Yesterday, I saw something really quite fascinating.
Anandtech, one of my formative technology review outlets, had secured Intel’s newest 8-core via an early retail sale and benchmarked it. The CPU itself isn’t even out until March 30th, and most tech press are under a news embargo, so they can’t share the information provided by Intel, but by getting a part at retail, Anandtech sidestepped all of that neatly and was able, legally by the terms of NDA, to publish a pre-release review.
Leading up to this point, the speculation has been that Rocket Lake CPUs would take back the performance crown from AMD, and Intel would be well on their way to fighting back the loss of marketshare they’ve had since AMD launched Ryzen in 2017.
And…well, that seems to be pretty incorrect, but it might not mean much anyways. Let’s discuss.
Rather than summarize the whole review (go read it at Anandtech, they’re a great site), I want to pull out what we expected and what we appear to have.
Intel’s Rocket Lake disclosures to date hinge on a few crucial points – that Rocket Lake offers a 19% IPC uplift over their past 5 years of CPU releases on the same 14nm silicon manufacturing process, that the clockspeeds of the final parts would be nearly identical to the 10th generation CPUs, and all of that paired with Xe-class integrated graphics and PCIE Gen 4.0 support would mean that these CPUs would be the new market leaders. Early leaked benchmarks seemed to back this up, with Rocket Lake taking higher single-threaded scores than the AMD competition, but losing in multi-threaded benchmarks, which has largely been the status quo from the Ryzen launch until Zen 3 parts launched late last year.
What the review from Anandtech suggest is that Intel is actually flat on their asses, losing still to the Ryzen 7 5800x in an 8-core versus 8-core showdown, but also in some cases even losing to prior-gen Intel parts, with both the 10700k and the 9900k winning a few benchmarks over their newer and supposedly better sibling. Yikes. In professional workloads like CPU rendering and other such tasks, the 11700k does often win over its older family members, but falls to the Zen 3-based 5800x, with few exceptions. Intel has one clear lead – in workloads that support the AVX-512 instruction set, but that is a category devoid of applications or uses, and while it may grow over time, adoption of AVX-512 has been anemic and there are not many signs of an impending growth spurt.
So what is my take?
Well, as an AMD fan trying to buy a Zen 3 CPU to finish my current build, I like this news. However, I think it is crucial to keep a clear mind and remove bias from the equation to examine this more critically.
Firstly, the biggest point I would make is one against taking this data at face value – which Anandtech makes themselves – while using a compatible motherboard of undisclosed make and model, the likelihood of microcode updates via future BIOS revisions is incredibly high, and those can and often will improve performance. Is the likely margin they’ll gain enough to offset some of the crippling defeats these benchmarks show? Probably not – you might see the 11700k overcome the older Intel parts, but the gap between these new parts and Zen 3 are pretty large. You’ll see some gains, and likely a lot of indirect gains that help the CPU slighty – better PCIE Gen 4 compatibility and management, tighter memory management and broader XMP compatibility, smoother interactions with the larger system, and things of that nature. However, you won’t find an extra 10-15% performance, or it is at least incredibly unlikely.
However, I think there is some good analysis in the article about why Rocket Lake likely underperforms expectations in these benchmarks, and it comes down to something that is simple in principle but a bit more complicated in reality. When Intel markets these CPUs as having an up to 19% IPC increase, what that means is that in an ideal workload will show that 19% boost clock-for-clock, while less optimal workloads will show less improvement. Anandtech’s findings are that the marketed IPC gain is almost entirely in floating-point work, with integer work taking a smaller increase. The improvements to both are still there, and still reasonably impressive – exiting the margin of error at the worst case (multi-threaded integer) and offering a solid boost, but if you were expecting 19% across the board flat, as Intel’s marketing would lead you to believe, you will end up disappointed, it seems.
There are other quirks of the design change to Intel’s first new desktop core in almost 6 years – latency. The major advantage of the Skylake-derived designs Intel has rolled with for so long now is that their use of Intel’s Ring Bus coupled with design maturity meant that core-to-core latency was pretty outstandingly low. The changes in core design for Rocket Lake mean new approaches to communication with both other CPU cores in the package, the L3 cache in the package, and memory outside of the CPU, all of which suffer from vastly increased latency. On paper, Anandtech’s result doesn’t sound that bad, an increase of between 1-34 cycles of latency. In practical terms, however, this creates an actual latency increase of nearly 1.5x, going from 18-24 nanoseconds of latency to 28-30! In gaming, this is the performance killer – Ryzen CPUs have been improving in gaming due to architectural improvements and higher IPC, but also largely because the core-to-core latency has been cut down with each generation. This regression is hitting Intel at the worst possible time for a few reasons – firstly, because it puts them down compared to AMD at a time where their victory is crucial, but secondly because this is the only part of this design we’ll get on desktop. After Rocket Lake, Intel’s next CPU release planned for desktop is Alder Lake, their first mainstream big/little core architecture, which will be wholly different yet again. As a result of that, developers will be less-incentivized to optimize around the limitations of Rocket Lake, and far more likely to continue to push Zen-friendly optimizations, especially in gaming where the Zen 2 cores are the next 5 years of console gaming as well as a growing segment of PC gaming!
It is also worth mentioning here that the i7-11700k is not the top Rocket Lake SKU or the fastest 8-core in the lineup, as both of those titles belong to the i9-11900k, which has a peak single-core clockspeed much higher than the 11700k tested, and a slightly-higher all-core turbo clock. A proper 11900k with optimized microcode and proper cooling will beat these numbers and will likely close in on the Ryzen 7 5800x. The real question is how close it will then get to the Ryzen 9 5900x or 5950x, both of which tend to boost slightly higher than the 5800x and offer more cores, more threads, and far more cache.
Lastly, Anandtech’s review notes some concerning power consumption and temperature figures. While the average gaming use is around 130-155w of power and a temperature of 60 degrees Celsius with a high-end air cooler, under AVX-512, the CPU appears to be allowed to run full speed, which is unusual for Intel’s AVX implementations, and under those conditions, it pulls 291w of power and peaked at 104 degrees Celsius, which is astoundingly bad for an 8-core desktop CPU. This also seems like a culprit of early microcode, and I suspect that release-ready microcode will place an AVX offset similar to past Intel CPUs to avoid letting them run all the way up on power and thermals. That microcode may also place limiters on turbo performance as well, which would be a downside for gaming and general non-AVX productivity tasks.
At the end of the day, though, the analysis here is kind of moot for a sort of current, topical reason. Right now, AMD sells every Ryzen 5000 chip they put out, except for the Ryzen 7 5800x, which is such a poor value that it is easily obtainable at most etailers in several countries for MSRP. Intel has an advantage that their competition does not, however – they fabricate their own chips. Instead of relying on third-party foundries to manufacture their silicon, they own a large network of fabs around the world in multiple countries. As a result, Intel is not bound to a limited contractual supply of wafers, but instead their own planning and manufacturing capacity resulting from said planning. Material shortages could still affect them, but in all likelihood, Intel is going to have a fairly good supply of these chips available right off the bat, releasing into a market starving for new CPUs. Provided final review samples and microcode don’t mirror these early results, there are a lot of people who are trying desperately to build a new gaming PC and will likely spring on these parts. Hell, even if they end up between the Ryzen 5000 CPUs and the 10th gen Intel Core parts, these will likely still sell in good volume because they would, in that case, still offer an improvement. It is hard to predict though, because while Intel can supply the market en-masse with 10th gen Core CPUs, they remain on shelves, indicating something of a lack of interest in those parts.
So for now, these benchmarks remain a bit of a curiosity – worrying for Intel and the public perception of their next-gen design, but ultimately difficult to put too much stock into as of yet.