The new height of home QLC-Intel® SSD 670P 2TB evaluation

44-layer stack QLC

QLC NAND has been around for almost four years. In the past four years, with the continuous evolution of 3D NAND stacking technology, the stacking layers of Intel® QLC NAND have also increased from 64 layers of the second generation 3DNAND. To the 144th floor of the fourth generation. The increase in the number of stacked layers means that the single-chip capacity is getting bigger and bigger, the read and write performance is getting stronger, and the durability is getting better and better. In terms of official indicators, it is a qualitative leap compared to the previous 660P and 665P. After the LOGO of the EVO certification platform, the packaging of its own products has gradually moved closer to the style of rectangular nesting. The Rocket Lake S processor not long ago is also good, and the 670P this time is also the same. From the outer packaging, you can see at a glance that it is a new product in 2021.

The back reads 5yr LIMITED WARRANTY, which is a five-year limited warranty . However, the upper limit of the write volume of the warranty is not directly indicated. The upper limit of the warranty write volume given by the previous generation 660P is 200PE, that is, the upper limit of the 2TB disk warranty is 400TBW. The 670P based on 144-layer stacked QLC has also been strengthened in life , the upper limit of warranty is increased to 340PE, and the upper limit of warranty of 2TB disk is 740TBW.Although there is still a gap compared to the 600PE warranty provided by various manufacturers for high-end consumer-grade TLC disks, there has been a substantial improvement of 70% compared to 660P, and the lifespan problem that QLC has been criticized for has been greatly improved.

The package contents are nothing special, a solid state disk body and a brochure

670P adopts the current M.2 solid state disk with the most mainstream specification size of 2280, and the label is attached to the front

All components are also concentrated There is nothing on the front and the back, because most motherboards and notebooks only provide single-sided heat sinks, so single-sided SSDs are obviously more suitable than double-sided ones. However, some manufacturers have begun to stick the labels on the back, so that the exposed particles on the front are more conducive to heat dissipation. I wonder if Intel® will learn from

in the future. After removing the labels, you can see that there are four main components on the 670P, one for the main control, One DRAM, two NAND. The

master and 660P are also from Huirong , and the model is SM2265G. At present, there is no detailed information about this master controller on Huirong's official website. However, based on the existing information, some guesses can be made. The 670P supports the NVME 1.4 version protocol, the package size of the main control chip, and supports external DRAM. SM2265G is likely to be a master controller obtained by downgrading PHY to PCI-E GEN3 version on the basis of SM2267, which is somewhat similar to the direct relationship between Phison E16 and E12

DRAM cache is a conventional configuration NT5CC128M16JR-EK 256MB capacity DDR3L-1866 DRAM particles.If it is a general high-end solution, each TB flash is configured with 1GB cache. Intel® only provides 256MB of cache for the 2TB 670P, which is also in line with its entry-level positioning

NAND number is 29F08T2AOCQKI, which is a 144-layer based The stacked QLC flash memory has a single capacity of 1TB, so it is possible to make a 2TB disk into a single-sided

From the layout point of view, the 670P is more loose than the 660P, because the 64-layer stacked QLC particles used in the 660P have a limited capacity and want to make it bigger. The capacity of the disk needs to arrange 4 grains on one side, so it is very crowded

Basic performance test

This time, the latest Intel® Rocket Lake S is used as the test platform, the CPU is I9-11900K, and the motherboard is ASUS ROG M13H. All disks to be tested are installed in the M.2 slot of the pass-through CPU

Crystal Disk Info can correctly read important information such as power-on time, write volume, temperature, etc., and you can also check that 670P does support NVME 1.4 version protocol
First of all we Let's take a look at the Crystal Disk Mark benchmark test

in an empty disk state. The sequential read reached 3461 MB/S, basically reaching the official index of 3500MB/S; the sequential write reached 3009MB/S, exceeding the official index of 2700MB/S; high queue The 4K random test read reaches 1284.73MB/S, which is about 313654 IOPS, reaching the official target of 310KIOPS; the high-queue 4K random write reaches 1408.09MB/S, which is about 343771 IOPS, and also reaches the official target of 340KIOPS. It seems that Intel has no false standard for the 670P indicators, and all of them can basically be achieved. Although there is still a certain gap between the extreme random performance and the current high-end TLC disk (the official read and write index is above 3000MB/S), it is worth mentioning that the random read speed of 4K single queue is as high as 85.18MB/S, which is even Exceeds many high-end TLC NVME SSDs.

Compared with the 660P 512G, the 670P 2T achieved an overwhelming victory. Most projects have doubled or more improvements, with the biggest gap being the seven-read-three-write (70% read/30% write) mixed test. Seven readings and three writings is a scenario simulation that is more in line with the current usage habits of solid-state drives, and can better reflect the operating status of solid-state drives in everyday situations. However, in terms of random writing of 4K Q1T1, 670P is not as good as 660P. The reason is probably that with the increase of the number of flash memory stack layers, the write delay also increases, so the random write backward situation occurs.

The above is the full disk test. Next, I loaded a large number of video files of 1500G for the 670P 2T, which took up about 78% of the space. See if there is any change in the benchmark test. The fluctuation of about 1% can basically be ignored, so in daily use, you don’t need to worry about the slowdown after half of the installation. Master Land 3", respectively record the loading time from clicking the benchmark button until the benchmark screen appears The improvement of
Due to the slow NAND write-through speed of QLC, I used long sequential reads to test the temperature when the 670P is fully loaded.670P temperature performance is very good. After installing the heat sink that comes with the motherboard, the full load temperature does not exceed 50 degrees, which is lower than 660P. What is the peak performance of 670P under different queue depths?

4K random reading, the peaks of IOMETER and CDM are almost in the early 300KIOPS, and they reached their peaks around Q8T8. The peak value of improving
4K random writing is also consistent with the CDM results, between 340 and 350 KIOPS. Writing can reach full speed earlier than reading, and it basically reaches its peak at Q4T4. In contrast, for some high-end NVME disks, although the peak value of IOPS is very high, it also requires the queue depth of Q32T4 or even higher to reflect it.
After this test, we can see that although the 4K random read and write peak value of 670P is not as good as that of some high-end TLC NVME disks, its starting speed is still relatively fast, and its performance is considerable in the case of low queue depth. For most home environments where the read and write queue depth is less than 4, the gap compared to high-end TLC disks is further narrowed. , all in the SLC CACHE range of 670P. SLC CACHE can be considered as a layer of magic shield for TLC/QLC solid state drives, improving the experience of daily use. But people are always curious. After seeing a beautiful woman, they want to know what her makeup looks like without makeup. Next, I will show you the SLC CACHE mechanism of 670P and the real QLC performance hidden behind the magic shield.

It is a description of 670P SLC CACHE given by Intel.670P uses the dynamic and static mixing SLC CACHE scheme. Large-capacity SSDs often use dynamic SLC CACHE, which decreases as available capacity decreases. In the 670P, the dynamic CACHE part remains unchanged when the free space of the entire disk is maintained above 50%, and when the free space is lower than 50%, it enters the full-disk simulated SLC CACHE mode, and the CACHE capacity decreases linearly with the decrease of the free space. When the free space is lower than 15%, the dynamic cache is basically exhausted.
At the same time, in order to ensure the minimum user experience, the 670P also has a fixed layer of static CACHE. For the 2TB version, there is 256GB of dynamic cache when the disk is empty, and the total capacity of 24G of static cache is 280GB.
At this point, a question arises in my mind. The fixed static SLC CACHE is generally provided by the OP (reserved space) of the solid-state disk. This part of the flash memory capacity is not included in the available space and is used alone to provide CACHE or help GC. use. However, the available space of the 670P 2TB version is 2048GB, that is, the 2TB of NAND is all used as storage space, and there is no external reservation. So I think it is very likely that the actual capacity of the QLC NAND 29F08T2AOCQKI is greater than 8Tb (1TB), so that the reserved space inside the NAND can ensure the normal operation of static CACHE

In the above test, I used to fill this 670P with video files About 78% of the space has been occupied, and the 670P 2T version can still provide SLC CACHE that can provide about 110GB, and the anti-speed rate is still very strong. At the same time,
can also see that after the SLC CACHE is exhausted, the QLC NAND direct write speed of 670P is about 250MB/S, which is still significantly improved compared to 670P in the early 100s. The problem is that if these CACHEs are exhausted in a short period of time, how can the SSD recover performance? The most common method in
is "wait"! Yes, waiting is one way.Let the master perform garbage collection by itself, write back the data files, and gradually release the CACHE space. But waiting is always frustrating, especially when you can't see the progress bar, people always feel like it's doing nothing and want to deal with it manually. The disk optimization tool provided by
Windows can handle this problem to a certain extent

but Intel provides a better solution - Memory and Storage Tool GUI (MAS). That is, the upgraded version of Intel® SSD Toolbox

The new version of MAS only supports detection of model, serial number, free space, firmware version and other information for non-Intel® SSDs, while for Intel® SSDs, it can provide diagnostic scans and firmware updates , Secure Erase and Optimization, I want to focus on the Secure Erase and Performance Booster, because the toolbox software provided by most manufacturers does not provide these two functions.

Secure Erase, as the name suggests The SSD can be restored to a "new disk" state, erasing everything and restoring factory performance quickly. The general security erasure needs to enter the linux environment, or some high-end motherboard BIOS will add a security erasure tool, such as the recently measured ASUS M13H and ASRock Z590 Taichi. The secure erase operation in
MAS is very simple, you just need to delete the partition corresponding to the SSD in disk management, then the erase button will light up, and then click various confirmations. The performance is correct, but the data in the disk will disappear, which does not seem to solve the problem mentioned above, and quickly restore SLC CACHE when there is data. This task is completed by the next feature - Performance Booster.This function is specially designed by Intel for QLC disks, which can quickly release SLC CACHE and restore performance in the shortest time. But after all, the 2T 670P has hundreds of GB of CACHE. If it is all exhausted, the recovery time will be calculated in minutes. Go ahead and do not think that the program is dead and force it to close.

After watching SLC CACHE, let's look at it again Let's take a look at the performance of the 670P under heavy load. Use IOMETER to perform 4K random writing with QD=32 for up to 3 hours on the 670P. Under this strong pressure, you can see the performance of the 670P in a harsh environment. The
670P does not perform very well in terms of dispersion, and it still does not enter a stable state for 3 hours. From the discrete graph, we can see that the GC strategy of 670P is relatively aggressive. It belongs to the kind of rushing to write and rush to recycle, so there will be regular dumping points. The last time I saw such a similar disk was the Plextor M9P PLUS. . But the lowest point of 670P and the zero point after all, that is to say, the disk at that moment is almost unavailable. At present, only data center-level products such as P4320/P4420 can run a relatively beautiful dispersion curve in QLC disks that I have seen. Consumer-level QLC disks are still difficult to withstand such heavy pressure. After all, the 670P is aimed at the home environment

summary

Practical first, no worries at home. This is my evaluation of the 670P disk.
Although there is still a certain gap between 670P and high-end TLC NVME disks in some indicators, its improvement over 660P is visible to the naked eye, and the lifespan and read and write performance are almost doubled.
's large CACHE capacity, good performance in CACHE, and optimization of low queue depth, these adjustments of 670P are precisely aimed at home users. In low-intensity home application scenarios, these optimization directions make the experience of 670P not much worse than that of high-end TLC disks.If the price is affordable enough, I think 670P will be a good cost-effective choice in the home environment!