Intel Optane SSD 905P Review: World's Fastest SSD Gets LEDs
Why you can trust Tom's Hardware
1TB Performance Testing
Comparison Products
You don't just reach into the NVMe SSD box and pull out any drive to compare to Intel's flagship Optane SSD. These drives have to go toe-to-toe with the champ, so we grabbed the powerful new 1TB Samsung 970 Pro and EVO SSDs as well as the other leading flash-based products. We also included the HP EX920 and the RGB Plextor M9Pe, which are two of the more exciting flash-based SSDs on the market. The Optane SSD 905P's biggest challenge comes from Intel's own 480GB 900P and its less-expensive 118GB 800P M.2 SSD.
The Optane SSDs will take some losses early in our tests. Intel's seven-channel controller dedicates one channel to ECC and metadata, which is a slight disadvantage compared to the NAND-based SSDs with eight channels. Most applications respond best to random performance, but some specialty applications buck that trend. For instance, editing 4K video uses sequential data, so simple renders would likely be faster with something like the 970 Pro. Other video-related tasks like effects could go either way. We'll explore those details further in our next Optane review.
Final Fantasy XIV Load Time
We're introducing this new test, which measures load times using Final Fantasy XIV's inbuilt benchmark, to show how some SSDs can load gaming scenes faster than lesser drives. To ensure accuracy, we're presenting an average of five test runs. We expanded our test pool for this test to give you a better understanding of performance with several types of drives.
It's important to understand that these are system-specific tests. We used an Intel Z270 platform and a Core i7-7700K processor running with the all-core enhancement BIOS setting. Over the next few weeks, we'll explore other systems to examine how increased processor core counts impact the results. For now, the results are clear: the Optane SSDs outperform the other products, but not by a large amount.
Sequential Read Performance
To read about our storage tests in-depth, please check out How We Test HDDs And SSDs. We cover four-corner testing on page six of our How We Test guide.
Most SSD vendors spec their drives with several workers pounding the drive at the same time. We managed to surpass Intel's 2,600 MB/s sequential read spec with a single worker. The Optane 905P accelerated to ~2,700 MB/s at queue depth (QD) 2 and stayed steady throughout our test. The NAND-based drives all reached higher peak throughput numbers, but most of them are slower than the 905P at low queue depths.
Sequential Write Performance
We also blasted through Intel's 2,200 MB/s sequential write claim. We knocked on the 2,400 MB/s door at QD2, but it still wasn't enough to take down the two Samsung 970 SSDs and the new SanDisk Extreme Pro.
Sustained Sequential Write Performance
The story changes when we move over to sustained sequential writes. The Samsung 970 Pro with MLC flash retains its high level of performance at QD1, but most flash-based drives can't sustain those speeds once the SLC buffer is full.
The Optane 905P doesn't have that problem, but its performance is less consistent than we expected. Both the 900P and 905P reach similar peaks throughout most of this test. The baseline performance is still very high for both drives, and we didn't see any significant performance degradation.
Random Read Performance
We see more dramatic changes when we move to the random workloads that are a better measure of desktop PC performance. All three 3D XPoint-based drives deliver superior performance at low queue depths. Lower latency typically equates to higher IOPS, so this high performance is largely due to 3D XPoint's superior latency. That's important because latency has a direct impact on responsiveness. The 905P is in a class of its own at QD4 and beyond.
Random Write Performance
Intel also stepped up its random write game with the new 905P. The 900P and 905P have similar performance at QD1 and QD2, but everything beyond that is a demonstration of the 905P's brute force.
70% Mixed Sequential Workload
We describe our mixed workload testing in detail here and describe our steady state tests here.
During this test, Intel shuts down all the negative talk about the 9-Series' sequential performance. If you spend $360 to $1,300 on an SSD, hopefully you abuse it with heavy workloads and merciless multitasking. You don't need to hit these drives with extreme workloads, though, to see more performance than the best flash-based SSDs. You just have to use applications that take advantage of what's available. At realistic queue depths, the 900P and 905P are much faster than any other SSD.
70% Mixed Random Workload
The same trend carries over to mixed random workloads. The Optane SSD 905P takes mixed random performance to new levels. At QD2, the 905P delivers three times more performance than the best flash-based SSDs. As we mentioned in a previous section, this is really about latency and the user experience.
Sequential Steady-State
Optane drives forgo TRIM because they don't have to clean cells with the latency-killing read, modify, write process before they write new data. That's a massive advantage over other drives. Users with heavy write workloads will experience the most benefit.
Random Steady-State
The Optane SSD 905P delivers similar performance to the 900P during our steady-state test. We didn't spot any noticeable difference between the two 9-Series drives, but the Optane drives are remarkable compared to other products.
PCMark 8 Real-World Software Performance
For details on our real-world software performance testing, please click here.
The Intel Optane SSD 905P either matches or leads in every application test, and the 900P isn't far behind. These tests are fairly lightweight and include many applications most of us use every day. Some use these tests to argue that Optane technology doesn't provide enough of a real-world advantage. Operating systems and file systems are designed for hard disk drives, which hinders application performance. As a result, it takes massive improvements in storage technology to see a performance improvement in common applications.
Application Storage Bandwidth
The three Optane SSDs land at the top of the performance list. The new 905P beats the 900P by a small amount, and both 9-Series models hover around twice the throughput of the 8-Series Optane SSD.
PCMark 8 Advanced Workload Performance
To learn how we test advanced workload performance, please click here.
We use the same applications in this test, but this time we precondition the drives with heavy data writes to force a steady-state condition. The drives run through several degrade and steady-state passes before we slow things down and intersperse five-minute pauses throughout the workload. The idle time allows the NAND-based SSDs to perform cleanup operations that restore their performance. Intel's Optane SSDs don't need the cool-down time because the drives overwrite old data without a latency-inducing erase pass.
Total Service Time
Service time is the measurement of latency plus disk busy time. The two Optane 9-Series SSDs lead the charge here, but the Samsung 970 drives aren't far behind.
Disk Busy Time
This chart measures disk activity during our tests. As you can see, the 905P only works for a small portion of the test. The drive blasts so much data to the system memory and CPU that it sits idle while the other components process the data. This impacts system power consumption in several ways.
The drive is so fast that it can drop back into a lower power state faster than the other drives. Also, the processor doesn't have to wait for data, so it can process it more effectively. This actually increases overall system power consumption for a brief time. If you measured the power consumption for five minutes, you would see larger spikes as the CPU processes data faster, but those spikes result in lower overall power consumption.
MORE: Best SSDs
MORE: How We Test HDDs And SSDs
MORE: All SSD Content
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Intel tempers expectations for next-gen Falcon Shores AI GPU — Gaudi 3 missed AI wave, Falcon will require fast iterations to be competitive
Minisforum's AM5 mini-PC gets Ryzen 9 9950X upgrade for $919 — adding 64GB RAM and 2TB SSD pushes the price tag to $1,199
Nvidia revives LAN party after 13 years to celebrate RTX 50-series GPU launch — GeForce LAN 50 is a 50-hour LAN party across four different cities
-
docswag What's even the point of the 960gb model if two 480gbs in RAID0 would end up cheaper? Usually higher capacity SSDs have lower price/GBReply -
svan71 each model is $200 more than i'm comfortable spending, at least they should be 512 and 1024. The storage numbers bs should be dealt with. I purchased a 4tb drive once formatted it's 3.6. Thats 400 gb lost simply making the drive usable.Reply -
dudmont 20957159 said:each model is $200 more than i'm comfortable spending, at least they should be 512 and 1024. The storage numbers bs should be dealt with. I purchased a 4tb drive once formatted it's 3.6. Thats 400 gb lost simply making the drive usable.
Optane doesn't need overprovisioning. No more lost space.
As to the issue of 2 in raid vs 1 big drive. It would be cheaper, but not faster(except in sequential, which doesn't hardly matter, as most users never have circumstances where sequential actually happens). -
Brian_R170 I'd like to see some more data on power consumption. My instinct says that since the idle power consumption of Optane is so much higher than a NAND SSD, a typical desktop system with Optane won't be able to compete on power. However, the last paragraph on performance testing gives some hope at least for systems that aren't idle.Reply -
Brian_R170 20957058 said:What's even the point of the 960gb model if two 480gbs in RAID0 would end up cheaper? Usually higher capacity SSDs have lower price/GB
There's the convenience of not having to deal with RAID. I know my time would be worth the extra $100 to never have to deal with it if I didn't have to.
Plus, in theory, 2 drives in RAID have 3 times as many points of failure. -
takeshi7 You should do the loading benchmarks in CPU reviews instead of SSD reviews. It's pretty obvious that the storage isn't the bottleneck because all of the numbers are so close regardless of whether it's the Intel 905p or just a SATA SSD. The CPU and/or platform must be the bottleneck.Reply -
Gillerer 20957159 said:... The storage numbers bs should be dealt with. I purchased a 4tb drive once formatted it's 3.6. Thats 400 gb lost simply making the drive usable.
The issue wasn't with "making the drive usable", but with the fact that a 4 TB drive has 4,000,000,000,000 bytes (usually with a few hundred thousand extra), which is industry standard and correct, since 1 tera = 10^12.
Where it breaks down is when Windows reports file and drive sizes in base 2 while still using base 10 prefixes. A Windows "terabyte" (really tebibyte) equals 1024^4 = 1.0995*10^12.
If you divide the 4,000,000,000,000 with 1.0995^12, you get 3.637. It *should* be labeled TiB, but Windows uses TB. -
Giroro 20957159 said:each model is $200 more than i'm comfortable spending, at least they should be 512 and 1024. The storage numbers bs should be dealt with. I purchased a 4tb drive once formatted it's 3.6. Thats 400 gb lost simply making the drive usable.
You don't lose 400 GB due to "formatting" - that's a marketing lie that generally hasn't been applicable since floppy disks. On floppies you could change the amount of space the disk could hold by altering the number/layout of the magnetic tracks - but people don't do that kind of thing with HDDs. SSDs are not magnetic so it is literally impossible to change the amount of flash memory available by changing the format. You can't change physical chips using software. Format in this case means the physical arrangement of data, not what filesystem is being used on the disk.
Now, formatting a disk in Windows generally means to initialize a specific file system (FAT, NTFS) on the disk.
When you format (add a file system) a drive with windows, you do lose a few hundred megabytes as the drive needs a space to save where the files are located. But Windows will display this as used space, not by reducing overall capacity.
To windows, 1 GB = 2^30 Bytes = 1073741824 Bytes
To storage marketers 1 GB is redefined to 10^9 = " 1 Billion Bytes" (There was a lawsuit to stop this because for most other industries it is illegal to redefine units of measure, but the storage companies won since "giga" can mean 10^9 in scientific notation)
What actually ends up on the (raw/"unformatted") drive is often a little bit less than the advertised "1 Billion Bytes", and I have not yet found any other explanation other than false advertisement. I don't think this is a case of over provisioning, but it's not like there's many people out there testing various drives for their true capacites.
Since the old definition of 1GB = 2^30 Bytes is very important for a lot of technical reasons, computer scientists invented a new word "Gibbibyte" to use instead. So now the abbreviation is 1GiB = 2^30 Bytes. Windows still uses the old abbreviation of GB (and maybe Mac/Linux?, I've seen it both ways in android) when they really mean GiB. There is a whole set of equivalents to expresses capacities in binary:
1 Tebibyte (TiB)= 1024 GiB
1024 Mebibyte (MiB) = 1 GiB etc.
Basically say the old words as if you were a baby, and you have the new words. Yes, it's very dumb sounding but extremely important for engineers to be able to speak about this stuff in a non-ambiguous way.
Anyways, the reason you have less capacity than you think you should, is because 1 GB is about 7% smaller than 1 GiB. But the gap gets worse as you move between orders of magnitude. So 1 TB is more like 9% smaller than 1 TiB.
Check an online calculator and you'll see 4TB = 3.63798 TiB (which you'll see as 3.6 TB in Windows) = 3725 GiB
TL;DR Hard drive storage capacities are a deliberately confusing mess, and marketers get away with their obsolete "formatting" explanation because most people don't know the difference. All the data you "lose" never existed in the first place and some judge said that's legal, because of technicalities and math. Also, OS makers aren't exactly helping the situation.