My initial investigation set out to determine whether AnandTech’s revisit of Broadwell in 2020 was reproducible. This extends that work by adding modern 8- and 4-core CPUs, and testing faster RAM. As we’ll see, Broadwell does notably better than Haswell, but can’t keep up with modern platforms in all games with fast(ish) RAM in 2022.
Broadwell in 2022
For today’s testing, we’re adding 2 new CPU platforms to our suite: the Ryzen 3 3300X to represent a fast quad-core CPU on a modern platform, and an i7-11700K to represent essentially some of the best gaming CPUs on the market. There are faster CPUs, especially if we look at Intel’s 12th gen Alder Lake parts that have just launched, but this is intended to be representative. Hopefully in the future I’ll get an Alder Lake CPU in for testing.
Expanded Configurations
In addition, we’re adding faster RAM speeds to the testing: 3200MT/s for DDR4 and 2200MT/s for DDR3. For the DDR4, I chose the stock i7-6700K speeds and the XMP settings on the 2x8GB kit I had on hand. The idea here was to choose speeds and timings that a layman could easily achieve. For the DDR4, that meant just enabling XMP. For the DDR3, since the kit I have is DDR3-1600 MHz, that meant looking up reasonable XMP timings for DDR3-2133/DDR3-2200 and seeing how close to that I could get quickly without trying to iterate too many times (spending no longer than 20 minutes optimizing timings).
Starting off with the expanded spec table, this is what we’re working with:
As an aside: I eventually tuned that DDR3 RAM in to use as my primary gaming PC (my husband took the i7-11700K, as it’s much much better in FFXIV, as we’ll see). That was eventually set to DDR3-2133 with major timings at 10-11-10-27. The “tuned” version was not fully tested in this suite, but in the few tests I did, they were so close to the more relaxed timings at DDR3-2200 that I didn’t deem it necessary to re-test everything.
Testing Methodology
The original results at their stock RAM speeds were all retested. For the DDR3 platforms, I switched to a new motherboard to allow for overclocking the RAM, so rerunning those benchmarks was required. The i7-6700K was also retested at its non-overclocked DDR4 settings.
A quick note: The i7-11700K does support up to DDR4-3200, but I still tested it at DDR4-2133, as that’s what the RAM defaulted to when I booted the system. Additionally, I’m trying to determine how RAM speed affects gaming workloads, so it made sense to also test what should be the fastest CPU for gaming at the slower RAM speed as well.
The games themselves were generally run at their highest settings. I wanted to answer the question “what is the real gaming difference between these CPUs given how I tend to set up my games?” This is typically how I play games – with the max settings I can manage while hitting the refresh rate I want (typically 60 fps). This is not the be-all and end-all of whether one CPU is the absolute very best or not.
Each benchmark was run 4 times and the results averaged to obtain the results below. They were tested in the same location, but there was some temperature fluctuations in that room throughout the testing. However, no appreciable difference was seen in GPU clock behavior – the ASUS ROG Strix 1080 Ti has plenty of cooling on an open bench. Ambient temperature fluctuations of 2-3 degrees Fahrenheit shouldn’t result in measurably difference performance.
The CPU cooler for the Intel parts was the Thermalright Peerless Assassin 120. Given that the CPUs were not overclocked, this should be sufficient to prevent thermal throttling. For AMD, the stock Wraith Spire cooler was used. I validated CPU clocks using this cooler, and no thermal throttling was seen at stock clocks.
1080p Results
General 1080p Scaling
Shadow of the Tomb Raider and FFXIV benefit the most from the 8-core 11700K, so we can posit that those two games are most sensitive to overall CPU throughput – either scaling to >4 cores, or taking advantage of faster single-core throughput.
Civilization VI is quite unique in this set, as it’s the only RTS here. The Ryzen 3 3300X outperforming the i7-11700K at all here is a bit of a surprise, but it was a consistent result. It’s possible there’s a confounding variable at play. The NVIDIA driver, Windows OS version, and RAM were identical, so I’m reluctant to try and guess what that could be. I wish I’d tested the cache access latencies on these platforms to have some more information to examine. AMD’s microarchitecture may have an inherent advantage over Intel’s in this title.
Shadow of War is a bit interesting – its results indicate there may be some advantage in that game to having the eDRAM present as the L4 victim cache. Then again, the chart topper here is the i7-6700K, so that can’t be said for certain. Typically Shadow of War is fairly light on its CPU requirements, and I was surprised to find any scaling at all here, even if it is “only” a 16.5% improvement from the worst to the best result.
Shadow of the Tomb Raider comes closest to what I would think of as the generally accepted gaming CPU ranking of the CPUs we have on hand. There’s not a massive difference between the top and bottom results, but it does show scaling. Note: When I tested the i7-11700K, the game itself was reporting that it was entirely GPU bound at 1080p.
Looking at the overall average across all games, the rankings aren’t too surprising. The 11700K and 3300X come out on top, with the rest of the ordering being what we would expect given the rest of the tech press’s CPU reviews.
RAM Scaling
Below are the charts for each game. These represent average framerates across four runs of the benchmark in each game. Every configuration is affected at least somewhat by the speed of the RAM in the system. The chart below helps illustrate this, as it shows the difference between “fast” and “slow” memory for each CPU.
The scaling on Skylake and Ryzen makes sense based on existing coverage of how Skylake and Ryzen scale with memory speed and latency. It’s interesting to see an effect on the older Haswell part – the i7-4790, even if that part is at the bottom of the charts and never matches the Broadwell Xeon. The older coverage from circa 2014 I can find of Haswell shows a smaller amount of inconsistent scaling with RAM speeds in older games.
As we might expect, the Broadwell part (Xeon E3 1285L v4) is markedly less affected by RAM speed than other CPUs. This is likely due to the eDRAM cache.
The i7-11700K Rocket Lake part is slightly less improved by memory speeds than the other CPUs, likely due to its 16MB of L3 cache. The 3300X on the other hand scales quite a bit because the RAM speed is tied to the CPU’s internal Infinity Fabric clock on this platform, meaning the internal communications in the CPU run faster as the RAM speed increases. The 11700K’s ring-bus runs on a clock separate from the DRAM, so it shows less improvement.
I think the most interesting results here, and the ones that we wanted to investigate in the first place, come from the i7-6700K. It has some clear differences between slow and fast memory, similar to the 3300X. In every game tested, the 6700K with 3200 MT/s RAM beats the Xeon, if only by small margins in some cases. The 2133MT/s RAM configuration loses to or ties the Broadwell Xeon in all of these tests.
1440p Results
At 1440p the bottleneck is moving more to the GPU, and there’s much less difference in most games between these CPUs. Civilization VI shows similar scaling to the 1080p results, which goes to show just how little it demands in terms of graphics, or possibly how much it demands of a CPU.
FFXIV again shows some scaling, but not nearly as much as it did at 1080p. In FFXIV, there appear to still be scenes where the game is held up by the CPU judging by its framerate graph during the benchmark (not shown here).
In most of the other CPU configurations, the differences are academic and may point to some other confounding variable(s) being present. I’m suspicious of the i7-11700K results, but I was unable to determine what, if anything, was consistently causing it to post results a few percentage points slower than most of the other CPUs here.
The chart below summarizes performance by looking at the average framerate across all games and comparing the CPUs with fast RAM vs slow RAM. This illustrates This chart shows the improvement seen for each CPU by using “fast” memory versus “slow” memory. At 1440p, there is some effect for the i7-4790, but relatively little difference for the other CPUs.
Interestingly, Shadow of War sees a decrease in performance on the 3300X and the 11700K with the faster RAM configuration. I don’t know why that would be. The effect is less than 3% degradation, but it’s measurable.
In general, RAM speed has a very small effect on average framerates at 1440p. FFXIV still shows that faster RAM can make a difference in games, but the improvement is at most 6%, which is seen in the 3300X results. The i7-4790 sees the largest improvement of any CPU in any game here in Civilization VI. As discussed before, RAM speed also affects the Infinity Fabric clock, so it makes sense that Ryzen would be more affected than the Intel platforms, even at 1440p.
The rest of the games show relatively little scaling, but I will note that the eDRAM appears to help the Xeon keep pace with more modern CPUs. The i7-4790 is at the bottom of most of the charts here, but shows the largest difference between its slow and fast RAM configurations. This should indicate that the eDRAM did improve some gaming workloads.
Compared to the other CPUs under test, the Broadwell part does well at 1440p. That said, there’s not a huge difference between the CPUs at 1440p. Some people may see a difference between the highest and lowest average FPS on this chart, but it would be difficult for the average gamer to notice.
2160p Results
I fully acknowledge that it’s essentially folly to test different CPUs at 4K, but given that I enjoy gaming at 4K when I can, and I already had everything set up, I figured collecting a round of results was warranted.
There’s next to no difference between the CPUs here, so there’s not much to analyze. Civilization VI and FFXIV show the largest differences across the tested CPUs, but even those differences are small. There are some unexpected results here in FFXIV, such as the 11700K performing worst, and the Xeon E3 1285L v4 showing 2 fps higher average framerates than all the other CPUs. Civilization VI shows some scaling, but nearly all the results are within 1 fps of each other at 93+ average fps.
This chart shows the improvement seen for each CPU by using “fast” memory versus “slow” memory. At 4K, it largely has no effect on average framerate.
I’ll be the first to admit that most of these games are not able to be played using this GPU at 4K. Four of the games have sub-60fps averages. However, the point of this was to see if there was any CPU scaling at this resolution with this 1080 Ti. I think I can conclusively say “not really, no”. All of the results are very close to one another.
The only anomaly is FFXIV, which for some reason consistently benchmarks a few percentage points higher on the Broadwell Xeon than on any other chip. This is a repeatable result. I also saw a similar result from 1440p in Shadow of War for the 11700K and the 3300X: the faster RAM ends up causing slightly lower average framerates.
Looking at our “average of averages” chart for 2160p, we can see that the Xeon does “best” on average across the 6 games tested here. The margins here are quite slim, and I don’t think in practice most would be able to tell the difference between any of these CPUs, barring differences in frametime consistency, which is not tested here.
Conclusions
RAM Scaling
The below chart compares the “slow” RAM configurations against the “fast” RAM configurations. This allows us to see how each individual CPU scales when RAM speed is increased. The Ryzen 3 3300X, the i7-6700K, and the i7-4790 all benefit reasonably well from faster RAM at 1080p. The Xeon generally doesn’t benefit enough to bother with, and at 1440p and above, generally none of the CPUs show significant benefits of faster RAM. This makes sense, as above 1080p the bottleneck moves more to the GPU.
These faster RAM kits were not available at the launch of Skylake, so Anandtech’s original testing of the Skylake parts did not use them. Anandtech also tests CPUs at JEDEC timings, which is much slower than what I test here. In 2022 it’s reasonable to assume that most enthusiasts have access to DDR4-3200, and the know-how to enable XMP in their personal systems. With faster RAM, the i7-6700K and the Ryzen 3 3300X matched or beat the Broadwell Xeon E3 1285L v4 at 1080p, whereas both CPUs fall behind the Broadwell CPU when paired with slower RAM.
Anandtech’s original article on Broadwell in 2020 is still valid and quite useful, as it shows how a OEM systems might perform, or what might happen if you neglect to enable XMP on your gaming system. The article is also useful for posterity’s sake: it shows what would’ve been seen at launch for each of the CPUs in Dr. Ian Cuttress’s CPU Overload project.
If you’re building a high end gaming PC, you’re already buying not-cheap parts, so buying not-slow RAM for a few dollars more just makes sense. I won’t say everyone needs DDR4-3200 CL14 RAM, but reasonably timed DDR4-3200 CL16 RAM isn’t expensive, and should be in all DDR4-based builds in 2022.
In short: buy at least DDR4-3200 and enable XMP in your BIOS! It’s free performance at this point going by DDR4 prices.
I’m hoping to get my hands on a fast GPU to see if high refresh rate gaming is affected any differently than the games and settings that were tested here. I would hypothesize that high refresh gaming will be more affected by RAM speed, but have no data to back that up at this time.
Broadwell in 2022
Based on the testing results, Broadwell looks to have held up quite well over the years. The i7-11700K is only 9.23% ahead of the Xeon E3 1285L v4 in average framerate across all games. Given that there is a 6-year gap between the release of these CPUs, I think Broadwell is still quite good for gaming, at least at these settings in these games. Broadwell is 14.56% ahead of the Haswell part when using 1600MT/s RAM as well.
Anyone on a Haswell system could potentially see a marked improvement from a new CPU, depending on the games they play. Upgrading an older system from Haswell to Broadwell is probably not worth it unless you can score a Broadwell chip for around $100 though, given how the new CPUs like the i5-12400 and i3-12100 are performing.
About 18-24 months ago, when you could find these Broadwell CPUs for a bit over $100, they made decent sense as an upgrade for those with older systems, especially those with an i5 from the Haswell generation. Currently on AliExpress and eBay, you can find the i7-5775C for $150, and the Xeon E3 v4 Broadwell models for about $160, so the pricing has increased significantly.
If you were looking to build a super budget gaming rig, picking up an i7 CPU alongside a used LGA 1150 motherboard that supports 5th gen CPUs, you could probably get away with getting CPU+motherboard+RAM for $250 or less if you were good at snagging deals. However, you’d also still be a little behind the i7-6700K with 3200MT/s RAM in gaming, which means you’d likely be similarly behind an i3-10100F. An i3-10100F build can also be done for about the same price, and with new parts instead of used. I don’t think Broadwell is a particularly good value if the CPU itself costs $150.
Limits of This Analysis
These results are also only valid for this particular GPU performance level. Once the GPU is more performant than a GTX 1080 Ti or roughly an RTX 3060, the CPU scaling may change, and may change significantly. Users who are looking at that level are already spending $600+ on a new GPU in this market, so they should consider also making sure they can make the most of their GPU by upgrading their CPU where applicable. I hope to get a new GPU to test this as well for the systems I still have on hand, but that’s much easier said than done these days.
As mentioned in the RAM Scaling section: while not tested here, I suspect that high-refresh gaming (120Hz and above) such as eSports titles (CS:GO, Overwatch, etc.) could be held back by a Broadwell CPU. Anandtech’s F1 2019 benchmark (known to be extremely CPU bound) shows this somewhat. The more modern chips are significantly ahead of the Broadwell CPU in those charts (the i7-5775C), but the Broadwell chip does alright still. However, once you move up to 1440p, you’re generally in a GPU-bound workload even in these kinds of games.
In the future, I would like to expand this analysis to get some smoothness metrics from a couple games to determine if Broadwell aids in smoothing out frame rate dips that might be seen on Haswell, and compare Broadwell’s in-game smoothness to the i7-11700K and the Ryzen 3 3300X.
[…] Further analysis can be found in Broadwell Revisit Pt 2: Now with Faster RAM. […]