You can't reach 250 MB/s. PCIe overhead is roughly 20%. And you're not helping it with adapter neither, which also will take away some of that bandwidth.

Do I need to wait for SSD SATA drives to advertise 1000 MB/s read/write speeds, for example?

That would be last gen SAS SSD, which has 1.2 Gb/s interface, not SATA. SATA is pretty much dead at this point, being stuck at 600MB/s for years now.

If you have enough PCIe lanes on the motherboard, your best choices are last gen SAS controller (which is compatible with SATA btw) or NVME.

The Serpent Rider wrote on 2021-04-23, 02:13:

You can't reach 250 MB/s. PCIe overhead is roughly 20%. And you're not helping it with adapter neither, which also will take away some of that bandwidth.

From my understanding the "overhead" is the "encoding scheme". PCIe 1.0a x1 spec is for 2.5 GT/s, which is 320 Megabytes per second. Subtract the 20% encoding overhead, we are at 250 megabytes per second. Why can't I reach 250 megabytes per second?

From my understanding the "overhead" is the "encoding scheme". PCIe 1.0a x1 spec is for 2.5 GT/s, which is 320 Megabytes per second. Subtract the 20% encoding overhead, we are at 250 megabytes per second. Why can't I reach 250 megabytes per second?

That's not the only overhead and much more complicated - https://www.xilinx.com/support/documentation/ … apers/wp350.pdf

As I said, the real overhead is 20% of advertised bandwidth. For example, with 4x PCIe 1.1 lanes you can get 800 MB/s on NVME SSD via passive adapter, i.e. direct link to PCIe. It gets only worse if you add additional actors between storage and the bus.

Looks like you may have edited your post after I replied, so I did not get your full message in my above response. From that paper, it seems like the data transfer overhead, or encoding overhead, is the most prominent overhead. The paper goes on to mention another approx 1% loss from packet and traffic overhead. Then some from link and flow control protocol overhead, but I'm not sure what the average is those.

I guess I could use one of those USB 3.2 SSD's to further extend the benchmarks, but these native USB SSD's are too expensive in the 4 TB size. For test purposes, cheapest, smallest (512GB), USB 3.2 gen2 with speeds of 1000 MB/s are $90 CAD. That's too rich just to satisfy my curiosity. Maybe someone else would be willing to benchmark one of these USD 3.2 SSD devices on a PCIe 1.0a x1 slot with a USB 3.0 card?

I guess I could use one of those USB 3.2 SSD's to further extend the benchmarks

I don't see any point in that. It still will be limited to roughly 200 MB/s, maybe just slightly more. That said, you can improve PCIe performance by just simply overclocking it. Most old-fashioned boards can handle 110 Mhz or 10% increase in bandwidth.

It should provide some insight as to how much bandwidth we are loosing from the SATA-to-USB converter and how much bandwidth is lost from these extra PCIe overheads. I find it hard to accept that these "extra overheads", like packet, traffic, link, and protocol are further reducing bandwidth by an additional 15%, that is on top of the 20% encoding overhead. From 320 MByte/s down to 200 MByte/s is a costly 38% reduction.

I may try to overclock PCI express it to 110 MHz like you suggest, but I'm not sure how happy the Adaptec 6805T RAID controller will be.

Best I could get using an x1 PCIe card with a Fresco Logic FL1100 chip was:

208 MB/s read
171 MB/s write

I tested several PCIe x1 cards with Fresco Logic FL1100, NEC D720201, AsMedia ASM1042A, VIA VL805, & Eltron EJ168A. The cards with the FL1100 consistently benchmark a little faster.