Intel Skylake DDR3 vs. DDR4 Comparison

[MAC]

When leaks about Intel Skylake processors first started appearing about 9 months ago, those of us in the tech community were extremely surprised to hear that this new generation would feature an integrated memory controller (IMC) that would support not only DDR4 but the older DDR3 standard as well. There had never before been an Intel processor with a similar dual mode IMC. Once actually launched, Intel's official datasheets revealed that Skylake only supported 1.35V DDR3, also known as DDR3L, which makes sense since that was well within the voltage range of enthusiast DDR4 memory kits and thus would not put undue strain on the IMC.

Source: 6th Generation Intel® Core™ Processor Family Datasheet, Vol. 1​

Nevertheless, motherboard manufacturers started announcing Z170 motherboards with standard 1.5V DDR3 memory slots, and they even validated those models with enthusiast memory kits with voltage requirements up to 1.65V. We actually reviewed one of those yesterday, GIGABYTE's Z170-HD3. Because of this development, Intel took the unusual step of explicitly telling certain websites that 1.5V and above DDR3 memory was not supported as they could not guarantee the long-term health of Skylake's IMC at those higher memory voltages.

Despite these clear warnings, we still thought that using DDR3 on Skylake would be pretty cool, since it meant that those who already had suitable DDR3 memory kits could upgrade to the Skylake platform without having to expend extra cash on pricey DDR4. Since then the price of DDR4 has crashed to near parity with DDR3, but those on tight budget understandably might still want to carry over their current DDR3 memory kits.

So there are certain risks in reusing your older DDR3 memory kit, but is there a performance penalty? That's what we are here to find out. When the Haswell-E LGA2011-v3 platform came out with native DDR4 support many people tried to compare DDR3 and DDR4 memory performance in really ham-fisted ways. However, the only way to perform an identical comparison is to have a platform that supports both DDR3 and DDR4 while keeping the same CPU, and we now have that opportunity with Skylake.

Will the cliché that DDR4's loose memory timings hurts its performance hold true? We don't think so, and we explained why in our very first DDR4 memory kit review:

"While these timings might seem incredibly loose when compared to what we are all familiar with on the DDR3 side, it is important to realize that secondary and tertiary memory timings have arguably become an increasingly more important part of memory performance due to improvements in the memory controller. When you combine this truth with the fact that DDR4 memory kits are able to run at more aggressive secondary and tertiary timings, the overall performance disadvantage (if any) is quite minimal."

As a result, what we are actually more concerned about is whether DDR3's lower clock speeds and resulting lack of bandwidth will hold back the performance potential of our flagship quad-core Core i7-6700K. Reusing parts and saving money is great, but not if it results in a bottleneck that chokes the life out of a cutting edge new processor.

 

[MAC]

DDR3 vs. DDR4 – More Than Just Frequency & Latency

DDR3 vs. DDR4 – More Than Just Frequency & Latency

Some people have shied away from DDR4 because of the loose timings, but that's somewhat shortsighted. As a general rule, memory timings increase as clock speed increases, so of course DDR4 has looser timings since it's clocked much higher. Having said that, true latency is a function of clock speed, which means that memory latency actually decreases as memory speeds increase. As long as you're increasing the speed more than you're loosening the timings, you’ll achieve lower latency, higher bandwidth, and just generally better memory performance. And that is what DDR4 was designed to do. Both Crucial and Corsair have really interesting articles dedicated to this topic. While these are both inherently biased sources - since they really want you to go out and buy DDR4 - both companies are also experts in their field.

Now you might be saying that DDR3 can match certain DDR4 frequencies and do so at tighter timings. That is absolutely true, but it can't do it at 1.2V or 1.35V. When you consider the 20% voltage reduction and all the other energy-saving features baked into DDR4, the overall power savings are upwards of 30%. If a reduction in power consumption doesn’t excite you, then maybe vastly increased module density will.

While DDR3 memory chips can contain up to 8 internal banks, DDR4 increases that to 16 internal banks. While more internal banks help with performance - by allowing the advance prefetch functions to reduce access latency via increased parallelization – it opens the door to higher density memory modules. Thus, the DDR4 standard supports high-density 128GB DDR4 memory modules, an example of which Samsung just unveiled a few days ago.

The point of this section is not to sell you on DDR4, it's just a refresher on why the new standard was needed to begin with. Having said that, once you see some of our future reviews, revealing DDR4's true capabilities and potential, we suspect that no one will need any further convincing. You can see a good example of that in the sneak peak posted above.

When it comes down to it, we can list all the reasons why DDR4 is a step in the right direction, but all we actually care about is performance. Does DDR4's inherent superiority actually make a difference on Skylake? Are these quad-core processors even able to make use of all that additional bandwidth? That's what we are here to find out.

 

[MAC]

Test Setups & Methodology

Test Setups & Methodology

For this review, we are going to be testing the performance of DDR3 and DDR4 on the Skylake-S LGA1151 platform using a variety of memory frequencies and timings. Excluding the motherboards and memory kits, the components and software are the same across both systems, and aside from manually selecting the frequencies, timings, and voltages, every option in the BIOS was at its default setting. The ASUS motherboard set a default Uncore/Cache frequency of 4.1Ghz instead of the default 4.0Ghz, so we lowered that in order to match Gigabyte's default 4.0Ghz setting.

Intel Core i7 LGA1151 DDR3 Test Setup​

Intel Core i7 LGA1151 DDR4 Test Setup​

For all of the benchmarks, appropriate lengths are taken to ensure an equal comparison through methodical setup, installation, and testing. The following outlines our testing methodology:

A) Windows is installed using a full format.

B) Chipset drivers and accessory hardware drivers (audio, network, GPU) are installed.

C)To ensure consistent results, a few tweaks are applied to Windows 7 and the NVIDIA control panel:

• UAC – Disabled
• Indexing – Disabled
• Superfetch – Disabled
• System Protection/Restore – Disabled
• Problem & Error Reporting – Disabled
• Remote Desktop/Assistance - Disabled
• Windows Security Center Alerts – Disabled
• Windows Defender – Disabled
• Screensaver – Disabled
• Power Plan – High Performance
• V-Sync – Off

D) All available Windows updates are then installed.

E) All programs are installed and then updated, followed by a defragment.

F) Benchmarks are each run three to eight times, and unless otherwise stated, the results are then averaged.

Here is a full list of the applications that we utilized in our benchmarking suite:

• 3DMark Vantage Professional Edition v1.1.3
• 3DMark11 Professional Edition v1.0.132.0
• 3DMark 2013 Professional Edition v1.5.915
• AIDA64 Extreme Edition v5.50.3600
• Cinebench R15 64-bit
• FAHBench 1.2.0
• Final Fantasy XIV: Heavensward Benchmark
• HEVC Decode Benchmark (Cobra) v1.61
• LuxMark v3.0
• PCMark 8
• SuperPi Mod v1.9 WP
• Sisoft Sandra 2015.SP3 20.28
• Valve Particle Simulation Benchmark v1.0.0.0
• WinRAR x64 5.30 beta 6
• wPRIME version v2.10
• X3: Terran Conflict Demo v1.0

That is about all you need to know methodology wise, so let's get to the good stuff!

 

[MAC]

System Benchmarks

System Benchmarks

SuperPi Mod v1.9 WP

When running the SuperPI 32MB benchmark, we are calculating Pi to 32 million digits and timing the process. Obviously more CPU power helps in this intense calculation, but the memory sub-system also plays an important role, as does the operating system. We are running one instance of SuperPi Mod v1.9 WP. This is therefore a single-thread workload.

wPRIME 2.10

wPrime is a leading multithreaded benchmark for x86 processors that tests your processor performance by calculating square roots with a recursive call of Newton's method for estimating functions, with f(x)=x2-k, where k is the number we're sqrting, until Sgn(f(x)/f'(x)) does not equal that of the previous iteration, starting with an estimation of k/2. It then uses an iterative calling of the estimation method a set amount of times to increase the accuracy of the results. It then confirms that n(k)2=k to ensure the calculation was correct. It repeats this for all numbers from 1 to the requested maximum. This is a highly multi-threaded workload.

Cinebench R15

Cinebench R15 64-bit
Test1: CPU Image Render
Comparison: Generated Score

The latest benchmark from MAXON, Cinebench R15 makes use of all your system's processing power to render a photorealistic 3D scene using various different algorithms to stress all available processor cores. The test scene contains approximately 2,000 objects containing more than 300,000 total polygons and uses sharp and blurred reflections, area lights and shadows, procedural shaders, antialiasing, and much more. This particular benchmarking can measure systems with up to 64 processor threads. The result is given in points (pts). The higher the number, the faster your processor.

WinRAR x64

WinRAR x64 5.30 beta 6
Test: Built-in benchmark, processing 1000MB of data.
Comparison: Time to Finish

One of the most popular file archival and compression utilities, WinRAR's built-in benchmark is a great way of measuring a processor's compression and decompression performance. Since it is a memory bandwidth intensive workload it is also useful in evaluating the efficiency of a system's memory subsystem.

FAHBench

FAHBench 1.2.0
Test: OpenCL on CPU
Comparison: Generated Score

FAHBench is the official Folding@home benchmark that measures the compute performance of CPUs and GPUs. It can test both OpenCL and CUDA code, using either single or double precision, and implicit or explicit modeling. The single precision implicit model most closely relates to current folding performance.

HEVC Decode Benchmark v1.61

HEVC Decode Benchmark (Cobra) v1.61
Test: Frame rates at various resolution, focusing on the top quality 25Mbps bitrate results.
Comparison: FPS (Frames per Second)

The HEVC Decode Benchmark measures a system's HEVC video decoding performance at various bitrates and resolutions. HEVC, also known as H.265, is the successor to the H.264/MPEG-4 AVC (Advanced Video Coding) standard and it is very computationally intensive if not hardware accelerated. This decode test is done entirely on the CPU.

LuxMark v3.0

Test: OpenCL CPU Mode benchmark of the LuxBall HDR scene.
Comparison: Generated Score

LuxMark is a OpenCL benchmarking tool that utilizes the LuxRender 3D rendering engine. Since it OpenCL based, this benchmark can be used to test OpenCL rendering performance on both CPUs and GPUs, and it can put a significant load on the system due to its highly parallelized code.

PCMark 8

PCMark 8 is the latest iteration of Futuremark’s system benchmark franchise. It generates an overall score based upon system performance with all components being stressed in one way or another. The result is posted as a generalized score. In this case, we tested with both the standard Conventional benchmark and the Accelerated benchmark, which automatically chooses the optimal device on which to perform OpenCL acceleration.

AIDA64 Memory Benchmark

AIDA64 Extreme Edition is a diagnostic and benchmarking software suite for home users that provides a wide range of features to assist in overclocking, hardware error diagnosis, stress testing, and sensor monitoring. It has unique capabilities to assess the performance of the processor, system memory, and disk drives.

The benchmarks used in this review are the memory bandwidth and latency benchmarks. Memory bandwidth benchmarks (Memory Read, Memory Write, Memory Copy) measure the maximum achievable memory data transfer bandwidth. The code behind these benchmark methods are written in Assembly and they are extremely optimized for every popular AMD, Intel and VIA processor core variants by utilizing the appropriate x86/x64, x87, MMX, MMX+, 3DNow!, SSE, SSE2, SSE4.1, AVX, and AVX2 instruction set extension.
The Memory Latency benchmark measures the typical delay when the CPU reads data from system memory. Memory latency time means the penalty measured from the issuing of the read command until the data arrives to the integer registers of the CPU.

The tests above shouldn't come as any surprise since most programs and benchmarks don't capitalize upon the massive bandwidth increase offered by DDR4. Instead, we see near-parity across nearly all tests but there do seem to be some differentiating factors when the system runs up against an application that requires additional calls to the memory subsystem. There aren't many of those instances around, and even fewer outside the professional market so in most home-use cases users won't see one bit of difference.

 

[MAC]

Gaming Benchmarks

Gaming Benchmarks

Futuremark 3DMark (2013)

3DMark v1.1.0
Graphic Settings: Fire Strike Preset
Rendered Resolution: 1920x1080
Test: Specific Physics Score and Full Run 3DMarks
Comparison: Generated Score


3DMark is the brand new cross-platform benchmark from the gurus over at Futuremark. Designed to test a full range of hardware from smartphones to high-end PCs, it includes three tests for DirectX 9, DirectX 10 and DirectX 11 hardware, and allows users to compare 3DMark scores with other Windows, Android and iOS devices. Most important to us is the new Fire Strike preset, a DirectX 11 showcase that tests tessellation, compute shaders and multi-threading. Like every new 3DMark version, this test is extremely GPU-bound, but it does contain a heavy physics test that can show off the potential of modern multi-core processors.

Futuremark 3DMark 11

3DMark 11 v1.0.5
Graphic Settings: Performance Preset
Resolution: 1280x720
Test: Specific Physics Score and Full Run 3DMarks
Comparison: Generated Score


3DMark 11 is Futuremark's very latest benchmark, designed to tests all of the new features in DirectX 11 including tessellation, compute shaders and multi-threading. At the moment, it is lot more GPU-bound than past versions are now, but it does contain a terrific physics test which really taxes modern multi-core processors.

Futuremark 3DMark Vantage

3DMark Vantage v1.1.2
Graphic Settings: Performance Preset
Resolution: 1280x1024

Test: Specific CPU Score and Full Run 3DMarks
Comparison: Generated Score

3DMark Vantage is the follow-up to the highly successful 3DMark06. It uses DirectX 10 exclusively so if you are running Windows XP, you can forget about this benchmark. Along with being a very capable graphics card testing application, it also has very heavily multi-threaded CPU tests, such Physics Simulation and Artificial Intelligence (AI), which makes it a good all-around gaming benchmark.

Valve Particle Simulation Benchmark

Valve Particle Simulation Benchmark
Resolution: 1680x1050
Anti-Aliasing: 4X
Anisotropic Filtering: 8X
Graphic Settings: High

Comparison: Particle Performance Metric

Originally intended to demonstrate new processing effects added to Half Life 2: Episode 2 and future projects, the particle benchmark condenses what can be found throughout HL2:EP2 and combines it all into one small but deadly package. This test does not symbolize the performance scale for just Episode Two exclusively, but also for many other games and applications that utilize multi-core processing and particle effects. As you will see the benchmark does not score in FPS but rather in its own "Particle Performance Metric", which is useful for direct CPU comparisons.

X3: Terran Conflict

X3: Terran Conflict 1.2.0.0
Resolution: 1680x1050
Texture & Shader Quality: High
Antialiasing 4X
Anisotropic Mode: 8X
Glow Enabled

Game Benchmark
Comparison: FPS (Frames per Second)

X3: Terran Conflict (X3TC) is the culmination of the X-series of space trading and combat simulator computer games from German developer Egosoft. With its vast space worlds, intricately detailed ships, and excellent multi-threaded game engine, it remains a great test of modern CPU performance.

Final Fantasy XIV: Heavensward Benchmark

Final Fantasy XIV: Heavensward
Resolution: 1680x1050
Texture & Shader Quality: Maximum IQ
DirectX 11
Fullscreen

Game Benchmark
Comparison: Generated Score

Square Enix released this benchmarking tool to rate how your system will perform in Heavensward, the expansion to Final Fantasy XIV: A Realm Reborn. This official benchmark software uses actual maps and playable characters to benchmark gaming performance and assign a score to your PC.

Even at extremely low resolutions, games and synthetic gaming benchmarks realize even less performance increase with DDR4 than the applications we tested on the previous page. The reason for this is pretty simple: in most cases games experience a GPU bottleneck long before the system memory gets factored into the equation. Granted, there can be a 10% difference between the two extreme ends of the spectrum but once again, that's only at lower detail levels in a very old game.

We actually loaded these tests with older game examples and lower resolutions in an effort to up-play any differences there may be between the two memory types. The fact that it didn't happen is proof positive that gamers won't see much, if any visible performance increase with DDR4 installed.

 

[SKYMTL]

Conclusion

Conclusion

Although the DDR4 standard was established all the way back in 2012, nobody really cared about it until its appearance on the Haswell-E LGA2011-v3 platform. Even then, since that high-end platform is such a niche product anyways, it is really the arrival of Skylake-S that has forced your average consumer to start paying attention to this new memory standard.

As always happens with new memory standards, DDR4 started out with a gross price premium over DDR3. However, a quick price check reveals that DDR4 memory kits have essentially come down to parity with their similarly-sized DDR3 counterparts. The reason for this is that the industry transition from DDR3 to DDR4 is heavily underway, largely motivated by the enterprise and mobile markets, and thankfully the DIY desktop market gets to benefit from the lower prices caused by larger production volumes. As a result, with the price obstacle out of the way and the DDR4 performance deficit argument largely settled, has DDR3's time finally come to an end?

When it came to pure memory performance, the memory bandwidth difference between our DDR3-2400 and DDR4-2400 systems was only about 2.5% when it came to read speeds, 7% for write speeds, and 4% for copy speeds, all in favor of DDR4 . With the timings that we set, which we felt accurately mirrored the average timings of currently available memory kits, the DDR3 system did achieve 2.5% lower latency figures. These aren't dramatic differences, but that's because we are comparing two unevenly matched configurations. Whereas DDR3-2400 is near the top-end of the spectrum for DDR3, DDR4-2400 is considered relatively lowly in the DDR4 realm. If we compare the DDR3-2400 to DDR4-3200, the memory bandwidth gains are 32% read, 41% write, and 32% copy. That is pretty significant, but does it translate to real-life performance gains?

Well, when we compare the system and gaming benchmark results between our DDR3-2400 and DDR4-2400 configurations, there is very little difference. The DDR3 system occasionally took the lead, but overall DDR4 consistently inched ahead. It might sound obvious, but to see the performance benefits of the additional memory bandwidth provided by DDR4 you need a memory bandwidth intensive workload. There aren't a ton of those on the consumer side, but WinRAR file extraction is one of them and the performance difference was significant. There was a 7.5% gap between the identically clocked DDRx-2400 configurations, and a 20% improvement between the fastest DDR3-2400 and DDR4-3200 results. Sure, we could have tried DDR3-2666 or even DDR3-2800, but the economics just don't make sense since those elite DDR3 kits are often twice as expensive as comparable DDR4 kits. Furthermore, no matter the results, we would never recommend that anyone actually go out and buy expensive DDR3 at this point anyways, since DDR4 is the future in terms of compatibility with upcoming platforms.

In the introduction, we wondered whether DDR3's low clock speeds and resulting lack of bandwidth would hold back the performance of a Core i7-6700K. Based on our results, we have to say both yes and no. There are definitely instances when run-of-the-mill DDR3-1600 or even DDR3-2133 will be a bottleneck, but the numbers just aren't significant enough for us to sincerely recommend that you need to buy a DDR4 kit if you're trying to upgrade on a tight budget. If you already have a highly clocked DDR3 memory kit, definitely consider using it in a Skylake build. You will save money and achieve comparable performance to DDR4, but you need to be comfortable with the possibility of potential IMC damage and accept that no matter what there is an eventual DDR4 RAM purchase in your future.