Hasitte
Posts: 10 +1
^^^ Now you are not making any sense!
Does one laugh or cry?
Overpriced, period. The PCIe buss does most of the work.
When you drive to work does the road do most of the work?
I laugh harder then I should!
^^^ Now you are not making any sense!
Does one laugh or cry?
Overpriced, period. The PCIe buss does most of the work.
When you drive to work does the road do most of the work?
I'm Curious - Where does the source of the file copy reside? RAM disk, SATA disk, or another PCIe disk?
This review seems wrong because the disk score reached 1.5GB/s on a sequential read which is PCIe 2.0. On a Xeon-based Linux server with sequential reads and O_DIRECT we can hit 2.9GB/s with this card reliably. With a second card we reached 1.5GB/s, like Steve did, and we were confused... until we plugged it into a Windows machine, removed it, and put it back in the server, achieving 2.9GB/s reliably like the first. We realized at one point we had it in a PCIe 2.0 slot briefly before our tests. It seems these cards can get locked into PCIe 2.0 negotiation and become lethargic even in a PCIe 3.0 slot. Initializing the card with the Windows+Intel NVMe driver woke it back up and it now speaks PCIe 3.0 for double the speed. So... the reviewer may want to recheck this; this card does in fact reach, or in our case exceed, the advertised speed.
2.9GB/s is well above the quoted speed by Intel for sequential reads so that seems unlikely. There are dozens of reviews online now and they pretty much all show the same performance we did, Intel themselves were happy with the results
Our AS SSD sequential read result was 2.2GB/s so the device is clearly using PCIe 3.0.
http://www.nvmexpress.org/blog/crystal-disk-marks-new-release-measures-true-performance-of-nvme/ said:Due to limited support for large queue depths and multiple workers, the full performance of NVMe SSDs could not be seen in Crystal Disk Mark 3.0.4 . For example, the maximum read bandwidth on this build for one NVMe drive was ~ 1.5 GB/s and the maximum IOPs was ~ 150K IOPs.
The developer has completed a new version, revision 4.0.3, that adds deeper queues and multiple worker threads to the benchmark and enables measurement of the true performance of an NVMe SSD. The same NVMe drive now measures over 2 GB/s for bandwidth and ~ 450K IOPs on random I/O. In order to fully saturate the NVMe drives, recommendations are a queue depth of at least 32 and a minimum of 8 worker threads for 4K random testing. Check it out at http://crystalmark.info/?lang=en.