Raptor Computing Systems Community Forums (BETA)
Third Party Hardware => General Discussion => Topic started by: tle on March 02, 2021, 06:40:43 am
-
I have been waiting for Samsung PRO 980 for quite awhile. Last week I replaced my old 950 with the 1TB 980 PRO.
The performance is quite decent, I do a simple dd test for a 3GB file, the avg write speed is 5Gb/s and read speed is ranging from 6.1 to 6.5Gb/s.
TRIM is also enabled on my end without any issue at all.
I have updated the wiki.raptorcs.com, so if you are looking for new NVMe, you could have confidence that Samsung PRO 980 is fully compatbile
-
Great, too bad there are no more PCIe ports on the Blackbird otherwise I would have mounted it myself. In any case, I am also satisfied with the SATA devices at the end but I hope that the next Raptor motherboard is in ATX format deriving from the Condor, a card that is now done and over and that I do not think will be lost and that they can reuse for another processor. They will also be able to implement dedicated NVme ports if they want ...
-
They'd still need PHBs to attach them to, and there is only a fixed number of PCIe lanes per processor. If you want oodles of NVMe, the big T2 (with both CPU sockets populated) is really your only option; everything else is a tradeoff. This T2 has two NVMe devices, plus video card, SATA (for the optical drives) and FireWire.
-
https://www.tomshardware.com/news/samsung-launches-pm9a1-ssd-cheap-980pro (http://)There is also the OEM version, Samsung PM9A1, equivalent to the Pro 980 but at OEM price (cheaper)[/url]
In terms of performance and the previous comment about using dedicated slots for each of these drives, here are some details:
- the fastest sequential read from these new Samsung devices is 7000 MB/s
- a PCIe 4.0 (Gen4) slot, with one lane (x1) carries 2GB/s
- the Talos II motherboard has a mysterious x4 micro PCIe connector (https://forums.raptorcs.com/index.php/topic,145.msg1146.html#msg1146), it appears to be x4 lanes, that could carry 4 * 2GB/s = 8GB/s for one of these SSDs or two of the previous generation SSDs without congestion
- the PCIe 4.0 slots, x8 can carry 8 * 2GB/s = 16 GB/s, that is enough for two of these new Samsung devices + some spare capacity for legacy SATA all running simultaneously on a single Tri-mode HBA
- the PCIe 4.0 slots, x16 can carry 16 * 2GB/s = 32 GB/s, that is enough for four of these new Samsung devices + some spare capacity for legacy SATA all running simultaneously on a single Tri-mode HBA
For people building workstations based on Blackbird or Talos II Lite it is probably desirable and feasible to have an x8 HBA for SSDs and use the x16 for a GPU
-
I am hoping the next revision of Blackbird would have built-in NVMe M2 slot so I could use the X8 slot for a SAS controller.
-
You still need PCIe lanes to connect it.
-
Just do a quick reading benchmark with Linux 6.5.3, it's around 3361.47 MB/sec which is a bit disappointing
```
$ uname -ar
Linux shrimp-paste 6.5.3-300.fc39.ppc64le #1 SMP Wed Sep 13 12:19:24 UTC 2023 ppc64le GNU/Linux
$ sudo hdparm -t /dev/nvme0n1
[sudo] password for tle:
/dev/nvme0n1:
Timing buffered disk reads: 10086 MB in 3.00 seconds = 3361.47 MB/sec
```
-
Looking at the history of kernel performance changes in Fedora 37, I get the impression that no one does performance testing before releasing a kernel. The variability is high. It's a shame that Fedora developers don't see this.
SSD-Samsung-980-Pro-History-2023-09-20.png (https://www.dropbox.com/scl/fi/9lphey55dagxsp7gg01o9/SSD-Samsung-980-Pro-History-2023-09-20.png?rlkey=q3cc7wwob2hyb393igxdzoq15&dl=0)
-
Looking at the history of kernel performance changes in Fedora 37, I get the impression that no one does performance testing before releasing a kernel. The variability is high. It's a shame that Fedora developers don't see this.
SSD-Samsung-980-Pro-History-2023-09-20.png (https://www.dropbox.com/scl/fi/9lphey55dagxsp7gg01o9/SSD-Samsung-980-Pro-History-2023-09-20.png?rlkey=q3cc7wwob2hyb393igxdzoq15&dl=0)
ppc64le has always been treated as 2nd class citizen mainly due to hardware availability. I am wondering the performance regression also impact the x86_64?
-
shrimp-paste [tle]$ uname -ar
Linux shrimp-paste 6.14.0-0.rc6.49.fc42.ppc64le #1 SMP Mon Mar 10 23:19:40 UTC 2025 ppc64le GNU/Linux
/dev/nvme0n1:
Timing buffered disk reads: 4972 MB in 3.01 seconds = 1650.85 MB/sec
shrimp-paste [tle]$ sudo hdparm -t /dev/nvme0n1
and it's getting worse in Fedora 42
-
Ugh. That's not encouraging.
-
Looking at the history of kernel performance changes in Fedora 37, I get the impression that no one does performance testing before releasing a kernel. The variability is high. It's a shame that Fedora developers don't see this.
I'm using Fedora right now and my impression is they just don't do testing in general, like for example shipping GIMP 3 with no way to go back without shipping one of its dependecies (babl (https://packages.fedoraproject.org/pkgs/babl/babl/index.html) 0.1.112, which looking at the tracker page still isn't released for Fedora).
-
I'm using Fedora right now and my impression is they just don't do testing in general,
Fedora users and even the Fedora Developers are the testers. Unpaid testers for Red Hat.