Native Z170 Hyper M.2 vs PCIe3 M.2

November 12, 2015

Article: Intel Z170 native M.2 vs PCIe3

With consumer grade NVMe SSDs finally arriving, it becomes
important that the performance of NVMe SSDs can be fully realised by the way they
are connected to the motherboard.

In this article I will examine the performance difference
(if any) of an M.2 NVMe SSD connected using the Intel Z170 ‘native’ hyper M.2
solution via the Z170 Skylake PCH (Platform Controller Hub),
and the same M.2 NVMe SSD connected using the main PCIe3 x16 sockets via the Intel
Skylake CPU.

NVMe (Non Volatile Memory Express)
SSDs are PCIe based and can be installed in a standard PCIe slot, M.2 socket, or
via the brand new U.2 connector. PCIe SSDs are not new, and have been around
for several years. However, the PCIe SSDs of the past required a special
controller which sat between the SSD hardware and the PCIe system bus, to allow
the SSD hardware and the PCIe bus to perform the translation and communication
between the two interfaces. This was of course a very complex and time
consuming task, which inevitably led to increased latency.

NVMe is a native solution, with its own highly optimised
protocol, which features a very much reduced command set, much lower latency
when compared to AHCI, and is specifically optimised for Non Volatile Memory
(FLASH memory).

M.2 is a small form factor, initially designed for small
form factor PCs, such as laptops, but have found their way into desktop PCs.
Initially, M.2 appeared ‘natively’ in desktop PC’s with the Intel Z97 chipset.
The Z97 native solution was very limited. It could use only two generation 2
PCIe lanes, and was thus limited to maximum bandwidth of 10Gbps, or 1GB per
second. In practice, and due to the overheads, this translated to a maximum
throughput of around 740 MB/s. Latency was also quite high with the Z97 native
solution and therefore, to get maximum performance from M.2 in a Z97 setup, M.2
SSDs were often mounted in a PCIe3 PCIe to M.2 adapter card, thereby utilising
the higher bandwidth and lower latency of main PCIe3 sockets on the motherboard.

However, mounting an M.2 SSD in a PCIe3 card, and then
connecting this to the main PCIe3 sockets on the motherboard does have its
limitations. If you’re using a main PCIe3 socket then you’re limited to a
single graphics card, and that single graphics card will be limited to using
eight PCIe3 lanes. This shouldn’t really be a problem as current graphics cards
should not be limited by the reduced bandwidth, but it does rule out using an effective
SLI or crossfire graphics solution.

The new Intel Skylake CPU requires an Intel 100 series
chipset to function, with the high performance desktop variant being designated
Z170. For a device to be ‘natively’ supported, it must either be a feature
built into the CPU, such as the DRAM controller, Intel HD graphics, DMI, and
16x PCIe3, or alternatively, a feature of the chipset itself, PCH (platform
controller hub).

Most Z170 motherboard manufacturers are connecting M.2 via
the platform controller hub, but a few are connecting M.2 via the main 16x
PCIe3 sockets. You should watch out for this because, as I explained above,
connecting via the main PCIe3 sockets will be stealing bandwidth from those
sockets. On the positive side, M.2 connected via the main PCIe3 sockets is
likely to enjoy lower latency, and therefore better performance.

My Asus Z170 Deluxe motherboard uses the native Z170 PCH
solution, and therefore makes the full bandwidth available to the two main
PCIe3 sockets for graphics cards, or other bandwidth hungry PCIe3 solutions. In
the picture below, we can see the layout of the Asus Z170 Deluxe.

Asus Z170 Deluxe

At the bottom right of the picture you see the native M.2
socket. This solution uses 4x PCIe3 lanes, with a maximum bandwidth of 32Gbps,
or 3.2 GB/s. You can also see that the Asus Z170 Deluxe supports Hyper M.2 X4
with the built in M.2 socket, via an PCIe3 x4 add in card. Worth noting as well
is that the Asus Z170 Deluxe supports the new U.2 standard, for 2.5 inch NVMe
SSD solutions.

In the block diagram below, you can see how native hyper M.2
is connected on the Asus Z170 Deluxe motherboard.

We can see clearly that the hyper M.2 socket is connected
via the Intel Skylake Z170 PCH (platform controller hub). Finally, in the
screenshot below, you can see both the U.2 and M.2 PCIe3 adapter cards supplied
with the Asus Z170 Deluxe. The hyper M.2/U.2 card allows you to mount an M.2 or
U.2 SSD in the main PCIe3 x16 sockets, should you prefer to take that route
rather than using the built in hyper M.2 socket.

Before I go any further, I should explain some important
things that have changed between the Intel Z97 and the Intel Z170 chipsets.

The platform controller hub on the Intel Z97 chipset is
connected to the CPU via an interface known as DMI 2, which allows maximum
bandwidth between the PCH and the CPU of 20 Gbps, or 2 GB/s without overheads.
On the Intel Skylake Z170 platform, DMI has been upgraded to DMI 3, which
allows 8GT per second, or 8 Giga transfers per second, with a maximum bandwidth
of 40 Gbps, (4GB/s).

Intel Z97 connected to all connected devices using PCIe
generation 2 lanes; with each lane allowing a maximum bandwidth of 500 MB/s.
Z97 had eight PCIe2 lanes available for connecting all the devices.

Intel Z170 PCH connects to all connected devices using PCIe3
lanes, with each lane providing a maximum bandwidth of 1GB/s, and Z170 has 20
PCIe3 lanes at its disposal. So the Intel Z170 platform controller hub is a
huge upgrade from a performance perspective.

So, let’s compare ‘native’ hyper M.2 via the Z170 platform
controller hub, and hyper M.2 connected via the main PCIe3 x16 sockets, to find
out which one performs the best.

Let’s head to the next page.........

Test machine

For this review I will be using a computer with the
following configuration:

Hardware:

Motherboard: Asus Z170 Deluxe (Intel Z170 chipset)
Processor: Intel 6th generation Core i7 6700K
CPU cooler: BeQuiet Dark Rock Pro 2
RAM: 16GB Corsair Vengeance LPX 2666MHz DDR4 (dual channel)
GFX: MSI GTX 950 Gaming 2G
Sound: Onboard Realtek ALC1050 HD audio controller
Hard disk OS: OCZ Vector 256GB SSD.
Case: Antec Performance One P280
PSU: Antec True Power modular 550W
Display: Dell UltraSharp U2412M 24” widescreen IPS LCD (HDCP
compliant)
Operating System: Windows 10 Professional 64bit

For these tests I will be using a Samsung
950 Pro M.2 NVMe SSD, which I reviewed just a few weeks ago. The Samsung
950 Pro M.2 NVMe SSD was tested in the ‘native’ hyper M.2 socket fitted to the
Asus Z170 Deluxe, then the same tests were carried out with the Samsung 950 Pro
M.2 NVMe SSD mounted in the Asus hyper M.2 to PCIe3 adapter card, and then
connected to the main PCIe3 x16 socket, connected via the Skylake CPU.

The NVMe drivers used to test the 950 Pro M.2 NVMe SSD were
Samsung’s own NVMe driver version 1.4.7.15

CPU power saving states was disabled for consistency, and
all the SSDs in this article were tested with all CPU power saving states
disabled.

Test applications

To test the performance of the Samsung 950 Pro M.2 NVMe SSD,
I will be using the following test applications in this review.

HD-Tune Pro
ATTO
Iometer
AS SSD Benchmark
Anvil’s
Storage Utilities
PC
Mark 8

Test procedures

Firstly I will check the performance with a few synthetic
benchmarks, and on the next page I will supplement this with a real world test
using PCMark 8 storage suite.

The exact same tests were carried out in the two different
methods of connecting the Samsung 950 Pro M.2 NVMe SSD.

Drive preparation for running the tests

To make the tests fair, the SSD was secure erased to bring
performance back to its factory default state before the start of each test.

Where I use graphs in this article to display results, I
will use the following colours to make it easier for our readers to see which drive
we are reviewing.

Samsung 950 Pro –
Native Z170 hyper M.2 connected.

Samsung 950 Pro –
PCIe3 hyper M.2 connected.

SATA SSD

Synthetic Benchmarks

HD Tune Pro

In this benchmark I am checking sequential reading speed.

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

Let's see how the two connection methods compare in the bar
graph below.

The result is almost identical, with PCIe3 having a slight
advantage.

ATTO disk benchmark

ATTO has become a standard tool for measuring the data
throughput of HDDs and SSDs. It measures the reading and writing performance,
using different file sizes and block sizes.

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

Let's find out how they compare.

ATTO Reading performance

ATTO - Reading
performance at various block sizes

The Samsung 950 Pro M.2 NVMe SSD connected via native Z170
hyper M.2 has a very slight advantage when reading data, but once again they
are very close.

ATTO Writing performance

ATTO - Writing
performance at various block sizes

This time the PCie3 connecting method has a slight
advantage, but yet again they are very close.

AS SSD Benchmark

AS SSD benchmark is a benchmarking tool specifically
designed to test SSDs. The application tests sequential reading and writing
performance, 4K random reading and writing performance.

AS SSD benchmark also tests 4K threaded performance. This is
very exciting, as this test is the first available test that I am aware of,
that simulates how a PC operating system actually works. A modern PC and OS,
such as Windows 10 does not just run a single thread at a time, it runs many
threads. The AS SSD benchmark "4K 64Thrd" tests run 64 threads
simultaneously throughout the test. If this result is good, then you can be
pretty sure the drive will perform extremely well as a system drive.

After the tests complete, AS SSD benchmark derives a total
score for the drive being tested. This is based on all aspects of the test
results, and gives an indication of how the drive is performing overall.

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

This time, PCIe3 connected is faster than native Z170 hyper
M.2 connected, but once again the advantage is slight.

Anvil’s Storage Utilities

As well as performing SSD endurance tests. Anvil’s Storage
Utilities has a very nice SSD benchmarking application. The SSD benchmark tests
many different aspects of SSD performance, including 4K random at different
queue depths, and also sequential performance, but more importantly than this,
all using real test data.

Another very nice feature of Anvil’s SSD benchmark is the
fact that you can change the compression levels of the test data. The
compression levels of the datasets used for the tests can be varied from 0%
compression right up to 100% compressed data, and there are even a few data
profiles already included, such as database (8%) compression, and also an application
profile (46%) compression, which is designed to simulate real application data
being read and written to the SSD. However, for this test I will only use the 0
fill option.

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

Let’s compare the results in the graph below.

Once again, the PCIe3 connecting method has an advantage,
but the Samsung 950 Pro connected via native Z170 hyper M.2 is not far behind.

Testing I/O Performance with IOMeter

IOMeter is probably the most versatile of all the synthetic
benchmarks. Its ability to be configured to generate a multitude of different
I/O traffic is unmatched. Another great feature of IOMeter, is the capability
to test any storage metric that you can think of, providing you know how to
configure the assignments. The reviewer also has complete control over things
like queue depth, block size, whether the traffic is random, sequential, or
even a mixture of both.

Partition alignment and sector boundaries

Windows 10, Windows 8.1, Windows 7, and Windows Vista will
automatically align a partition to 4k boundaries during partition creation,
Windows XP won’t. It is imperative that an SSD’s partition is aligned. Windows
XP is also restricted to sector boundaries, while Windows 7 and 8 will use 4k
boundaries if they can. The Samsung 950 Pro M.2 NVMe SSD is 4k boundary aware,
and will use these boundaries if possible. Of course it will also remap LBAs
for compatibility with the sector boundaries so that the drive can be used with
Windows XP.

IOMeter 4K random write test with repeating data.

The first test involves creating continual 4KB random files
on the target drive with IOMeter. I use a 4KB file size, as it is believed that
Windows will create and modify many of this size of file constantly in the
background during a typical Windows session. It is said that most 4K random
writes take place at a queue depth of only one, and I have been requested to
include this test in my reviews.

Queue depth 1

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

MB/s

Yet again, the PCIe3 connecting method is the faster of the
two.

Latency

This graph shows the difference in latency between the two
connecting methods, and also includes an SATA SSD for comparison.

The Samsung 950 Pro performs slightly better when connected
via PCIe3 in this test, so it’s no surprise to see that latency is also lower.
By including an SATA SSD in this test (the Samsung 850 Pro)it is interesting to
see how much lower the latency is with NVMe SSDs.

Queue depth 4

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

This time the native Z170 hyper M.2 connection proves to the
faster of the two, but once again they are close.

Queue depth 32

Samsung 950 Pro –
Native Z170 hyper M.2 connected

Samsung 950 Pro –
PCIe3 connected

Once again, the native Z170 hyper M.2 connection method
proves to be faster.