Ozzuneoj wrote on 2025-01-25, 17:17:
Oh duh, I guess it didn't register in my mind that those were old generic cards. I saw the Sintechi thing and wasn't paying attention to the brand names or sizes of the cards. Yeah, it is unlikely that 512MB and 4GB SD cards have anything resembling modern specs, so those are probably good baseline SD performance numbers. I don't know if 10-15 year old microSD cards perform any differently, but it looks like any of these are probably more than enough for DOS gaming.
Exactly. The difference between the slowest devices and the fastest devices is going to be 25% on an ISA DOS system and within 50% on most VLB systems. So not particularly noticeable. You are probably CPU bound on everything anyway.
The next batch of systems is the spooky WDMA forest ( late 486 to socket 7 430VX ). If your controller, driver, and storage get along in WDMA mode, it's a different world in a multi tasking OS than if you are stuck in PIO. Warning: Figuring out what magical components to combine to make WDMA work when using 21st century storage devices is not always straight forward. There were a lot of different IDE controllers during this time.
The next batch of systems is sunny PIIX4 UDMA2 land (ATA 33, 430TX through Slot 1). Assuming you get UDMA2 working, other than some the slow CF outlier, everything else is going to be pretty close again around 20MB/s.
After that is 80 conductor cable territory (socket 370 through the end of PATA ). I lump UDMA 4,5,6 together here because there is often only small differences between ATA 66 and ATA 133. Systems frequently have resistance in other parts of the stack, either between the CPU and the south bridge or within in the storage device itself. But this is the area where bridged sata and fast pata ssd's separate themselves from the Sintech devices. So the benchmarks I did are really for the people that have 1GHz and faster systems (and don't use VIA!)
Ozzuneoj wrote on 2025-01-25, 17:17:
The next thing I would be curious about would be performance consistency over time or as the drives fill up. For example, test a few SD and CF cards vs SSDs and HDDs in the following scenarios:
There are already a lot of variables between driver, controller, storage device and bridges that make this a tough problem. And determining what is actually going on is often a black box problem. There's no nice tool that says "WDMA2 accomplishment unlocked!" Throwing in the "over time" aspect to what is already a multi-dimensional problem gives me headaches.
Ozzuneoj wrote on 2025-01-25, 17:17:
Also, I have to wonder, is that an nForce2 chipset thing that allows UDMA5 performance in DOS, or is it specific to that motherboard and BIOS? My Windows XP test system has an Abit NF7-S 2.0 nForce2 Ultra 400 board (no IGP). I would be curious to see if it can do something similar.
So most motherboard BIOS aim for compatibility and they stick to old standards. Nvidia got daring, took an Oak PCI IDE controller, put it into their silicon, and modified their BIOS to that particular controller to do UDMA transfers. Probably adds a little latency, but really opens up the through put. Broke some Linux installs in the process. If you get a Promise PCI IDE controller with BIOS, it's option rom will do similar stuff. But in XP or Win95, you are going to be using protected mode drivers, not the BIOS, which is why not many BIOS's do this.
I did go to the back of the drawer and pulled out some old SD cards. 2 were DOA. one was funky enough in the benchmarking that I wouldn't use it for a build. Seemed like a lot of retries were occurring. The last one was 256GB. it worked OK, in the range of the other devices, but using a 256GB SD on a Sintechi device often leads to suffering, due to the incomplete LBA48 implementation.
