Notifications
Clear all

OWC Accelsior 4M2 vs. Sonnet’s M.2 4x4 PCIe

(@jdouglass)
Trusted Member Customer

I’ve been doing some quick research on installing a PCIe NVMe card to create a faster raid setup. I noticed that the Accelsior says it achieves 6000MB/s in raid 0 with what I assume are just the included NVMe blades. But if I were to install something like the WD Black blades that can get up to 3400 R/W could I reach up to 12000MB/s as the Sonnet card claims to reach?

If I understand the bandwidth of PCIe 3 correctly, it can support nearly 16 GB/s. If that’s the case what is the bottleneck with the Accelsior card preventing it from attaining the faster speeds?

Quote
Topic starter Posted : 14/01/2020 9:42 pm
(@softraid-support)
Member Admin

The short answer is the Sonnet card is a full length, 16 lane card.
The Accelsior is an 8 lane card, so it fits in any slot.

While the Sonnet requires a (ful length) 16 lane PCI slot, the extra lanes enable the additional performance.

I do not have any WDs for testing, so do not know what the actual vs. theoretical performance would be in the Accelsior or the Sonnet cards. There is significant overhead in PCI, though, where 16GB's is not possible. What the actual limits are, I am not sure at this time.

I just received a couple Accelsiors here so I can test striping 2 units together, and post later on the results.

ReplyQuote
Posted : 15/01/2020 1:06 pm
(@jdouglass)
Trusted Member Customer

I would certainly be interested in your test results as I am planning to stripe two units myself

ReplyQuote
Topic starter Posted : 15/01/2020 2:22 pm
(@softraid-support)
Member Admin

So far, peformance is not scaling, so am investigating with engineers.

It may take a few days before I get answers.

The advantage of the smaller Accelsior cards is they do not use up the critical full size 16 lane slot, which many need for a second video card.

However, I would expect in the Mac Pro to have fully scalable performance with dual cards. There seems to be a performance cap for some reason I cannot fathom.

ReplyQuote
Posted : 15/01/2020 5:33 pm
(@softraid-support)
Member Admin

Testing so far shows 2 cards limited to about 6,700GB/s writes and 7,000GB/s reads. Three cards showed the same performance. So there is a bottleneck somewhere.

(Apple's RAID demonstrates the same limitation, this is not a SoftRAID issue)

I am reseaching whether this is a limitation in the Mac Pro or the cards. It may take time to find out.

This is all "bleeding edge" technology, so there is still a lot to learn about optimizing performance in the Mac Pro 2019.

ReplyQuote
Posted : 16/01/2020 2:34 pm
(@jdouglass)
Trusted Member Customer

Testing so far shows 2 cards limited to about 6,700GB/s writes and 7,000GB/s reads. Three cards showed the same performance. So there is a bottleneck somewhere.

(Apple's RAID demonstrates the same limitation, this is not a SoftRAID issue)

I am reseaching whether this is a limitation in the Mac Pro or the cards. It may take time to find out.

This is all "bleeding edge" technology, so there is still a lot to learn about optimizing performance in the Mac Pro 2019.

I would be putting them into a Mac Pro as well so this interesting. Right now I am using a few of the Express 4m2’s and found that each nvme slot is capped at about 700mb/s no matter what blade you use. Could that be the same situation?

ReplyQuote
Topic starter Posted : 16/01/2020 3:06 pm
(@softraid-support)
Member Admin

No, because 2 4m/2's will provide double the performance on two thunderbolt buses.

The reason for the PCI limitation is there is no real need for faster PCI, when 4 blades will easily flood the Thunderbolt bus. The price of the enclosure would be much higher and there would be no user benefit, except for those who use it in JBOD mode, which is a minority.

ReplyQuote
Posted : 16/01/2020 8:35 pm
Share:
close
open