feipoa wrote:

krcroft wrote:

Here's a link to SanDisk's A2-rated Extreme Pro - https://www.amazon.com/dp/B07G3H5RBT/ Yes, you will find many more "Extreme Pros" that cost less, but they won't be A2 rated!

I noticed that the card in that photo says SDXC. Does SCSI2SD support SDXC, or only SDHC? If SDXC works, then here's an even cheaper A2, https://shop.sandisk.com/store/sdiskus/en_US/ … -UHSI-Card-64GB

Yes, might as well get the smallest A2 card if it's only for testing. That's a great price, although on the specifications tab it shows:

Capacity: 32 GB
Read Speed: up to 90MB/s(5)
Write Speed: up to 60MB/s(5)
Form Factor: microSDHC
Video Speed: C10, U3, V30(4)
SD Adapter: Yes

... 32GB? Also doesn't mention Application speed ratings or show A2 anywhere.. (although the product number and photo indicate 64GB and A2). I would assume the spec page is just in error given the photo and product code look good.

This card should be so fast in your 486 that everything upstream from it become the greatest rate limitters, so you're mostly measuring the latency of the software, memory to cpu latency, isa bus latency, internal scsi to sd latency, and finally the card itself.

Do you have a DOS benchmark that tests random read and write? That's where you should see the largest differences among all of your cards.

Ahh, that's a good point. I went to the Amazon site for the same Extreme card and it shows 160 mb/s read and 60 mb/s for write, while the Extreme Pro card shows 170 mb/s write and 90 mb/s write. So perhaps I should source an Extreme Pro card. I wish the card sizes were smaller, like 8-32 GB. That would save $, but I'm guessing SanDisk's market analytics team decided it wouldn't be sufficiently profitable.

If I can find a cheap enough CF SanDisk Extreme Pro card at 160 MB/s UDMA7, the comparison between SCSI2SD and ACARD-CF would be pretty fair. While I saw some marketing brocure for a 16 GB Extreme Pro 160 MB/s card, I can only find as small as 32 GB for sale.

I also noticed that there aren't any A2 cards which are UHS-II. Different focus for the target application I guess.

Did you view the Excel file? Is the Random Read, Random Write, Sequential Read, and Sequential Write benchmarks provided by WinBench96 insufficient? If so, perhaps you can recommend another program for DOS/Win3.11? Once I can compare the more extreme SD/CF cards, I'll move the assembly to a faster computer and retest.

Sorry, I should have double checked your original post - yes, the 200 to 4096 byte random IO tests are perfect.

Interesting they don't have an A2 + UHS-II card; I wonder if there's some technical trade off where they give up some peak sequencial throughput for improved IOPS / latency.
When you test on your fast system, I've read some reports that the card-reader itself can bottleneck these cards; if you can find one bundled with SanDisk's USB3.1 reader, then that's probably the best you can do to eliminate this risk.

feipoa wrote:

Have you tried formatting an XDHC A2 card as FAT16 or FAT32 and work in systems designed around SDHC?

I have, many times, and it works, but as I've said before there isn't any guarantee. Let me try to put it in points:

- The cards can be the same brand and electrically the same, may even be the same silicon actually since any dies with defects will be sold as smaller capacity cards, but the control firmware might be different.
- The above is even more true for A1/A2 cards vs non-A, clearly there is some difference that makes the random I/O faster on one but not the other model.
- There are only a few big Flash manufacturers so even brand-name cards of the same model can actually be different chips inside, or possibly with different FW.
- As mentioned earlier the cards themselves can have various quirks, same goes for the devices they will work in, you could just be unlucky and end up with somehow incompatible combo.
- And then there are fake cards, especially of the more expensive and faster models - a huge subject that I'm not going to touch here.

So point is, I always recommend cheaper cards. If you need that capacity you are forced get a (possilby expensive) SDXC and reformat it to FAT. You'll have to experiment yourself, in general it should work but sometimes just doesn't. I have some CF cards that don't work in older DOS machines - either not detected at all, or have weird DMA quirks that confuse the IDE controller so much even floppy drives stop working. So it's not a problem exclusive to SD cards.

Thanks a lot for your insight. Sounds like in terms of compatibility, everything is a maybe.

I would say that nothing done on Vogons is really "need"ed. People here like to experiment and have fun with it. That is often the more interesting avenue than the end goal. In this case, the end goal was to get a replacement SCSI solution, e.g. SCSI2SD, but more interesting for me is how much incremental gain can be had on obsolete hardware. Then see how much the 486SXL is slowing down this gain compared to a significantly faster system.

This is interesting. I have a 430HX board +P233 out for testing and I decided to plug in the Voigon 133x CF card to the onboard IDE port and run Winbench96. The Disk Winmark scored 651 on the P233. From the charts on page 2 of this thread, the same Voigon 133x card scored 656 on a 486SXL-40 system. What can we attribute the higher score to on the significantly slower system? SCSI? Is it that the 486SXL is not CPU limited in these tests? The 486SXL is using ISA SCSI, while the 430HX board is using onboard PCI IDE.

I did order an A2-class SD card so we can see how much more can be squeezed out of this SCSI2SD.

feipoa wrote:

What can we attribute the higher score to on the significantly slower system?

With a 1% difference, my guess is they are statistically equivalent and they are within the random variability of performance. For example, if you reran the same test 3 times on the same system, how much does each test differ?
The good news is that for this CF card, the 486 doesnt add enough latency for it to impact performance, likewise this CF is similarly maxed out in the pentium and there isn't any extra performance the pentium can extract with its faster CPU and higher IDE spec.. so can say that that this CF is the bottleneck within both systems.

feipoa wrote:

I did order an A2-class SD card so we can see how much more can be squeezed out of this SCSI2SD.

Looking forward to the results!

feipoa wrote:

Voigon

Off-topic: The name of this manufacturer is amusing 🤣

krcroft wrote:

feipoa wrote:

I did order an A2-class SD card so we can see how much more can be squeezed out of this SCSI2SD.

Looking forward to the results!

Also ordered a faster CF card!

Stiletto wrote:

feipoa wrote:

Voigon

Off-topic: The name of this manufacturer is amusing :lol:

Yeah, made me wonder where the name came from...

Eeek, I can't quote reply anymore. Does everyone else see my above response as in my photo below? What did I do wrong?

feipoa wrote:

Eeek, I can't quote reply anymore. Does everyone else see my above response as in my photo below? What did I do wrong?

Yes, I'm seeing the quoting problem per your screenshot; although I'm not sure why the forum isn't rendering it properly!

Just for fun, I tested my two A1 cards ...

in my Raspberry Pi 3 to see how they compare against the cards here: https://www.pidramble.com/wiki/benchmarks/microsd-cards

(click to make it bigger)

1Card                hdparm buffered    dd write      4KB rand read   4K rand write
2SanDisk Ultra A1     20.67 MB/s        18.62 MB/s    8.86 MB/s       3.87 MB/s
3SanDisk Extreme A1   22.10 MB/s        20.44 MB/s    8.95 MB/s       3.94 MB/s

The A1 cards come in near the top of the pile and win the random write test, which is where users feel the most pain when performance falls off.

Dividing the random results by 4K, we get ~2200 read and ~960 write IOPS, exceeding the A1 spec (1500 read, 500 write). Side note: if you're wondering why the entire lot of very fast cards perform so poorly in large-block sequential throughput, it's due to the Raspberry Pi organization using an extremely slow bus-attached SD controller.

What's also interesting is that Samsung's two "+" models (EVO+ and PRO+) and the 16 GB SanDisk Extreme provide enough IOPS to be A1 rated. That said, this benchmark only exercised 100MB of uncached IO, and I suspect these cards woulnd't be able to sustain an A1 load for the entire capacity of the card with zero let up (which is what I personally saw when using non-A1 cards in my Pi.. huge write stalls after extended writes; the card would go into some kind of garbage collect mode).

Your inbound A2 card guarantees sustained 4000 read and 2000 write IOPS, which would deliver 16MB/s random read and 8MB/s random write - and dominate the benchmark above. If we cut the transaction size back from 4K to 200-bytes as used by WinBench 96 (in your original post), the results become 781KB/s random read and 390KB/s random write. That assumes zero upstream latency, so you'll probably see some decent percent of these numbers.

(Posting these RPi benchmarks isn't at all meant to compare against your numbers given the CPU speeds, bus speeds, controllers, software, and filesystems are completely different, but figured it would be interesting to compare A1 vs non-A1 cards under a different test environment).

feipoa wrote:

Also ordered a faster CF card!

You don't mess around 😎! The anticipation continues to build 😀

feipoa wrote:

Eeek, I can't quote reply anymore. Does everyone else see my above response as in my photo below? What did I do wrong?

Quote_no_quote.png

You clicked the "Disable BBCode" checkbox on the Full Reply screen while writing your post, I fixed it.

krcroft wrote:

feipoa wrote:

Also ordered a faster CF card!

You don't mess around :cool:! The anticipation continues to build :happy:

Yeah, and they would have been here by now had I just signed up for my free trial of Amazon Prime!

Stiletto wrote:

feipoa wrote:

Eeek, I can't quote reply anymore. Does everyone else see my above response as in my photo below? What did I do wrong?

Quote_no_quote.png

You clicked the "Disable BBCode" checkbox on the Full Reply screen while writing your post, I fixed it.

Ahhh, a case of the mouse being quicker than the eye!

feipoa wrote:

Ahhh, a case of the mouse being quicker than the eye!

Maybe clicker than the eye?

One week later and the items haven't arrived yet. Amazon must really want you to sign up for Prime! by using these delayed shipment schemes. I didn't even get an "It has shipped notice" until yesterday.

Here is an update using the following microSD and CF cards.

The sequential write speed of the A2 SD card improved over the non-A2 SD card, but squential & random read decreased, as well as the random write speeds, yet the overall disk WinMark score improved slightly, 1%. It would have been nice to see all the results increase with the newer SD card, but I guess there are some trade-offs in using A2 cards. I'm curious how a UHS-II 300 MB/s card would perform, e.g. https://www.amazon.ca/Sandisk-Extreme-Pro-Mem … ctronics&sr=1-6 Perhaps in 5 years I can get this card for $10 and test it out.

As for the CF card, there was only a 6% improvement in the WinMark score with the 150 MB/s SanDisk Extreme Pro. I have my doubts that I would be able to get any CF card faster than the SD cards on this system.

With the values shown in the attached table, which card would you leave in the system - SD: Sandisk Extreme, class 10, UHS-I, U3 or MicroSD: Sandisk Extreme Pro, V30, UHS-I, U3, A2?

I would go with the A2 card given it wins the random tests, which is the the most likely IO pattern for common OS usage. It also roughly ties the leads on sequential IO, which is less important.

All of these storage options are maxing out your system except for the SCSI HDD which only delivers half the random IO rate as the A2 card.

I would also go with the A2 because its lower-bound performance is guaranteed; if you could tell WinMark to benchmark each storage's entire addressable size, the other cards would falter at some point. Some would never recover and deliver a trickle for the remainder of the test. Others might trip up countless times but regain speed in between.

The A2 will provide smooth performance for the entire span of storage. I would only wager that the SCSI HDD would keep a similar steady, albeit slower, pace.

If you have a fast modern system with good USB 3 ports & drivers, it would be interesting to see them stretch their legs (with something like http://www.hdtune.com/, which has random, multi-random, and sequence tests that span the entire storage)

Thanks for the analysis. I'll use the A2 card.

I have an AM2-based system running XP SP3 with a PCIe USB 3.0 card, but my USB-to-SD card reader is from the USB 2.0 era. Should I bench it with hdtune?

feipoa wrote:

Thanks for the analysis. I'll use the A2 card.

I have an AM2-based system running XP SP3 with a PCIe USB 3.0 card, but my USB-to-SD card reader is from the USB 2.0 era. Should I bench it with hdtune?

You certainly can't go wrong with that A2 card; it doesn't get much better than that for OS-pattern IO!

I also haven't splurged on a descent UHS-I or II bus reader either, but will probably make the jump when UHS-III cards come out. For now, I wouldn't bother with the benchmarks.

The reader itself will impose a large amount of latency. It will also force your cards to step down their bus speeds from UHS-II (or UHS-I) to "High Speed" (which is 25 MB/s), in addition to USB 2.0's overhead and latency. USB 2 will also cap throughput around 30MB/s, but the SD-HighSpeed's 25 MB/s is even worse 😀

Apparently Lexar makes some blazing fast UHS-II readers which, at half-duplex, can move 312 MB/s in one direction... and some of these cards can come pretty close to that! I think my SSD would have a hard time keeping that pipe saturated 😀

Here's a bit more info on those HS / UHS-I / UHS-II / UHS-III buses: https://www.sdcard.org/developers/overview/bus_speed/