After reading AnandTech’s article by Dr. Ian Cuttress, A Broadwell Retrospective Review in 2020: Is eDRAM Still Worth It?, I was quite surprised at how Intel’s Broadwell processors have held up over the years compared to their predecessor, Haswell. I wanted to see if I could reproduce those results, and happened to have a lot of Haswell-era PC parts around, so I was a few CPU purchases away from being able to test those results.
I jumped on Aliexpress and eBay to see what I could find for a reasonable price. I managed to snag two Broadwell CPUs: both an Intel i7-5775C and an Intel Xeon E3-1285L v4 for $115 each. A year later those each go for upwards of $140, which is certainly unusual for computer parts – they rarely appreciate in value over time. I like to think Dr. Cuttress’s article had a little something to do with that.
Broadwell: Background
Broadwell was always destined to be a neglected child of Intel’s CPU generations. First, it was late to market. One could argue this was the first late chip from Intel that was the beginning of the ongoing trend of missing their goals on new process technologies. This was the first 14nm chip from Intel, and this 14nm process has since become a meme in the computer hardware community (how many pluses are we up to?). Second, Broadwell was limited in its consumer desktop release. Broadwell-E was apparently good enough to release many high-end desktop (HEDT) CPUs, but the mainstream platform, Broadwell, only ever got two socketed CPUs for consumers: the i7-5775C and the i5-5675C. There were a further six Xeon SKUs as well, but taken as a whole, this is still far fewer than a “normal” consumer release on the mainstream platform for Intel.
Third and finally, this was the second time Intel was releasing a chip with an integrated GPU they claimed could compete with AMD’s offerings. This was also the first time Intel was releasing a socketed chip with such GPU. AMD had been outdoing Intel for a while in integrated GPUs on socketed desktop parts, even with their now-infamous Bulldozer CPU core architecture holding back their overall market penetration.
Embedded DRAM (eDRAM)
Intel was competing with AMD partly via the use of embedded DRAM (or eDRAM), which sat on-package next to the CPU die. The thinking at the time was that DDR3 couldn’t provide sufficient memory bandwidth for graphics-intensive workloads, so an additional cache was added that provided 50 GB/s of bandwidth in each direction. This cache could be used in conjunction with the DDR3 system memory, giving a theoretical maximum throughput of 75.6 GB/s (assuming dual-channel 1600MHz DDR3 memory is used).
Intel first shipped eDRAM in select BGA variants of their Haswell processors, clocking the eDRAM interface at 1600MHz. Intel has continued to ship eDRAM on several of its Skylake parts (and derivatives), however eDRAM never made it into another socketed processor from Intel aside from Broadwell. It seems that the largest customer of these eDRAM SKUs was Apple for their 13” Macbook Pro line. A few Intel mini computers (NUCs), and an assortment of devices from smaller manufacturers also used these SoCs with higher powered iGPUs, but they were relatively uncommon.
Given that Intel’s GPU architecture was essentially unchanged from 2015’s Skylake through 2020’s Comet Lake, it’s not a surprise that Intel continued to ship eDRAM on its parts through 2020. The GPU was designed to take advantage of that kind of bandwidth, and getting that level of bandwidth on the Skylake memory controller in a power-constrained environment is difficult. You would need DDR4-3200 to achieve the same amount of raw bandwidth, and even that would be at a higher latency than Intel’s eDRAM.
It can be assumed that Apple elected to favor power efficiency over memory bandwidth (and capacity) on the 13” Macbook Pro, given that the 13” Macbook Pro shipped with LPDDR3 through the end of its life in 2019. Looked at another way: the eDRAM on Skylake and its derivatives allowed Apple to use lower power LPDDR3 without sacrificing GPU power (in theory). This assumes that the eDRAM consumes less power than the delta in power between DDR4 and LPDDR3.
These days, any good dual-channel DDR4 implementation can match the bandwidth of the eDRAM in Broadwell, albeit at likely at a slightly higher latency. DDR5 and LPDDR5 are slated to take things even further (see: Valve’s semi-custom SoC in the Steam Deck with 88GB/s of bandwidth). That’s not to say the idea of a higher-than-RAM bandwidth additional cache on the CPU is being abandoned; quite the opposite in fact.
The next big eDRAM-like technology will be AMD’s V-Cache that they’re planning to bring to market in 2022 (although that’s technically just extra L3 cache at essentially the same latency as the existing L3 cache), and of course high-bandwidth memory (HBM). Intel just announced HBM enabled SKUs for their new Sapphire Rapids CPUs, so maybe one day we’ll see it in a consumer part.
Today’s Test
This brings us to today’s investigation: Haswell vs. Broadwell in games in 2021. It’s time to see if Dr. Cuttress’s findings hold true. To not bury the lede: yes, they do (generally).
For today’s test, we’ll be using the Xeon E3-1285L v4 to represent Broadwell and the i7-4790 to represent Haswell. These are not far off their corresponding flagship parts, and they clock quite close to one another (100MHz difference in all-core turbo speeds), so they should be interesting to test against each other.
Results
Average Framerate
On average at 1080p, Broadwell shows a 14.6% lead over Haswell in average framerate, which is notably more than its IPC or other core improvements would normally account for. In their initial Broadwell review, AnandTech also noted that the IPC (instructions-per-clock) improvement was in the mid-single-digits range, certainly below 10%.
Naturally, Broadwell’s advantage in average framerate starts to slip away as we move to 1440p (averaging 9.0% over the Haswell CPU), and vanishes entirely at the tested 4K resolution (results show 1.09% improvement for Broadwell).-
This is to be expected: as you increase the dependency on the GPU, the absolute speed of the CPU becomes less important. It’s why my favorite (tongue-in-cheek) advice to those looking to upgrade their CPU is: “Just play at 4K – then it won’t matter what CPU you have.”
That said, I do think a 9% advantage at 1440p is noticeable. That advantage increases in more CPU-intensive games like Civilization VI and FFXIV as well, but some games show almost no improvement (Shadow of War and Borderlands 3).
Anecdotally, Star Wars: Jedi Fallen Order also showed a noticeable bump from swapping to Broadwell from Haswell, but that game wasn’t in the cards to test this round since it doesn’t include a reproducible built in benchmark. I may add it for future testing if I can figure out a good way to reproduce the results, as it’s an interesting data point.
Minimum Framerate
Unfortunately for this data set, the minimum framerates and 99th percentile framerates were only gathered for a few games. The scaling here is a bit of a mixed bag: at 1080p, only FFXIV shows more uplift in minimum framerates than it does in average framerates.
At 1440p, FFXIV and Civilization VI show better-than-their-average scaling, but Shadow of War does not. Interestingly, at 4K, all 3 games show small but reproducible better-than-their-average scaling.
This is interesting behavior I think, but I can’t make heads or tails of a real pattern based on the data collected. Suffice to say: there’s something here, but without further investigation, we can’t posit exactly what trends may exist.
Conclusion
All in all this makes me very glad I snagged a couple of these very unique CPUs to use. I’ve always had a soft spot for interesting, unique hardware (ask me how my 2 months with a 2020 Surface Duo went), and it seems like these are a decent improvement from the Haswell systems I was running previously. They’re certainly the best CPUs for gaming for the LGA 1150 socket, barring my testing of an i7-4790K, hopefully coming soon. They have the added bonus of being rare and neat.
Check back for my next post to see how these two compare to Zen 3 and Skylake quad core parts. Additionally, I plan to do some memory latency and bandwidth testing, as there was some discussion that the original AnandTech article revisiting the Broadwell CPUs was unfair to the modern chips due to the latencies and speeds used for the memory.
Further analysis can be found in Broadwell Revisit Pt 2: Now with Faster RAM.
[…] initial investigation set out to determine whether AnandTech’s revisit of Broadwell in 2020 was reproducible. This […]