out there was a significant category of software that had issues that could only be resolved by turning off the cache; it's very possible that a given machine that one might install a 68040 accelerator in might have some ROM or hardware based quirk that crops up with what A/ROSE does.

 

[IIfx]

I got the card! ;D


 
Packaging and manual:
https://imgur.com/a/jorev
Sorry it is rotated the wrong way, evidently the stock Windows 10 image viewer doesn't permanently rotate the image.

The driver disk included in the box broke AppleTalk in my copy of System 7.1.1 Pro. Perhaps that's due to my odd setup with a Radius Rocket. Further investigation is required.

 

[CelGen]

I have two A/Rose cards myself. If I recall they sold the ethernet card, a multiport serial card, a token ring card and a 3270/5250 emulation card under the standard. I've seen the first three quite regularly but I have not seen the terminal emulation card in over a decade. They must be quite uncommon.

Think of that card as the 90's version of the Killer NIC. https://en.wikipedia.org/wiki/Killer_NIC

 

[olePigeon]

OK, I tested two ethernet cards.  One was a  Farallon Ethermac NB, the other the Apple NB card.

I timed the transfer of the System 7.5 Update 2.0 from my MacBook Pro to my IIci using Fetch using FTP over ethernet.  Macintosh IIci has a regular 5400 RPM SCSI HDD, 128MBs of RAM, and a Daystar Turbo 040 @ 40MHz.

Unfortunately, these times are stunted.  I don't know if it's my cache or what, but every megabyte or so, the HDD would thrash and pause the transfer completely.  Maybe some sort of processing by Fetch.  I'd like to try a different program and see how it goes, but it needs to be MacTCP compatible (which rules out Transit.)

15,939,712 bytes

Farallon - 2 minutes 11.32 seconds.

Apple NB - 2 minutes 2.11 seconds.

Before the first HDD thrash came, the Farallon got to about 130KB/sec.  The Apple NB got to around 190KB/sec.

My HDD is a serious bottleneck for Fetch, but the Apple NB is absolutely faster.

 

[IIfx]

Thanks for sharing your test results olePigeon! Interesting to see it is faster, albeit only slightly faster.

I have the card installed and configured in my IIfx. I am unable to commence testing as my main bridge Mac has gone down hard due to an HDD failure. The IIfx has 82mb of RAM, the stock 40mhz 68030, and an IBM DSAS-3540 500mb SCSI HDD. The IBM has a 14msec seek time for write, 4500rpm rotational speed, and 10MB/sec interface transfer rate. The media transfer rate per IBM is 32-44Mb/sec. It has a 192KB buffer. Not the best HDD for testing NIC speed but it seems to be fast enough for the IIfx.

Ideally I will do some testing of it this next weekend.

 

[olePigeon]

Well, you could use FTPD-enable on OS X (a nifty AppleScript that starts and stops the FTP daemon) and then use an FTP program on the 68k machine like I did.

 

[IIfx]

Ok, I finally got my IIfx running in a stable manner. Sometimes it's cranky because the PSU needs a recap to supply enough power for the tons of power hungry nubus cards. Some days it chooses to work perfectly.

Non Scientific AFP test:

Server: Mac OS 9.2.2 on QSG4 867

16.8MB AudioMedia II recording from the IIfx to the G4: 1min 30sec

15.2MB copy of Ultimate Doom from the G4 to the IIfx: 2min 21sec

10.6MB copy of Wolfenstein 3D from the G4 to the IIfx: 1min 39sec

Observations: File transfer via AFP feels more responsive. System lag on the system is greatly reduced during heavy network traffic. There is an occasional hiccup as the system deals with writing files to disk. Sharing monitor rarely indicates more than 1/2 AFP load on the IIfx side. It's not super responsive but it doesn't render the Mac uselessly slow while performing network transfers.

Subjective improvement? Yes!

 

[IIfx]

The Ethernet NB actually makes a tremendous difference that is not immediately evident just checking raw speed. While the IIfx's 40mhz 030 was slammed unstuffing a huge file for a CAD/CAM package I was able to transfer more games over the LAN to the Mac. The transfer progressed at the normal speed. With a normal NIC the transfer would have been slow at best, if not unresponsive. Apple's marketing hype for the NB isn't just hype. Offloading network processing to a dedicated 68000 has a real effect!