Rocket Lake has been a bit of a mystery ever since we first heard about the chip. There have been questions about whether it was a true 14nm backport of a 10nm design, whether it would keep the full measure of clock speed Intel squeezed out of 14nm, and whether it will compete effectively against AMD’s Zen 3 architecture. We can’t answer that last question until the chips launch, but luckily we can take a crack at the first two.
First of all, Rocket Lake isn’t based on Willow Cove, the higher-frequency CPU that Intel debuted earlier this year. Rocket Lake’s CPU core, code-named Cypress Cove, is based on Sunny Cove, the CPU that powers its Ice Lake line of processors. This isn’t necessarily a problem; Intel was attempting to fix problems with scaling to high frequencies when it built Willow Cove. Because it’s based on an older 14nm process, Sunny Cove shouldn’t suffer the same declines.
Meet Cypress Cove
Cypress Cove is a backport of Sunny Cove, aka Ice Lake. Intel claims it is looking for double-digit IPC scaling, which we assume refers to IPC scaling without counting clock speed improvements over previous chips. Advantages of Cypress Cove include:
New support for AVX-512 baked into Intel desktop CPUs.
Integration of Intel’s new Xe graphics core
Support for up to 20 PCIe 4.0 lanes
The Xe-LP graphics are only listed as being 1.5x faster than Gen9 integrated performance. This says nothing good about the actual onboard GPU built into the desktop versions of these chips.
To put this in perspective, Tiger Lake’s Xe solution is ~2x faster than Ice Lake’s GPU. As this graph from Anandtech shows, Ice Lake’s Gen11 GPU was substantially faster than any Gen9 solution Intel ever shipped. There are limited numbers, but they range from a 2x advantage to a 1.27x advantage.
If Intel was putting a full Xe configuration on its die, we should be able to expect something like a 2.54 increase in raw GPU horsepower. TGL laptops can carry up to 96 EUs (execution units). We know that Xe is going to be much faster than Gen 9 thanks to efficiency improvements, but it sounds like Intel is going to use between 24 EUs and 32 EUs in its desktop chips. If it used 32 EUs instead of the 24 it had fielded previously, this would represent a 1.33x performance improvement on width alone. We know, however, that Xe makes some significant changes to Intel’s overall GPU design, which is probably where the rest of the improved performance is coming from. If Intel was using a 96 EU configuration in Rocket Lake, we’d expect performance to be 2-2.5x faster than Gen 9, not just 50 percent.
I had floated the idea earlier this year that Intel might compete against AMD by positioning its own CPUs at price points where the built-in Xe graphics could carve into markets where people want better graphics on a budget. As an added bonus, setting up a comparison against an AMD APU with a 64-96 EU Rocket Lake would probably have gone well for Intel, given how Tiger Lake compares against the Ryzen 4000 series.
But if Intel is sticking to 32 EUs on desktop, that’s not going to happen. Gen9 simply wasn’t fast enough for a 1.5x performance increase to get you much in the way of useful gaming performance. If you start at 30fps, a 1.5x increase gets you to 45fps — a visible gain while playing. If you start at 15fps and pick up the same +50 percent, you’ve hit 22.5 frames per second — well below the playable threshold for the majority of titles.
Intel, apparently, still views its integrated graphics as worthless to its desktop line, which is deeply unfortunate. Companies like Topaz Video Enhance AI are actively working to improve Intel GPU support, and a more robust iGPU in desktop would have helped this a great deal.
Also, unlike Comet Lake, Rocket Lake will drop back to just eight cores. How much of an impact this will have on the product remains to be seen. Imagine, for example, that Intel hits the ~1.16 IPC uplift for Ice Lake and keeps every bit of its frequency at 14nm — or maybe even hits slightly higher clocks, since it doesn’t have to power 10 chips in the same die. That would be a 1.16xc IPC jump combined with, say, a 5 percent gain in real clock speeds, or about 1.22x uplift overall. An eight-core chip with a 1.22x performance uplift would be just slightly slower (theoretically) than a 10-core chip that lacked these advantages.
A few other notes: Intel doesn’t rule out releasing chips without integrated graphics, similar to its current “KF” CPUs, and it says that the x4 PCIe lane off the CPU is “only validated for discrete graphics, storage, or Intel Optane.” Not sure if anyone is likely to build a PCIe 4.0 sound card any time soon, so this is unlikely to be an issue. There’s also backward compatibility with the existing 400-series chipsets, with a new 500 series chipset coming as well, with unspecified features.
Intel’s single-thread performance jump would put it closer to par with AMD, but its multi-threaded performance might not get much better. Matching the performance of a 10-core with an eight-core is intrinsically impressive, but cleanly exceeding the performance of your 10-core with an eight-core in well-threaded code will require more than just a 1.16x IPC boost combined with a 5 percent frequency gain. A 10 percent improvement in sustained clocks combined with 1.16x IPC would do it, but it’s not clear if Intel’s 14nm process has any headroom left.
Getting AVX-512 into desktops is critical to encouraging developers to use it, but I think the most important narrative around Rocket Lake for Intel is that it’s no longer recycling an endless succession “14nm++++.” While these new chips are still built on 14nm, they will introduce real, distinct improvements to the Intel ecosystem, with everything from PCIe 4.0 to Xe support in an iGPU.
I’m not sure that’ll be enough — but it’s absolutely wonderful to see real competition in the CPU market again.