Corsair MP700 PRO 2TB PCIe Gen5 NVMe SSD

[jeisom]

That worked. Thanks.

[schmidtbag]

It appears that samsung may have engaged in some unhealthy "lets boost some already meaninglessly high and irrelevant metrics" firmware practices, leading to diminishing performance in other areas across the board.

Yeah they have an annoying tendency to do crap like this. That's why their phone cameras have these sensors with absurd MP count that accomplish absolutely nothing.

So it comes with a massive block of a cooler, needs a fan and an extra power supply for it, looking like a prototype straight out of a lab, and its performance only stands out for a couple of sequential read and write tests, but is otherwise worse than an older and much cheaper WD Black SN850? WTF??

I was thinking the same thing verbatim. What a stupid product.

[jeisom]

Here is the result of a pair of FireCuda 530 2TB in RAID0. System is 5900x with kernel 6.5.12 on XFS. At time of writing they are $149(USD) though I got mine at $130 each.

SQLite 3.41.2
Threads / Copies: 2
Seconds < Lower Is Better
Solidigm P44 Pro 1TB .... 11.239 |====
Inland PCIe 5.0 2TB ..... 33.500 |===========
WD_BLACK SN850 2TB ...... 9.798 |===
Samsung 990 PRO 1TB ..... 138.169 |============================================
Corsair MP700 2TB ....... 45.192 |==============
Corsair MP700 PRO 2TB ... 37.018 |============
Firecuda530-2TBx2-RAID0 . 22.177 |=======


SQLite 3.41.2
Threads / Copies: 4
Seconds < Lower Is Better
Micron 7450 Max 3.2TB ... 2.241 |=
Solidigm P44 Pro 1TB .... 14.799 |====
Inland PCIe 5.0 2TB ..... 49.124 |=============
WD_BLACK SN850 2TB ...... 14.408 |====
Samsung 990 PRO 1TB ..... 162.114 |============================================
Corsair MP700 2TB ....... 69.412 |===================
Corsair MP700 PRO 2TB ... 67.995 |==================
Firecuda530-2TBx2-RAID0 . 29.020 |========


SQLite 3.41.2
Threads / Copies: 8
Seconds < Lower Is Better
Micron 7450 Max 3.2TB ... 3.436 |=
Solidigm P44 Pro 1TB .... 21.733 |====
Inland PCIe 5.0 2TB ..... 66.331 |=============
WD_BLACK SN850 2TB ...... 22.512 |=====
Samsung 990 PRO 1TB ..... 217.247 |============================================
Corsair MP700 2TB ....... 97.323 |====================
Corsair MP700 PRO 2TB ... 92.354 |===================
Firecuda530-2TBx2-RAID0 . 37.414 |========


SQLite 3.41.2
Threads / Copies: 16
Seconds < Lower Is Better
Micron 7450 Max 3.2TB ... 6.182 |=
Solidigm P44 Pro 1TB .... 37.925 |======
Inland PCIe 5.0 2TB ..... 97.601 |==============
WD_BLACK SN850 2TB ...... 37.040 |=====
Samsung 990 PRO 1TB ..... 296.780 |============================================
Corsair MP700 2TB ....... 150.985 |======================
Corsair MP700 PRO 2TB ... 146.675 |======================
Firecuda530-2TBx2-RAID0 . 42.288 |======


Flexible IO Tester 3.35
Type: Random Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 4KB - Job Count: 32 - Disk Target: Default Test Directory
MiB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 3952.72 |======================
Solidigm P44 Pro 1TB .... 5408.20 |==============================
Inland PCIe 5.0 2TB ..... 5835.11 |================================
WD_BLACK SN850 2TB ...... 4738.27 |==========================
Samsung 990 PRO 1TB ..... 4688.10 |==========================
Corsair MP700 2TB ....... 5842.19 |================================
Corsair MP700 PRO 2TB ... 6103.32 |=================================
Firecuda530-2TBx2-RAID0 . 8057.10 |============================================


Flexible IO Tester 3.35
Type: Random Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 4KB - Job Count: 32 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 1011896 |======================
Solidigm P44 Pro 1TB .... 1384497 |==============================
Inland PCIe 5.0 2TB ..... 1493786 |================================
WD_BLACK SN850 2TB ...... 1212995 |==========================
Samsung 990 PRO 1TB ..... 1200154 |==========================
Corsair MP700 2TB ....... 1495600 |================================
Corsair MP700 PRO 2TB ... 1562448 |=================================
Firecuda530-2TBx2-RAID0 . 2062618 |============================================


Flexible IO Tester 3.35
Type: Random Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 4KB - Job Count: 1 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 229812 |======================
Solidigm P44 Pro 1TB .... 408631 |========================================
Inland PCIe 5.0 2TB ..... 218115 |=====================
WD_BLACK SN850 2TB ...... 385203 |=====================================
Samsung 990 PRO 1TB ..... 401965 |=======================================
Corsair MP700 2TB ....... 225921 |======================
Corsair MP700 PRO 2TB ... 217759 |=====================
Firecuda530-2TBx2-RAID0 . 463440 |=============================================


Flexible IO Tester 3.35
Type: Random Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 4KB - Job Count: 32 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 863869 |=========================
Solidigm P44 Pro 1TB .... 1414209 |=========================================
Inland PCIe 5.0 2TB ..... 534024 |===============
WD_BLACK SN850 2TB ...... 1027419 |==============================
Samsung 990 PRO 1TB ..... 1521820 |============================================
Corsair MP700 2TB ....... 494129 |==============
Corsair MP700 PRO 2TB ... 518317 |===============
Firecuda530-2TBx2-RAID0 . 438738 |=============


Flexible IO Tester 3.35
Type: Sequential Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 1 - Disk Target: Default Test Directory
MiB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 6537.42 |======================
Solidigm P44 Pro 1TB .... 7036.01 |========================
Inland PCIe 5.0 2TB ..... 9402.69 |================================
WD_BLACK SN850 2TB ...... 7012.42 |========================
Samsung 990 PRO 1TB ..... 7165.40 |========================
Corsair MP700 2TB ....... 9436.50 |================================
Corsair MP700 PRO 2TB ... 11295.84 |======================================
Firecuda530-2TBx2-RAID0 . 12745.57 |===========================================


Flexible IO Tester 3.35
Type: Sequential Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 1 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 3269 |========================
Solidigm P44 Pro 1TB .... 3518 |==========================
Inland PCIe 5.0 2TB ..... 4701 |===================================
WD_BLACK SN850 2TB ...... 3506 |==========================
Samsung 990 PRO 1TB ..... 3583 |==========================
Corsair MP700 2TB ....... 4718 |===================================
Corsair MP700 PRO 2TB ... 5648 |==========================================
Firecuda530-2TBx2-RAID0 . 6373 |===============================================


Flexible IO Tester 3.35
Type: Sequential Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 32 - Disk Target: Default Test Directory
MiB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 2159.80 |========
Solidigm P44 Pro 1TB .... 7045.07 |==========================
Inland PCIe 5.0 2TB ..... 8527.66 |================================
WD_BLACK SN850 2TB ...... 7013.81 |==========================
Samsung 990 PRO 1TB ..... 4709.29 |==================
Corsair MP700 2TB ....... 8446.21 |================================
Corsair MP700 PRO 2TB ... 10452.83 |=======================================
Firecuda530-2TBx2-RAID0 . 11497.78 |===========================================


Flexible IO Tester 3.35
Type: Sequential Read - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 32 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 1080 |=========
Solidigm P44 Pro 1TB .... 3522 |=============================
Inland PCIe 5.0 2TB ..... 4264 |===================================
WD_BLACK SN850 2TB ...... 3507 |=============================
Samsung 990 PRO 1TB ..... 2354 |===================
Corsair MP700 2TB ....... 4223 |===================================
Corsair MP700 PRO 2TB ... 5226 |===========================================
Firecuda530-2TBx2-RAID0 . 5749 |===============================================


Flexible IO Tester 3.35
Type: Sequential Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 1 - Disk Target: Default Test Directory
MiB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 5189.86 |====================
Solidigm P44 Pro 1TB .... 6280.23 |=========================
Inland PCIe 5.0 2TB ..... 9649.16 |======================================
WD_BLACK SN850 2TB ...... 6343.61 |=========================
Samsung 990 PRO 1TB ..... 6495.44 |==========================
Corsair MP700 2TB ....... 9624.62 |======================================
Corsair MP700 PRO 2TB ... 10913.22 |===========================================
Firecuda530-2TBx2-RAID0 . 10619.74 |==========================================


Flexible IO Tester 3.35
Type: Sequential Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 1 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 2595 |======================
Solidigm P44 Pro 1TB .... 3140 |===========================
Inland PCIe 5.0 2TB ..... 4824 |==========================================
WD_BLACK SN850 2TB ...... 3172 |===========================
Samsung 990 PRO 1TB ..... 3248 |============================
Corsair MP700 2TB ....... 4812 |=========================================
Corsair MP700 PRO 2TB ... 5457 |===============================================
Firecuda530-2TBx2-RAID0 . 5310 |==============================================


Flexible IO Tester 3.35
Type: Sequential Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 32 - Disk Target: Default Test Directory
MiB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 5236.48 |=====================
Solidigm P44 Pro 1TB .... 6306.77 |=========================
Inland PCIe 5.0 2TB ..... 9681.01 |======================================
WD_BLACK SN850 2TB ...... 5853.69 |=======================
Samsung 990 PRO 1TB ..... 6431.04 |=========================
Corsair MP700 2TB ....... 8740.71 |==================================
Corsair MP700 PRO 2TB ... 10941.60 |===========================================
Firecuda530-2TBx2-RAID0 . 10700.29 |==========================================


Flexible IO Tester 3.35
Type: Sequential Write - Engine: IO_uring - Buffered: No - Direct: Yes - Block Size: 2MB - Job Count: 32 - Disk Target: Default Test Directory
IOPS > Higher Is Better
Micron 7450 Max 3.2TB ... 2618 |======================
Solidigm P44 Pro 1TB .... 3153 |===========================
Inland PCIe 5.0 2TB ..... 4840 |==========================================
WD_BLACK SN850 2TB ...... 2927 |=========================
Samsung 990 PRO 1TB ..... 3215 |============================
Corsair MP700 2TB ....... 4370 |======================================
Corsair MP700 PRO 2TB ... 5471 |===============================================
Firecuda530-2TBx2-RAID0 . 5350 |==============================================


Dbench 4.0
Client Count: 12
MB/s > Higher Is Better
Micron 7450 Max 3.2TB ... 6665.79 |============================================
Solidigm P44 Pro 1TB .... 2849.04 |===================
Inland PCIe 5.0 2TB ..... 2034.43 |=============
WD_BLACK SN850 2TB ...... 4282.48 |============================
Samsung 990 PRO 1TB ..... 548.72 |====
Corsair MP700 2TB ....... 1635.23 |===========
Corsair MP700 PRO 2TB ... 1670.62 |===========
Firecuda530-2TBx2-RAID0 . 1288.26 |=========


CockroachDB 22.2
Workload: KV, 50% Reads - Concurrency: 128
ops/s > Higher Is Better
Micron 7450 Max 3.2TB ... 132046.7 |===========================================
Solidigm P44 Pro 1TB .... 111179.2 |====================================
Inland PCIe 5.0 2TB ..... 88227.5 |=============================
WD_BLACK SN850 2TB ...... 119112.4 |=======================================
Samsung 990 PRO 1TB ..... 15273.9 |=====
Corsair MP700 2TB ....... 69290.3 |=======================
Corsair MP700 PRO 2TB ... 71457.8 |=======================
Firecuda530-2TBx2-RAID0 . 62376.2 |====================


PostgreSQL 16
Scaling Factor: 1000 - Clients: 1000 - Mode: Read Write
TPS > Higher Is Better
Micron 7450 Max 3.2TB ... 55904 |==============================================
Solidigm P44 Pro 1TB .... 28129 |=======================
Inland PCIe 5.0 2TB ..... 39686 |=================================
WD_BLACK SN850 2TB ...... 36504 |==============================
Samsung 990 PRO 1TB ..... 25760 |=====================
Corsair MP700 2TB ....... 32172 |==========================
Corsair MP700 PRO 2TB ... 31796 |==========================
Firecuda530-2TBx2-RAID0 . 32542 |===========================


PostgreSQL 16
Scaling Factor: 1000 - Clients: 1000 - Mode: Read Write - Average Latency
ms < Lower Is Better
Micron 7450 Max 3.2TB ... 17.89 |====================
Solidigm P44 Pro 1TB .... 36.19 |=========================================
Inland PCIe 5.0 2TB ..... 25.20 |============================
WD_BLACK SN850 2TB ...... 27.82 |===============================
Samsung 990 PRO 1TB ..... 40.77 |==============================================
Corsair MP700 2TB ....... 31.57 |====================================
Corsair MP700 PRO 2TB ... 31.94 |====================================
Firecuda530-2TBx2-RAID0 . 30.73 |===================================

EDIT: https://openbenchmarking.org/result/...NE-2311268NE91
Many of the benchmarks were very competitive with PCIe5 drives and got twice storage for the price. In the tests that weren't I wonder if it was the kernel, the filesystem, the RAID or the more core more better. Or a combo of those.

[jeisom]

So it comes with a massive block of a cooler, needs a fan and an extra power supply for it, looking like a prototype straight out of a lab, and its performance only stands out for a couple of sequential read and write tests, but is otherwise worse than an older and much cheaper WD Black SN850? WTF??

The SN850 does really well in the Database tests, exceptionally so in DBench and SQLite which most consumers don't do on their consumer hardware. Yes SQLite is used, but not in a way that would need that much tuning. I suspect the other boards are tuned for workloads the average user does most often. If these were enterprise drives for use in enterprise I'd be more interested in the DB performance for instance.

[Mahboi]

Oh look!
Is it a tank? Is it a Carrier?
No!
It's a Gen 5 SSD!

[coder]

Michael, there are some really weird results in this test run. In particular, I'm struck by how the Micron 7450 Max 3.2TB fell apart on some benchmarks that it absolutely should've handled with ease.

Consider these two:

Are we to believe that a current, high-end Micron Datacenter SSD truly falls apart on sequential reads, when hit with 32 threads instead of 1? I don't.

How did you mount this thing? Was there direct airflow on it? Did you check for thermal throttling? These datacenter SSDs burn up to 15 W, even using 5 W at idle, and are designed to be mounted in a server chassis with substantial linear airflow.

According to the specs, it should be capable of 6.8 GB/s sequential reads @ 128 kB/read and queue depth of 32. That directly contradicts your measurement. Weirdly, we don't see a similar anomaly on write performance, nor in any of the multi-client database tests.

Specs:

https://media-www.micron.com/-/media/client/global/documents/products/technical-marketing-brief/7450_nvme_ssd_tech_prod_spec.pdf

[coder]

It appears that samsung may have engaged in some unhealthy "lets boost some already meaninglessly high and irrelevant metrics" firmware practices, leading to diminishing performance in other areas across the board.

I didn't expect its "pro" / top tier drive to be the runt in this bunch.

It has truly become the intel of ssds ... overpriced and underperforming.

You and schmidtbag sure are quick to jump to conclusions. I look at the same glaring performance discrepancies as you guys and wonder if it's not a problem with the test setup.

As I mentioned above, we should inquire about cooling - was it installed somewhere with direct airflow or with a heatsink (either its own or the motherboard's)? Also, I'd have to ask if the drives were all reformatted, immediately prior to the test, which is beneficial in part because mkfs implicitly calls TRIM -- an empty and trimmed drive will have the maximum space for its SLC buffer. Finally, was it running the latest firmware? There was a major bug in the 990 Pro's initial firmware that caused premature wear and I think likely also hampered performance.

Where it seems to have the greatest trouble are on benchmarks I think probably involve a fair amount of writes. Here, it's probably noteworthy that it's one of only two 1 TB drives included in the benchmarks. Typically, a drive's SLC write buffer is proportionate to its capacity, putting 1 TB drives at a disadvantage. Also, if the drive wasn't reformatted (i.e. a fresh filesystem installed), then its capacity for buffering write could be further reduced. If we're indeed seeing the effect of its write buffer filling up, that might not happen with the 2 TB version.

All the other reviews I've seen of the 990 Pro tested the 2 TB version, and found it to be one of the better client SSDs out there - even besting PCIe 5.0 models, in some cases. None have shown such weird outlier performance as this article.

• https://www.storagereview.com/review...ssd-review-2tb
• https://www.techpowerup.com/review/s...ro-2-tb/4.html

The above TechPowerUp review shows the 2 TB version of the 990 Pro has a maximum SLC write buffer just shy of the Solidigm P44 Pro 1 TB drive included in these tests. If the 1 TB 990 Pro's write buffer is half that big, that would indeed put it at a substantial disadvantage on sustained write workloads.

I'm not saying Michael's test results are invalid - just that they're strange enough to warrant further investigation.

[coder]

Is is possibile to temp throttle an SSD on Linux, so that it slows down when it exceeds certain temperature?

SSDs' own controller chips will throttle them based on the NAND temperature, in order to protect it. As NAND storage cells get hotter, they lose charge faster. This means they need to get refreshed more frequently, ultimately hurting the endurance of the drive.

[coder]

I think the drive temp stuff is getting a problem. Ridiculous.
What will happen soon is to have water-cooling solutions for NVMe SSD disks connected to a long cable inside a 3.5" housing next to the computer etc.

Part of the problem is PCIe 5.0, which requires more power than older standards. However, another issue is that the controller chips for these SSDs are made on comparatively old process nodes (12 nm for Phison's PCIe 5.0 controller). Silicon Motion just announced a new controller to be made on TSMC N6 that should shave a couple Watts off the drives' peak power consumption.

https://www.guru3d.com/story/silicon...t-than-phison/

These PCIe 5.0 drives currently max out around 13 W, which is close to the 15 W maximum specified for the M.2 form factor. The main reason they require such elaborate cooling solutions is to protect the NAND, which is a lot more temperature-sensitive than the controllers. So, I think your scenario of water-cooled SSDs isnt't likely to happen, save for a few niche products that will mainly be used in boutique gaming PCs.

For those who really care about scalable and sustained performance, datacenter SSDs (like the Micron 7450 Max, in this review) are a better way to go. You can mount them at the front of your case, where they'll get plenty of airflow.

[yump]

[/INDENT]
Are we to believe that a current, high-end Micron Datacenter SSD truly falls apart on sequential reads, when hit with 32 threads instead of 1? I don't.

How did you mount this thing? Was there direct airflow on it? Did you check for thermal throttling? These datacenter SSDs burn up to 15 W, even using 5 W at idle, and are designed to be mounted in a server chassis with substantial linear airflow.

According to the specs, it should be capable of 6.8 GB/s sequential reads @ 128 kB/read and queue depth of 32. That directly contradicts your measurement. Weirdly, we don't see a similar anomaly on write performance, nor in any of the multi-client database tests.

Specs:

https://media-www.micron.com/-/media..._prod_spec.pdf

I have a theory. IIRC, NVMe protocol has multiple queues, one per CPU. One queue with 32 I/Os in flight is not the same as 32 queues with one I/O in flight. Especially considering the large block size of 2 MB. I only have one NVMe SSD to check, but if I look at /sys/block/nvme0n1/queue/max_sectors_kb, mine has a maximum actual I/O size of 512K. My understanding (which fio+iostat agrees with) is that a 2 MiB read gets split by the kernel and sent to the SSD as 4 512 KiB reads.

Suppose what this enterprise SSD sees is 32 queues each with 4 requests in flight. Suppose also that its scheduler is tuned for peak latency or fairness instead of throughput. If it round-robins between the queues, that could effectively turn a 2 MiB sequential workload into a 512 KiB random workload.