A straightforward interpretation of the name “Joint Electron Device Engineering Council” tells us a lot about why memory standards improve so slowly. As a multi-member council of memory manufacturing and electronics engineering companies, the standards it sets are defined by ease of adoption. True innovation by one of its members is likely to remain non-standard for months or years as everyone else plays catch-up. Even after a newer standard like DDR3-1600 CAS 9 is finally adopted, most manufacturers revert to older settings, such as DDR3-1600 CAS 11, in order to maintain the highest level of compatibility with previous-generation components.
Of course, this doesn't seem to stifle the enthusiast market. Intel's XMP feature followed Nvidia’s EPP technology to add overclocking profiles on top of the JEDEC-specified table of SPD values. Unfortunately, XMP isn't prolific. Notebooks and tier-one desktops typically don't support the feature, so the potential of high-performance RAM is typically ignored by those machines.
While OEM systems (like the Erazer X700 we just reviewed) continue to restrict RAM selection to SPD settings, Intel’s Haswell architecture more forcefully pushes platform vendors to add JEDEC’s latest standards to their firmware. No doubt, that's largely related to the fact that Intel's mobile Haswell-based CPUs don't support 1.5 V memory.

For enthusiasts upgrading today, this means new systems finally support the rather old DDR3-1866 CAS 10 and DDR3-1600 CAS 9 ratings. In talking to the folks at G.Skill, we agreed that even enthusiasts might not be aware of the changes imposed by Haswell. So, the company sent over some of its low-voltage, high-performance SO-DIMM kits to let us quantify the impact for ourselves.
| G.Skill DDR3 16 GB SO-DIMM Specifications | |||
|---|---|---|---|
| Data Rate | Primary Latencies | Rated Voltage | |
| G.Skill Ripjaws F3-1866C10D-16GRSL | DDR3-1866 | 10-10-10-32 | 1.35 V |
| G.Skill Ripjaws F3-1600C9D-16GRSL | DDR3-1600 | 9-9-9-28 | 1.35 V |
| G.Skill Standard Series F3-1333C9D-16GSL | DDR3-1333 | 9-9-9-24 | 1.35 V |
G.Skill gave us a look a three kits of two 8 GB SO-DIMM modules from basic DDR3-1333 to enhanced DDR3-1866. The most common memory we've run across is DDR3-1600 CAS 11, but we didn’t have any 16 GB kits to compare. Previously, however, we've found similar performance between DDR3-1333 CAS 9 and DDR3-1600 CAS 11, so we’ll soldier on with the parts from G.Skill's care package.
We noticed that the company recently announced a DDR3-2133 Ripjaws SO-DIMM kit, but that it tops out at 8 GB (two 4 GB modules). Hopefully it goes to market with a larger 16 GB version in the future.
Small outline DIMMs are usually packed fairly tightly, so G.Skill doesn't bother with heat spreaders. Instead, we find large stickers covering both sides of each module, making it look like there's a spreader covering the ICs.

G.Skill’s dual-channel kit (part number F3-1866C10D-16GRSL) includes two F3-1866C10D-8GRSL modules programmed for DDR3-1866 CAS 10, DDR3-1600 CAS 9, and DDR3-1333 CAS 7. Many mobile CPU-based platforms don't even let you choose between those modes, but the best boards will automatically choose the highest compatible data rate.

Our test platforms had no trouble selecting the memory’s top configuration.

Like the DDR3-1866 kit we just looked at, G.Skill's Ripjaws DDR3-1600 SO-DIMMs use large stickers to look like heat spreaders, without significantly increasing module thickness.

The F3-1600C9D-16GRSL package comprises two F3-1600C9D-8GRSL modules with DDR3-1600 CAS 9, DDR3-1333 CAS 7, and DDR3-1066 CAS 6 values in their SPD table.

Again, our Haswell-based platforms had no trouble choosing the top performance level for these modules, without any manual intervention on our part.

The ability to support high data rates on a CPU with an integrated memory controller requires a modern processor and properly programmed firmware. One of those requisites is often missing from previous-generation machines. G.Skill supports the broader market by producing lower-cost modules for those systems.

Sold under the G.Skill part number F3-1333C9D-16GSL as a dual-channel kit, CPU-Z misreports the tRAS of its F3-1333C9D-8GSL modules at 25 cycles in one of our test systems. Had our platform not been DDR3-1333-capable, DDR3-1066 CAS 7 is also supported.

Regardless of what CPU-Z said about the SPD table, our test platforms correctly set the 9-9-9-24 timings we were looking for at the memory’s highest-rated DDR3-1333 data rate.

| Test System Configuration | |
|---|---|
| Socketed CPU | Intel Core i7-4770K (Haswell): 3.5 to 3.9 GHz, 8 MB Shared L3 Cache, LGA 1150 |
| Embedded CPU | Intel Core i7-4770R (Haswell): 3.25 to 3.9 GHz, 6 MB Shared L3 Cache, BGA 1364 |
| Socketed CPU Cooler | Thermalright MUX-120 w/Zalman ZM-STG1 Paste |
| Embedded CPU Cooler | Gigabyte Brix proprietary |
| Socketed Graphics | Intel HD Graphics 4600: 1250 MHz GPU |
| Embedded Graphics | Intel Iris Pro Graphics 5200: 1300 MHz GPU |
| Socketed Motherboard | ASRock Z87-M8, LGA 1150, BIOS P1.20 (08/22/2013) |
| Embedded Motherboard | Gigabyte M4HM87P-IA, BIOS F1 (10/25/2013) |
| Hard Drive | Samsung 840 Pro MZ-7PD256, 256 GB SSD |
| Sound | Integrated HD Audio |
| Network | Integrated Gigabit Networking |
| Software | |
| Operating System | Microsoft Windows 8 Professional RTM x64 |
| Graphics | Intel Graphics Driver 15.31.17.64.3257 |
| Chipset | Intel INF 9.4.0.1017 |
Curious to find out how much power benefit we’d extract from low-voltage SO-DIMMs, we needed a test platform that supported both 1.35 V and 1.50 V standards. That meant deviating away from a mobile or embedded Haswell-based CPU, since those only support 1.35 V modules. A member of its M8 barebones PC kit, ASRock’s Z87-M8 motherboard supports both voltage levels on two SO-DIMM slots by using a socketed processor.

We’ve heard that SO-DIMMs aren't very overclockable, but G.Skill is confident that its modules have headroom to spare when they're paired with an LGA 1150-based platform. While we weren’t able to follow through with a stable 2133 MT/s data rate, we were at least able to use the higher 1.50 volts to push CAS 9 timings.

Everything I’ve read so far lead me to believe that lower thermal and power limitations would slow the embedded version of Intel’s Core i7-4770 by up to 5%. Conversely, I’ve also been expecting that the company's Iris Pro graphics 5200 would deliver about 20% additional performance. The only problem is that CPU model doesn't let us compare different memory voltage levels. Moreover, I didn't have one at my disposal.

That is, until something "fell out of Gigabyte’s trunk" following BlizzCon. This thing packs the wallop of a huge copper sink to prevent thermal throttling, along with support for both 2.5” and mSATA SSDs. Those two advancements roughly doubled the thickness and weight of this enthusiast-oriented Brix compared to familiar models.
| Benchmark Settings | |
|---|---|
| Games | |
| Battlefield 3 | Campaign Mode, "Going Hunting" 90-Second Fraps Test Settings: Medium Quality Defaults (No AA, 4x AF) |
| F1 2012 | Steam Version, In-Game Test Test Settings: High Quality Preset, No AA |
| Far Cry 3 | V. 1.04, DirectX 11, 50-Second Fraps "Amanaki Outpost" Test Settings: Medium Quality, No AA, Standard ATC., SSAO |
| Metro: Last Light | Steam Version, Built-in Benchmarking Tool DX11, Low Quality, 4x AF, Low Blur, No SSAO/Tesslation/PhysX |
| Adobe Creative Suite | |
| Adobe After Effects CS6 | Version 11.0.0.378 x64: Create Video which includes 3 Streams, 210 Frames, Render Multiple Frames Simultaneosly |
| Adobe Photoshop CS6 | Version 13 x64: Filter 15.7 MB TIF Image: Radial Blur, Shape Blur, Median, Polar Coordinates |
| Adobe Premeire Pro CS6 | Version 6.0.0.0, 6.61 GB MXF Project to H.264 to H.264 Blu-ray, Output 1920x1080, Maximum Quality |
| Adobe Acrobat 11 | Version 11.0.0.379: Print PDF from 115 Page PowerPoint, 128-bit RC4 Encryption |
| File Compression | |
| WinZip | Version 17.0 Pro: THG-Workload (1.3 GB) to ZIP, command line switches "-a -ez -p -r" |
| WinRAR | Version 4.2: THG-Workload (1.3 GB) to RAR, command line switches "winrar a -r -m3" |
| 7-Zip | Version 9.28: THG-Workload (1.3 GB) to .7z, command line switches "a -t7z -r -m0=LZMA2 -mx=5" |
| Synthetic Benchmarks and Settings | |
| 3DMark Professional Edition | Version: 1.1, Benchmark Only, Ice Storm, Cloud Gate tests |
| 3DMark 11 | Version: 1.0.1.0, Benchmark Only |
| PCMark 8 | Version: 1.0.0 x64, Full Test |
| SiSoftware Sandra | Version 2013.10.19.50, CPU Test = CPU Arithmetic / Cryptography, Memory Test = Bandwidth Benchmark |
Sandra's Memory Bandwidth module shows the benefits of faster data transfer, with DDR3-1866 outperforming DDR3-1333 by nearly 40%.

Sandra also demonstrates the superior latency performance of our LGA 1150-based platform compared to the embedded version, though it's unclear if the soldered-on CPU's slightly different specifications (like a smaller shared L3 cache) have an impact on timings.

The socketed platform is so much quicker that its DDR3-1600 setting significantly outpaces the embedded platform’s DDR3-1866 results, and its DDR3-1333 setting matches the embedded unit’s DDR3-1600 benchmark numbers. Since this is the same memory, does this spell doom for the embedded platform's application performance?
Though the same memory serves on-die graphics, 3DMark 11 actually suggests that CPU-heavy tasks gain most from increased DRAM data rates.

I began 3DMark Professional Edition testing with the app's low-end Ice Storm test, only to have a "what the...?" moment after switching over to integrated graphics. Looking for performance differences attributable to memory, I was stunned to see just how badly Intel gimped the GPU on its LGA 1150-based processor compared to the embedded version. We told Intel this before Haswell launched, but I believe the company underestimated the market for a socketed processor with Iris Pro graphics 5200.

Cloud Gate pulls the Core i7-4770K's HD Graphics 4600 into the low 20s, while the embedded CPU's Iris Pro graphics 5200 engine cruises through unscathed. Again we see only minor performance differences attributable to memory data rate.

I chose 1280x720 for game testing because that resolution puts relatively little demand on our on-die GPUs and because it upscales nicely on 1920x1080 displays. I then increased the detail levels I was using to maximize my enjoyment of each title.

Faster memory boosts the Core i7-4770K's HD Graphics 4600 engine by around 11% in Battlefield 3; the Core i7--4770R's average FPS only rises by 3.2%. The big difference can be seen in minimum FPS, which picks up by 10%.

Far Cry 3 doesn’t need super-fast RAM, but it also doesn’t like DDR3-1333 on the weaker Core i7-4770K.

Would you believe it? A demanding title like Metro: Last Light actually runs on platforms with integrated graphics. Don’t let the Low quality preset fool you; the game still looks gorgeous. Faster memory helps where it matters most: the minimum frame rate figures.

We've often found F1 2012 to be limited by either CPU or DRAM performance. Of course, when you're relying on integrated graphics, that can hold you back, too. The HD Graphics 4600 engine's performance is crippled by anti-aliasing in this title. Unfortunately, the game looks pretty terrible at low resolutions with AA disabled. DDR3-1333 kneecaps the LGA 1150-based platform, while the Core i7-4770R with Iris Pro graphics 5200 cruises through as if it were a completely different technology.
Our single-threaded Acrobat workload sees some benefit from faster memory. The Acrobat test also gains from the Core i7-4770K's higher thermal ceiling and larger shared L3 cache.

We typically disregard performance differences of one second or less, since rounding can exacerbate tiny deltas. We don’t see a consistent trend in memory performance in After Effects; we only know that maximum memory capacity plays a big role in how long this test takes to finish.

Data rate has no discernible effect on Photoshop’s CPU-heavy filters, but faster memory does help its OpenCL-accelerated performance. The LGA 1150-based Core i7-4770K is particularly limited by memory throughput.

Our Adobe Premiere Pro workload doesn't correlate clearly to memory performance. Instead, we end up with a five-second margin of error.

WinRAR, WinZip, and 7-Zip benefit greatly from fast memory when it comes to compressing files on the Core i7-4770K and -4770R.



Continuous testing showed a one to four watt difference between 1.50 and 1.35 volt memory settings. This difference remained consistent through all speeds, though it’s only shown for the latency-optimized setting. Data rate appears to have a larger effect on memory power consumption, but only when the RAM is loaded and unloaded at unrealistic capacity levels.

We were hoping that the Core i7-4770K could realistically represent the BGA-based Core i7-4770R at both voltage levels, but that’s just not the situation. The LGA 1150-based CPU enjoys a performance jump averaging 17% in the transition from DDR3-1333 to DDR3-1866. The same data rate increase only affected the Core i7-4770R by a mere 4%.

You may have noticed that we have two baselines in the charts. That's an important fact, particularly when you look at the efficiency calculations. Both platforms compare the efficiency of DDR3-1866 and DDR3-1600 results to their own DDR3-1333 baseline.

The embedded platform with Intel's Core i7-4770R becomes 2.4% more efficient at DDR3-1600, but a jump in power demand drops its DDR3-1866 result to sub-baseline levels. The socketed Core i7-4770K also enjoys its highest efficiency at DDR3-1600, but stepping up to DDR3-1866 helps performance enough to partly offset increased power consumption.
Our performance data tells us that the best market for fast SO-DIMMs is likely compact machines with LGA 1150-based CPUs, even though the memory form factor is more specifically intended for notebooks, many of which instead employ low-voltage processors soldered down. Power measurements further specify that low-latency DDR3-1600 is the best memory for our specific Core i7-4770R-based platform.
Switching our attention over to graphics performance, it was more startling to compare HD Graphics 4600 to Iris Pro graphics 5200 than it was to gather the memory-oriented benchmarks, warranting further exploration. It'll be more difficult to explain why the -4770R saw so much less benefit from higher data rates than the -4770K. Further testing of the embedded system would require me to acquire a Core i7-4930MX model that’s otherwise identically equipped.