Two in one: Intel Optane Memory H10 (part 1)

Part 1 >> Part 2

SSD caching has been around for a long time, and allows you to squeeze out maximum performance from fast storage devices. In recent years, Intel Optane products using 3D XPoint nonvolatile memory have ruled the realm of small, expensive, and very fast drives. With Intel's third-generation Optane Memory caching memory, Optane’s performance will be able to qualify for a new product segment.

The first drives with Optane Memory were tiny NVMe SSDs, designed to speed up access to slower SATA drives, especially mechanical hard drives. Currently, Intel continues to use Optane SSDs for caching other NVMe SSM drives, focusing on a combination of Optane flash memory and QLC NAND. Now they put both types of SSDs on one M.2 module to create a new Optane Memory H10.

Intel Optane Memory H10 for the first time allows Intel to embed its Optane Memory caching technology in ultrabooks, where there is only room for one SSD, there is nothing to say about SATA. The squeezing of two disks into a single-sided 80mm M.2 module was made possible, in particular, due to the high density of Intel's four-bit Flash memory QLC NAND from Intel. Intel 660p QLC SSD has a lot of free space on the board in versions 1 TB and 512 GB, Optane cache memory can potentially compensate for the shortcomings of QLC NAND in performance and durability. Even with the placement of two types of memory on one module, the design of the H10 is very simple.

Optane Memory H10 does not use any new ASIC or other hardware to make the Optane and QLC disk modules look like a single device. Caching is fully software controlled, and the host system independently accesses Optane and QLC H10. Each half disk has two dedicated PCIe lanes. Previously, all Optane Memory SSDs were PCIe x2 devices, so they lose nothing in the new device. But the Intel 660p uses a 4x Silicon Motion NVMe controller, which is now limited to two PCIe lines. Although, to be honest, 660p almost never required more bandwidth than x2 can provide, so this is not really a bottleneck.

With slow QLC SSDs and fast Optane SSDs on a single Intel device, complex decisions had to be made when determining disk performance characteristics. For two large capacities of H10 (512 GB and 1 TB), the sequential read speed of more than 2 GB / s is claimed. This reflects the ability of the Intel Optane Memory software to extract data from both the QLC and the Optane H10 simultaneously. Recording can also alternate between different types of memory, but the maximum rating does not exceed any obvious limit for the performance of any of the devices. The specifications for random I / O for H10 are between the performance of devices with Optane Memory and 660p SSD, but much closer to the performance of Optane. Intel is not trying to advertise the ideal cache hit rate, but they expect it to be sufficient for real use.
The Optane cache should reduce the load that an entry in the QLC H10 creates, but Intel still evaluates the endurance of the entire device as the same 0.16 disk rewriting per day. This is the same endurance as the 660p QLC solid-state drives.

Intel Optane Memory H10 marketing photos show an SSD with a two-color printed circuit board to emphasize the dual nature of the drive, but in fact the drive board is one-color. The layout of the board is unique due to two controllers and three types of memory, but it also clearly shows the two discrete products on which it is based. Half of the drive with QLC NAND is closer to connector M.2 and is equipped with an SM2263 controller, plus one DRAM and one NAND packet each. Familiar Silicon Motion testing / debugging connections are located at the interface between NAND and Optane. The Optane side contains a small Intel Optane controller, a single XPoint 3D memory package, and most power management components. Intel SSD 660p and earlier solid state drives with Optane Memory had a lot of free space on printed circuit boards; Optane Memory H10 is full, and probably has the most parts among any M.2 SSDs on the market.

At first glance, the Optane Memory software has hardly changed; except that now there is more flexibility in the choice of devices for caching. (Intel has announced enhanced support for Optane Memory for Pentium and Celeron processors on platforms that have already been supported by Core processors.) If the boot volume is cached, Intel software allows the user to cache selected files and applications in the cache to prevent them from being removed from the cache. In addition to this option, there are no cache behavior settings.

Some OEMs selling systems equipped with Optane Memory reported memory amounts as the sum of DRAM and Optane. This would make sense if we are talking about Optane DC memory modules connected to the processor’s memory controller; what is misleading is the fact that the Optane product in question is an SSD.

Initially, Optane Memory H10 will be an OEM-only part. The new SSD will be available to consumers only as a component of new systems, primarily laptops. Intel is still considering the possibility of releasing H10 for retail sale – both as a separate product and as part of the NUC kit, the timing was not announced. Their motherboard partners set the stage for H10 support for almost a year, and many desktop 300 series motherboards already support H10 with the latest publicly available firmware.

Platform Compatibility

Placing two PCIe devices on one M.2 card is new, to say the least. Intel has already installed two SSD controllers on a single printed circuit board – in corporate high-end SSDs, for example, P3608 and P4608, but these drives use PCIe switch chips to split the x8 host connection into two x4 for each of the two NVMe controllers on the board . This approach inflates the TDP of the device to 40 W, which is not at all useful within the limitations of M.2.

There are several PCIe expansion cards that allow you to connect four M.2 PCIe SSDs through a single PCIe x16 slot. Some of these cards include PCIe switches, but most use a host system that supports split PCIe ports to separate one x16 port into four independent x4 ports. Ordinary consumer CPUs do not usually support this, and are limited to x8 + x4 + x4 partitions or simply x8 + x8, and only when the lines are redirected to different slots to support the use of multiple GPUs. The latest processors for servers and workstations are more likely to support port splitting up to x4, but motherboard support for enabling this feature is not available everywhere.

Even on processors, where the x16 slot can be divided into four ports x4, further separation up to ports x2 is rarely or not at all possible. Chips that support many PCIe lanes, such as the narrow x2 or x1 ports, are southbridge / PCH chips on most motherboards. As a rule, they do not support ports wider than x4, because this is the normal width of the connection to the processor.

Based on the above, we tried the Optane Memory H10 with almost all the PCIe 3.0 ports we have, using all the necessary adapters. Our results are shown below:

The Whiskey Lake laptop provided by Intel for this review is, of course, fully compatible with the Optane Memory H10 and will soon be available for purchase with a new drive. Compatibility with older and non-Intel platforms is basically what was expected: only the NAND H10 side is available. On these motherboards, you cannot use two PCIe devices that have a common M.2 x4 slot – they are not able to detect and initialize both devices. There are a few exceptions to note:

First, the H370 motherboard on our Coffee Lake system was supposed to fully support H10, but GIGABYTE released a broken firmware update that supposedly added support for H10: both parts of NAND and Optane H10 became available when using slot M.2 that connects to PCH but it is not possible to enable caching. There are many 300 series motherboards that have successfully added support for H10, and I am sure that GIGABYTE will soon release a revised firmware update for this particular board. Connecting the H10 to a PCIe x16 slot, which is connected directly to the CPU, does not provide access to the Optane side, reflecting the CPU’s lack of support for splitting a PCIe port down to x2 + x2.

The only modern AMD system that we had on hand was the Threadripper / X399 motherboard. All the PCIe and M.2 slots we tried made the Optane H10 visible, but could not detect NAND.

We connected H10 through two different brands of a PCIe 3.0 switch. Avago PLX PEX8747 provided access only to the NAND side, which was to be expected, since it only supports bifurcation of PCIe ports up to ports x4. The Microsemi PFX PM8533 switch supports split up to x2, and we hoped that it would give access to both sides of the H10, but instead only got access to half of the Optane. The Microsemi switch and the Threadripper motherboard may only need a firmware update to work with both H10 halves; earlier generations of Intel PCH may have this potential, but Intel will not provide such updates. Even if these platforms could access both H10 halves, they will not be supported by Intel's Optane Memory caching drivers. True, there is third-party caching software.

Test system

Our primary consumer SSD testing system is the Skylake desktop. It is equipped with a Quarch XLC programmable power supply for detailed power measurements, and is used for ATSB IO trace tests and FIO synthetic tests. But our system is older than Optane Memory, and Intel and their motherboard partners did not want to release firmware updates to support Optane Memory caching on Skylake generation systems. As a result, with the help of our test bench, we can access only half of the disk – to QLC NAND.

As happens with the release of the new Optane Memory, Intel sent us a whole system with the already installed and configured new Optane Memory H10. Therefore, the testing system is now an HP Specter x360 13t notebook with an Intel Core i7-8565U Whiskey Lake processor and 16 GB of DDR4 memory. In previous years, Intel has provided desktop systems for testing Optane Memory products, but the main advantage of the H10 is that it is one M.2 module that is suitable for small systems, and therefore a 13-inch laptop was chosen. Intel confirmed that Specter x360 will soon be available for sale with the Optane Memory H10 as one of the storage options.

The HP Specter x360 13t has only one M.2 type-M slot, so we used Coffee Lake and Kaby Lake systems, which Intel provided for previous Optane Memory releases, to test multiple disk caching configurations or options using SATA. The test results of applications such as SYSmark and PCMark are strongly influenced by differences in processor power and RAM between these machines, so we need to list three sets of ratings for each tested disk configuration. However, our AnandTech Storage Bench I / O tests and our synthetic tests using FIO give almost identical results for these three systems, so we can make direct comparisons, and for each such test we list one set of points for each storage configuration.

Intel Optane Memory caching software is designed only for Windows, so our synthetic Linux-based FIO testing had to be adapted for use in Windows. The configuration and testing procedure is as close as possible to our usual methodology, but some important differences mean that the results of this review cannot be compared directly with the results of our usual SSD reviews and the results published in Bench. In particular, in some cases it was not possible to perform a secure erase or formatting of NVMe from Windows. Our testing typically involves erasing the disk between the main phases in order to restore performance without waiting for the SSD background garbage collection to complete the cleanup and freeing of the SLC cache. In a review of synthetic tests based on Windows, the tests that record the smallest amount of data were run first, and the tests that require filling the entire disk were run last.

Optane memory caching requires the use of Intel drivers. Our usual procedure for Windows-based tests is to use Microsoft’s own NVMe driver, but not to use vendor-specific drivers. Optane caching configuration tests in this review were carried out with Intel drivers, but all tests with one drive (including tests on only one side of Optane Memory H10) use the default Windows driver.

Our standard Skylake test bed is designed to test NVMe solid-state drives in the main PCIe x16 slot connected to the processor. Optane memory caching requires disks to be connected via a chipset, so there is a small chance that congestion on the x4 DMI channel may affect the fastest disks, but H10 is unlikely to come close to saturating this connection.

We try to include detailed power consumption measurements in almost all of our performance tests, but in this review we have to skip many of them. Our state-of-the-art power measurement equipment cannot supply power to the M.2 slot in a laptop, it requires a regular PCIe x4 slot.

Application Tests

With such a complex multi-tier storage system, like Intel Optane Memory H10, the most accurate benchmarks will be tests that use real-world applications. SYSmark 2018 from BAPCo and PCMark 10 from UL are two competing sets of automated application tests. Both have a common goal – to provide an estimate of the overall system performance, as well as several additional assessments covering various common use cases. PCMark 10 is a shorter test, and it provides a more detailed breakdown of the various parameters. It also loads the GPU much harder, since 3D rendering is included in the standard test suite, and some 3DMark tests are included in the advanced test. The advantage of SYSmark 2018 is the use of full commercial versions of popular applications, including Microsoft Office and Adobe Creative Suite, and it includes the ability to measure the overall power consumption of the system during the test. The disadvantage of these tests is that they cover only the most common cases of everyday use, and do not imitate heavy multitasking. None of their subtests use storage vigorously, so points often change slightly when comparing fast and slow SSDs.

BAPCo SYSmark 2018

BAPCo's SYSmark 2018 is an application-based benchmark that uses real-world applications to simulate the behavior of business users, with additional indicators of productivity, creativity and responsiveness. The results reflect the overall system performance and are calibrated against a reference system whose performance is equal to 1000 points in each of the scenarios. Evaluating, say, 2000, will mean that the system under test is twice as fast as the reference system.

The desktop Kaby Lake and the Whiskey Lake laptop change places depending on the subtest; sometimes the laptop is ahead due to its additional RAM, and sometimes the desktop is ahead thanks to a higher TDP. These differences usually have a greater impact than the choice of storage, although the test for responsiveness shows that the parameters of the hard disk alone are not enough. Evaluating Optane Memory H10 with caching enabled is not much better than using only a part of QLC, and with the Optane cache it works at about the level of other, simpler SSD configurations.

Power consumption

SYSmark power consumption estimates show total system consumption, with the exception of the display. Our test system, Kaby Lake, consumes about 26 watts at idle and reaches peak power values ​​in excess of 60 watts, according to our results. SATA SSDs rarely consume more than 5 watts and are idle with a fraction of a watt, moreover, SSDs conduct most of the tests in idle mode. This means that energy use indicators will be very similar. The laptop consumes significantly less energy, despite the fact that the display consumption was also taken into account. None of the energy-intensive storage options (hard drives, Optane 900P) can fit into this system, so power consumption is also quite close to each other.

Optane Memory H10 was the most voracious version of M.2. Turning off the Optane cache saves some energy, but not enough to catch up with good TLC-based drives. Optane SSD 800P has higher energy efficiency than most flash drives, but its low capacity is an obstacle to real use.

UL PCMark 10

Charts


Extended


Standard


Essentials


Apps Start-Up


Video conferencing


Web browsing


Performance


Spreadsheets


Writing


Digital Content Creation


Photo editing


Rendering visualization


Video editing


Gaming


Graphics


Physics


Combined

Кэш-память Optane обеспечивает достаточный прирост показателей в тесте PCMark 10 Extended, чтобы вывести H10 в лидеры среди твердотельных накопителей M.2, протестированных на ноутбуке Whiskey Lake. Субтесты Essentials показывают наибольшее влияние хранилища Optane, в то время как задачи, требующие больших вычислительных ресурсов, получили мало пользы. H10 выглядит примерно так же с включенным кэшированием или без него.

Заполнение хранилища

Этот тест начинается с недавно очищенного диска, и заполняет его последовательной записью 128 КB на глубине очереди 32, фиксируя скорость записи каждого сегмента 1 GB. Этот тест не является представителем какого-либо обычного использования потребителем, но он позволяет нам наблюдать за изменениями в поведении диска по мере его заполнения. Это может позволить нам оценить размер кэша записи SLC и получить представление о том, сколько производительности остается в тех редких случаях, когда система продолжает записывать данные после заполнения кэша.

Charts


Intel Optane Memory H10 512GB


Intel SSD 660p 1TB


Intel Optane SSD 900P 280GB


Samsung 970 EVO 500GB


Intel Optane Memory H10 512GB (32GB Optane)


Intel Optane Memory M10 64GB


Team MP34 512GB


Crucial MX500 500GB


Intel Optane Memory 32GB


MyDigitalSSD SBX 512GB


Western Digital WD Black 7200RPM 1TB


Intel Optane SSD 800P 118GB


WD Black 1TB 7200RPM + Optane Memory 32GB

Во время непрерывной записи кэш Optane на Intel Optane Memory H10 не особо меняет ситуацию с поведением только QLC части (Intel 660p) — сам кэш Optane способен разве что на 350 МB/s. Кэш записи SLC на стороне NAND является более важным фактором, который помогает поддерживать высокую скорость записи после переполнения кэша Optane, равного 32 GB. Но в конце концов все кэши заполняются, и очень медленная скорость записи чистого QLC предстает во всей красе.

Общая средняя скорость записи при полной записи Optane Memory H10 очевидно ниже, чем у любого другого накопителя в этом наборе. Intel 660p на 1 ТB и так немного медленнее, чем жесткий диск на 7200 об / мин, а наш образец H10 имеет вдвое меньше QLC для работы.

Charts


Intel Optane Memory H10 512GB


Intel SSD 660p 1TB


Intel Optane SSD 900P 280GB


Samsung 970 EVO 500GB


Intel Optane Memory H10 512GB (32GB Optane)


Intel Optane Memory M10 64GB


Team MP34 512GB


Crucial MX500 500GB


Intel Optane Memory 32GB


MyDigitalSSD SBX 512GB


Western Digital WD Black 7200RPM 1TB


Intel Optane SSD 800P 118GB


WD Black 1TB 7200RPM + Optane Memory 32GB

Размер кэш-памяти Optane на H10 составляет 32 GB, но при тестировании случайных операций чтения он подходит только для рабочих наборов объемом 6–8 GB, после чего кэш начинает очистку, а производительность падает примерно до уровня обычного диска QLC. Похоже, что Intel могла зарезервировать большую часть кэша Optane для использования в качестве буфера записи, и это может нанести ущерб интенсивным нагрузкам чтения.

Спасибо, что остаетесь с нами. Вам нравятся наши статьи? Хотите видеть больше интересных материалов? Поддержите нас оформив заказ или порекомендовав знакомым, 30% скидка для пользователей Хабра на уникальный аналог entry-level серверов, который был придуман нами для Вас: Вся правда о VPS (KVM) E5-2650 v4 (6 Cores) 10GB DDR4 240GB SSD 1Gbps от $20 или как правильно делить сервер? (доступны варианты с RAID1 и RAID10, до 24 ядер и до 40GB DDR4).

VPS (KVM) E5-2650 v4 (6 Cores) 10GB DDR4 240GB SSD 1Gbps до лета бесплатно при оплате на срок от полугода, заказать можно тут.

Dell R730xd в 2 раза дешевле? Только у нас 2 х Intel TetraDeca-Core Xeon 2x E5-2697v3 2.6GHz 14C 64GB DDR4 4x960GB SSD 1Gbps 100 ТВ от $199 в Нидерландах! Dell R420 — 2x E5-2430 2.2Ghz 6C 128GB DDR3 2x960GB SSD 1Gbps 100TB — от $99! Читайте о том Как построить инфраструктуру корп. класса c применением серверов Dell R730xd Е5-2650 v4 стоимостью 9000 евро за копейки?

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *