- last edited on by
ROGBot
Continuous improvement drives the PC hardware industry forward. Take DDR4 memory, for instance. When it was introduced on the desktop, memory kits operated at a frequency of 2133MHz. Clock speeds increased incrementally over time, and nowadays, the top DDR4 kits can reach up to 4133MHz and beyond. But how much does faster memory really matter on a high-end desktop platform like Intel’s X299? We plugged a stack of DDR4-4133 sticks into ROG’s new Strix X399-E Gaming motherboard to find out.
Plotting the path to DDR4-4133
Before diving into performance, let’s take a retrospective look at the evolution of DDR4 memory support on Intel processors. Haswell-E was the first to adopt DDR4 memory in 2014. The technology was still in diapers at the time, and Intel officially supported speeds up to 2133MHz. While you could still push the memory frequency higher through overclocking, success was hit or miss.
The maximum officially supported speed remained unchanged when Intel launched its Skylake CPUs the following year. But the integrated memory controller in those chips was clearly stronger, and it wasn’t unusual to see enthusiasts running high-speed kits ranging from DDR4-3200 all the way up to DDR4-3866. Released in early 2017, Kaby Lake brought the first change to Intel’s CPU specs, with the stock memory speed bumped from 2133MHz to 2400MHz. Although this wasn’t an earth-shattering breakthrough, it was satisfying to see progress being made. Just a few months later, Kaby Lake-X and Skylake-X processors arrived with memory controllers tuned for DDR4-2666, which is where we stand today.
DDR4-2666 may be a step up from previous implementations, but it’s still far from the top speeds available with the fastest kits. The question is whether it’s worth spending more on higher-frequency RAM.
Corsair’s high-end Vengeance LPX 32GB DDR4-4133 (CMK32GX4M4E4133C19R) memory kit will help us find some answers. It comes with 19-26-26-45-2T timings and operates at a voltage of 1.4V. Underneath the bright red heat spreaders are delicious Samsung B-Die DDR4 chips that are notorious for scaling to high frequencies and handling lots of voltage.
The quad-channel kit provides one module for each memory channel on our Intel Core i7-7820X octa-core CPU. With double the number of DRAM channels available in Intel’s mainstream desktop processors, this Skylake-X chip isn’t exactly starved for memory bandwidth at stock speeds. Higher DDR4 frequencies may have less impact as a result.
The ROG Strix X299-E Gaming motherboard served as the foundation of our test system because it supports DDR4 memory speeds up to 4133MHz right out of the box. We got our Vengeance LPX kit up to the advertised speed with ease; all we had to do was select the appropriate XMP profile in the UEFI BIOS, and the motherboard took care of the rest.
On this platform, BCLK adjustments are mandatory for memory speeds of DDR4-4133 and higher. The board automatically increased the BCLK to 103.4MHz, which overclocked our i7-7820X CPU slightly, from 4.5GHz to 4.653GHz. To avoid giving that setup an unfair advantage, we used the same BCLK throughout our testing.
We tested performance at DDR4-2666 and DDR4-4133 speeds in various scenarios, including synthetic tests, real-world usage, and gaming. For DDR4-4133, we used the Corsair kit’s default 19-26-26-45-2T timings. But we tightened those latencies to 15-15-15-35-2T when running at DDR4-2666, because those are the most common timings for kits at that speed.
Test system and methodology
CPU: Intel Core i7-7820X
CPU Cooler: EKWB Predator 240
Motherboard: ASUS Strix X299-E Gaming
Memory: Corsair Vengeance LPX 32GB (4 x 8GB) 4133MHz
Storage: PNY CS1311 960GB
Video Card: ROG Strix GTX 1080 Ti
Power Supply: Seasonic Prime 750W
Operating System: Windows 10 Pro 64-bit with Creators Update
Drivers: NVIDIA 385.28 WHQL
Display: ROG Swift PG27AQ
We updated our benchmark software and game clients to the latest available versions before testing. For games, we tested at 1920 x 1080 resolution with a Strix GTX 1080 Ti to ensure there was no graphics bottleneck. The only in-game option we disabled was V-Sync. We used Fraps to capture individual frame times and then converted the data to FPS for easy interpretation.
AIDA64
Enthusiasts and overclockers embrace AIDA64 for evaluating memory bandwidth. The test does a good job of assessing read, write, and copy performance, so it’s a fitting place to start.
The results reveal the advantage of running faster memory. Write performance benefitted the most with a 43.5% increase going from DDR4-2666 to DDR4-4133. Read and copy performance increased by 27.7% and 26.8%, respectively.
The impact on memory latency was less impressive, with only a 9.1% reduction. Having to run the DDR4-4133 setup with looser timings likely contributed to the smaller difference here.
SiSoftware Sandra
Sandra is another popular utility for benchmarking system performance. Its memory benchmark is based on STREAM and measures sustained memory bandwidth instead of burst or peak speed.
Looking closely at the graph, you can see a consistent performance increase in each case. The improvements are all within 21.5-23.5%.
ROG RealBench
Synthetic tests can be great for showing performance gains, but the results don’t always translate to real-world usage. For a better sense of performance with everyday scenarios, we used ROG RealBench to test the impact of faster memory on image editing, video encoding, and general multitasking.
RealBench yielded mixed results. Image editing performance improved by 10%, while heavy multitasking registered a modest 3.4% boost. Video encoding performance wasn’t affected.
ATTO Disk Benchmark
With eight DRAM slots, the Strix X299-E Gaming lets you stick an insane amount of memory into your motherboard. It supports up to 128GB of DDR4, so running a RAM drive doesn’t sound like a ludicrous idea at all. We installed the latest revision of ROG RAMDisk II and created a 16GB RAM drive to evaluate the impact of faster memory when it’s configured as storage. ATTO measures read and write performance across different block sizes.
Faster memory had little impact on write speeds with the really small blocks, from 512B through 64KB, and with extremely large blocks, from 16MB up. The results bookended by those ranges are more mixed, with DDR4-4133 slightly ahead with some block sizes but trailing behind with others. It’s worth noting that the transfer rates are much lower than what we saw in the memory benchmarks, so frequency may not be the bottleneck here.
Read performance was consistent from 512B to 2KB, but the DDR4-4133 config surged ahead from there up to 64KB. DDR4-2666 came out ahead with most of the larger block sizes, although performance clearly plateaued for both setups. Again, overall speeds were much lower than in the memory benchmarks.
CrystalDiskMark
CrystalDiskMark is useful for testing the sequential and random performance of storage devices like our RAM drive.
Peak throughput was once again much lower than what we observed in synthetic memory benchmarks, suggesting that frequency didn’t bottleneck the performance of our RAM drive. The results were very close, with DDR4-2666 actually pulling slightly ahead in a couple tests.
CrystalDiskMark logged much higher sequential speeds with writes than with reads, but they were still well below the peak bandwidth measured in previous synthetic tests. This benchmark was basically unaffected by the system’s memory speed.
Overwatch
While increasing the memory speed to 4133MHz didn’t impact the average frame rate in Overwatch, the performance from frame to frame was more consistent, with tighter oscillation between highs and lows. Switching to DDR4-4133 also raised the minimum frame rate from 127 to 140 FPS.
PlayerUnknown's Battlegrounds (PUBG)
For the most part, frame rates in PUBG were a lot higher with DDR4-4133 compared to DDR4-2666. Gameplay felt smoother, and the average frame rate was 11% higher with the faster RAM.
Tom Clancy's Ghost Recon: Wildlands
Memory speed has no effect on Wildlands performance. Frame rates and variance are similar for both memory configurations.
Warhammer 40,000: Dawn of War III
In Dawn of War III, the DDR4-4133 configuration registered slightly higher frame rates. The average increased by 7.7%, and the dips weren’t as low as with DDR4-2666. Faster memory didn’t eliminate occasional frame drops completely, though.
The Witcher 3: Wild Hunt
We didn’t notice any memory-related changes in performance in The Witcher 3.
Crysis 3
In Crysis 3, frame rates oscillated slightly more with DDR4-2666. The minimum and average frame rates were largely unaffected by the memory speed, though.
Conclusion
Common sense tells us that faster is better. That statement remains true in most circumstances, but we must look at different angles when we evaluate memory performance. Our tests show that jumping from DDR4-2666 to DDR4-4133 produces big gains in synthetic benchmarks. However, there’s less of an advantage with real-world workloads, if there’s any impact at all. That’s the case with Skylake-X and Intel’s X299 platform, at least. Memory speed can make a bigger difference on other platforms, especially in dual-channel configurations with less bandwidth overall.
For X299, most users should prioritize memory capacity over speed. Separate the part of your budget that’s reserved for memory and pick the capacity that matches how you plan to use the system; content creators and power users should have more memory, while gamers can get away with less. Once you’ve decided on the capacity, purchase the fastest kit your budget allows.
In Dawn of War III, the DDR4-4133 configuration registered slightly higher frame rates. The average increased by 7.7%, and the dips weren’t as low as with DDR4-2666. Faster memory didn’t eliminate occasional frame drops completely, though.
The Witcher 3: Wild Hunt
We didn’t notice any memory-related changes in performance in The Witcher 3.
Crysis 3
In Crysis 3, frame rates oscillated slightly more with DDR4-2666. The minimum and average frame rates were largely unaffected by the memory speed, though.
Conclusion
Common sense tells us that faster is better. That statement remains true in most circumstances, but we must look at different angles when we evaluate memory performance. Our tests show that jumping from DDR4-2666 to DDR4-4133 produces big gains in synthetic benchmarks. However, there’s less of an advantage with real-world workloads, if there’s any impact at all. That’s the case with Skylake-X and Intel’s X299 platform, at least. Memory speed can make a bigger difference on other platforms, especially in dual-channel configurations with less bandwidth overall.
For X299, most users should prioritize memory capacity over speed. Separate the part of your budget that’s reserved for memory and pick the capacity that matches how you plan to use the system; content creators and power users should have more memory, while gamers can get away with less. Once you’ve decided on the capacity, purchase the fastest kit your budget allows.
4000-17-17-15-36-1T (tRFC : 305, tRFE : 32767)
AIDA64 Memory Benchmark
My System Memory Latancy : 49.7 (4000Mhz C17)
Your Testsetup Memory Latancy : 71.6 (4133Mhz C19)
RealBench
My System Realbench Encoding : 23,9077 (4000Mhz C17)
Your Teststeup Realbench Encoding : 30,8062 (4133Mhz C19)
More Important Memory Latancy
AIDA64 Memory Benchmark
My System Memory Latancy : 49.7 (4000Mhz C17)
Your Testsetup Memory Latancy : 71.6 (4133Mhz C19)
RealBench
My System Realbench Encoding : 23,9077 (4000Mhz C17)
Your Teststeup Realbench Encoding : 30,8062 (4133Mhz C19)
More Important Memory Latancy
YC2CHUR wrote:
4000-17-17-15-36-1T (tRFC : 305, tRFE : 32767)
AIDA64 Memory Benchmark
My System Memory Latancy : 49.7 (4000Mhz C17)
Your Testsetup Memory Latancy : 71.6 (4133Mhz C19)
RealBench
My System Realbench Encoding : 23,9077 (4000Mhz C17)
Your Teststeup Realbench Encoding : 30,8062 (4133Mhz C19)
More Important Memory Latancy
Your latency gain was from outofspec tref/trfc and rtl pairing with your cas and mesh clock. Most of your latency gain was from mesh clock and tref. Not much to do with ram speed entirely.,
His testing method is correct as its purely on ram speed for x299.*
Nate152 wrote:
Hi Chino
Alot of testing went on here I see, nicely done.
So faster ram can increase performance in some games with the biggest gain being with Minimum FPS.
That's important, min fps increase much slower than max.
That could prevent stuttering and similar issues and make games more fluid.
X299 have nice memory speed improvements compare to X99.
APEX have 4 DIMM slots and I saw on GSkill sites some insane fast QVL memories, both are 32GB but extremely expensive.
F4-4133C19Q-32GTZKKF Black
F4-4133C19Q-32GTZSWF Silver
Definitely worth upgrade from X99 but more for i9-7900X and up.
i7-7820X would be great for us with smaller budget, bad because Intel didn't want to offer full PCI-E lanes.
Korth, I think you missed Chinos point by your last paragraph. What he was getting at was depending on your heaviest use to budget capacity over speed and specifically stated gaming only rigs can get by with less. Me I run minimum 64GB for my heavy workloads like raw photo batch processing and video encoding where 16GB is a joke and will slow the process horrendously. I've also noted by personal experience that faster clock speed is not always better. The latency plays just as much of a roll as the clock cycles do. What good is having a clock cycle that's twice as fast if the latency is waiting for twice as long to commit a charge?
Korth wrote:
@JustinThyme -
Chino's summary: "[X299] content creators and power users should have more memory, while gamers can get away with less ... pick the capacity that matches how you plan to use the system ... once you've decided on the capacity, purchase the fastest kit your budget allows."
My response: "I'd say 16GB is plenty for any of the games ... even at "lowly" DDR4-2666 speeds ... the rest of the budget is better allocated towards getting faster GPU card(s) or faster CPU."
Your heavy workloads require a "minimum 64GB" because "16GB is a joke and will slow the process horrendously". My situation is similar. While we do use our systems for gaming we didn't allocate our budgets/priorities to gaming.
But gamers only need 16GB. There is quite seriously no need (and little ability) to fill up more memory unless running games (and operating systems) in multiple VMs or using tons of RAMDisk/RAMCache to fill it all up. So a gamer should choose 16GB capacity, then purchase the fastest kit budget allows.
My argument is that prices of faster memory slide ever upwards along an exponential curve. The most extreme fast top-end memory kit can cost as much as an entire computer in itself, each rung down from the top of the ladder drops the price by hundreds of dollars. While at the low-end the price curve is flatter, it's good practice to buy the fastest memory kit which hasn't begun the steep price climb. Definitely worth paying five or ten bucks more for DDR4-2666 instead of DDR4-2400. But the rest of the budget, for gamers, is generally far better allocated towards faster GPU/CPU than on faster memory - assuming it's about getting maximum performance on a budget, buying the specs which really matter the most.
You bring up good points about latencies. It's much less of an issue on quad-channel architectures which already (nearly) double total memory bandwidth. I never liked the high-latencies married to high-frequencies in DDR4, everybody handwaves them under "real performances" measured on synthetic benchmark scores. But these latencies do have "real performance" impacts on some kinds of work, at least with some kinds of hardware. I'm still experimenting to find the "best" balance with my Quadro cards, very finicky things when installed on a consumer platform.
I don't think we are understanding each other very well.
1st and foremost if you just dumped $3K+ on an X299 rig whats another couple hundred bucks for the best ram you can get? :rolleyes:
I need the capacity but what is missing from a lot of memory discussion is the real factor of how with increased speed the latency also increases. It does no good to have all the clock cycles if you spent them waiting to commit a charge.
Just one example same manufacturer Gskill Trident Z
3200 MHz CL 13
13/3200x1000=4.06 milliseconds
3466 MHz CL 14
14/3466x1000=4.03 milliseconds
4266 MHz CL 19
19/4266x1000=4.45 milliseconds
This was just a quick grab some results are even more extreme but in this case the 3200 MHz and 3466 MHz memory are actually faster then the 4266 by nearly 10%.
Compare any of these to Chinos test sample and they are all faster.
Corsair 4133MHz CL 19
19/4133x1000=4.59 milliseconds.
In the end, speed and throughput isn't about clock speeds. One must take all the attributes into account and do the math. :cool:
