Intel Core i9-11900K Rocket Lake Review: 14nm’s Final Burn

0
70


Intel’s new Rocket Lake desktop platform has arrived at a critical moment for the company. Just weeks ago, the CPU manufacturing giant unveiled a plan to overhaul its manufacturing business, open a client foundry business, build new fabs, and strike deeper partnerships with rival foundries like Samsung and TSMC. These initiatives are all part of new CEO Pat Gelsinger’s goal of reclaiming manufacturing leadership over the next four to five years.

Like the IDM 2.0 plans Intel unveiled earlier in March, Rocket Lake represents a fundamental break from the past. The 2011-day reign of Skylake is over. For the first time in five and a half years, Intel has a new desktop microarchitecture. One can imagine Intel engineers breaking out the champagne when Rocket Lake taped out, purely for the joy of building something else.

Rocket Lake, in a nutshell.

Unlike Skylake, Kaby Lake, Coffee Lake, Whiskey Lake, and Comet Lake, Intel’s Rocket Lake SoC uses the Cypress Cove microarchitecture. Cypress Cove is a backport of Intel’s 2019 Sunny Cove CPU design. Sunny Cove was built on 10nm and wasn’t capable of reaching high clocks. Backporting the design to 14nm allowed Intel to recover the clock it lost in the 14nm – 10nm transition.

Intel has promised Cypress Cove delivers a 1.19x IPC improvement compared to the 10th Gen Core family. Other new features debuting with Rocket Lake include a new, Xe-based GPU core (not something we’re discussing in this piece), support for the AVX-512 instruction set, an improved memory controller with support for higher-clocked DRAM, and an additional four lanes of CPU-attached PCIe 4.0 storage relative to what’s available on the Core i9-10900K. Rocket Lake also uses an x8 PCIe 3.0 chipset link if paired with a 500-series chipset. 10th Gen CPUs installed in the same boards use x4 links.

The Core i9-11900K uses an LGA1200 motherboard socket, which it shares with the 10th Gen CPU family. There’s some cross-compatibility between Intel’s 400-series motherboards and 11th Gen CPUs; check this article for details.

Competitive Positioning

Note: All references to MSRPs throughout this story are strictly hypothetical. AMD’s Ryzen 5000 CPUs are still carrying a premium and 11th Gen is too new to make any judgments about availability.

After stepping up to 10 cores last year with the Core i9-10900K, Intel is backing down to eight once more with the Core i9-11900K. During our briefing, the company tried to claim it was forced to reduce its CPU core counts because eight cores represented “the most we could fit.” This is ridiculous twaddle.

While we don’t have the exact die size of the Core i9-11900K, the 10-core Intel Core i9-10900K has a 206.1mm sq die. Back in the 22nm days, Intel’s Xeon Phi had a 705mm sq die, and the company hasn’t torn out its large-die manufacturing equipment, so reports of a supposed die size restriction are nonsensical. Our results suggest Intel dropped back to eight CPU cores because Cypress Cove’s power consumption on 14nm made more than that a bad idea, not because it couldn’t build a larger chip. It’s fine to say that a given core count represented the best overall design point.

The Core i9-11900K enters this fight hemmed in on every side. AMD’s Ryzen 7 5800X is theoretically less expensive, at $450. Intel chose to price the Core i9-11900K at $539 where it competes against the 12-core Ryzen 9 5900X at $549. Lower in the stack there are very well-positioned chips such as the Core i5-11600K, but we’ll be considering that CPU in a separate article.

The Core i9-11900K needs to match or exceed the performance of its predecessor and justify its price premium over the Ryzen 7 5800X. Ideally, it would carry the fight to the Ryzen 9 5900X, but asking an eight-core chip to beat a 12-core is a lot and Intel historically prices their CPUs at a premium compared with AMD.

Tests and System Configuration

We’ve ported our CPU results over to the same graphing engine we debuted for the 6700 XT launch last week. Our results are therefore organized a bit differently. The tabs in the graph below will display a different set of benchmark results depending on which option you select. You can click on the color buttons to remove a result if you’d like to compare specific solutions.

Application and game-specific configuration details are given below:

Blender Benchmark 2.0.5: We display the total wall-clock render time for multiple Blender scenes, including BMW, Fishy_Cat, Koro, Classroom, Pavilion, and Victor. This is a standalone benchmark tool provided by Blender that anyone can use. Rendering times are reported in minutes.

Cinbench R20 and R23 Cinebench R20 and R23 are an older and a newer version of the same benchmark. Maxon, the developer of Cinema4D, has created Cinebench to serve as a simple rendering test for their larger application. Rendering time is reported in scores and higher scores are faster.

Corona Render 1.3: Corona Render is an unbiased, physically based renderer with plugins available for applications like 3ds Max. Render times are reported in minutes.

Handbrake 1.33: We’ve temporarily reintroduced Handbrake as a video transcoding benchmark. We convert the 4K film Tears of Steel using the Fast 1080p30 preset, but we altered the preset to retain the source video’s frame rate. Performance is reported in both H.264 and H.265.

IndigoBench: IndigoBench tests IndigoRenderer, another unbiased, photorealistic GPU and CPU rendering package. Performance is reported in Msamples/sec and higher scores are faster.

V-Ray: V-ray, as the name implies, is a flexible ray tracing solution deployed across the 3D industry. We measure CPU performance in vsamples/sec, with higher scores indicating higher performance.

y-cruncher: We use y-cruncher to calculate the first 5 billion digits of Pi and report the wall-clock time required. We use an optimized executable corresponding to each CPU’s microarchitecture. The 11900K is tested with AVX-512 enabled, both Zen 5000 CPUs run Zen 2-optimized code, and the Core i9-10900K uses an executable optimized for Broadwell and Skylake. There is no Zen 3-optimized version (yet).

Next up, games:

Assassin’s Creed: Origins: Ultra Detail, DX11.

Borderlands 3: Ultra Detail, DX12.

Far Cry 5: Ultra Detail, High Detail Textures enabled, DX11.

Metro Exodus: Extreme Detail, Hairworks and advanced physics disabled, DX12.

Shadow of the Tomb Raider: Tested at High Detail, with SMAATx2 enabled. Uses DX12.

Strange Brigade: Ultra Detail, Vulkan.

All systems were tested with an MSI RTX 3080 Suprim X using Nvidia’s 461.92 driver series. The Core i9-11900K was tested on an Asus motherboard, while the AMD Ryzen 7 5800X and Ryzen 9 5900X were both tested in an MSI X570 Godlike motherboard. Both the Intel and the AMD systems were configured with 32GB of Crucial Ballistix DDR4-3600 RAM in four slots.

We had no stability problems with our Asus ROG Maximus XIII Hero beyond a dead M.2 slot. We opted to test the board using a PCIe 3.0 riser card. This worked but blocked our ability to test PCIe 4.0 southbridge support.

Performance Analysis:

Our performance data shows that the Core i9-11900K mostly matches the multi-threaded performance of the Core i9-10900K, while substantially exceeding it in single-threaded tests. In some cases, the Cypress Cove-powered Rocket Lake is even faster than its 10-core predecessor. Single-threaded performance is about 5 percent higher than AMD in tests like Cinebench and no less than 20 percent and 23 percent higher than the Core i9-10900K. Single-threaded workloads are a minority of tasks these days, but the uplift here is very real.

Intel is generally faster than the 5800X, though there are exceptions, like NeatBench 5 and 7zip’s compression test. While the 11900K is capable of besting the Ryzen 7 5800X, it’s got no chance against the Ryzen 9 5900X. Unfortunately for Intel, the 5900X is actually its closest price match. The Core i9-11900K cannot match the Ryzen 9 5900X in any of our multi-threaded application tests.

Let’s turn our attention to gaming:

The 5900X doesn’t offer much uplift in gaming compared with CPUs such as the Core i9-10900K. In 1080p, it leads the Ryzen 7 5800X by about 4.6 percent and the Ryzen 9 5900X by 2.6 percent. In 4K, there’s a 6 percent gap between the 5800X and the 11900K and a 4.5 percent gap between it and the 5900X.

As for power consumption, here’s how these four CPUs compare. Power consumption is higher here than you may see in other reviews because we’re using an MSI RTX 3080 Suprim X. This particular RTX 3080 is tuned for maximum performance and its power consumption figures reflect it.

Intel’s Power Consumption: Up, Up, and Outré

We’ve measured power in y-cruncher, using a binary tuned for each CPU. This means AVX-512 for Core i9-11900K, Broadwell/Skylake tuning for the 10900K, and Zen 2 enhancements for the AMD CPUs. There isn’t a Zen 3-optimized executable yet.

“Rocket Lake” is an incredibly appropriate name for this SoC. While we’re reporting wall power and not CPU-specific power consumption, y-cruncher has a text GUI and no graphical UI. The Core i9-11900K draws 1.73x more power than the 5800X and 1.25x more than the Core i9-10900K, even though it has fewer CPU cores. If we were using AMD’s more power-efficient X470 platform we might be able to fit two Ryzen 7 5800X’s into the same power envelope as one Core i9-11900K.

The difference between the Core i9-11900K and the Ryzen 7 5800X is large enough to make a difference to your yearly power bill, depending on how much you use the chip. If you ran both CPUs all-out for four hours per day and turned them off at other times, you’d pay $70.08 per year for the 11900K and $38.16 for the AMD CPU. If we assume 24/7 continuous operation at $0.12 per kilowatt-hour, the Core i9-11900K would run $420.48 per year against the AMD 5800X at $229.08. Neither figure accounts for the increased HVAC costs during the summer or the potential reduction in heating costs during the winter.

Conclusion: The Core i9-11900K Is Intel’s FX-9590 Moment

The Core i9-11900K’s market position reminds me of AMD’s FX-9590 launch. Believe it or not, I mean that as a compliment.

The FX-9590 was an AMD Piledriver CPU with a 4.7GHz base clock and a 5GHz boost. AMD launched it in the spring of 2013 to give its users a high(er)-performance alternative to the FX-8350. AMD couldn’t challenge Intel’s absolute performance or power efficiency at the time, but it still decided to ship the highest-performing CPU it could as an option for its own fans. I praised AMD for catering to its own fans in 2013 and I’m giving Intel a nod for doing the same thing here.

The Core i9-11900K does some things right. It sets a new high mark for single-threaded desktop CPU performance, holds an edge in gaming, and modestly leads the Ryzen 7 5800X in a number of our CPU tests. It offers PCIe 4.0 and support for AVX-512. We’re glad to see the latter rolling out in more systems, as it increases the chance of wider software support.

These strengths, unfortunately, do not outweigh the CPU’s strategic weaknesses. Outside 1T performance, the Core i9-11900K is a sidegrade from the Core i9-10900K. Intel customers still using 4-6 core CPUs would see much larger performance improvements, but AMD’s Ryzen family offers 12 cores in the same price band where Intel sells eight. The Core i9-11900K cannot match the multi-threading performance of the Ryzen 9 5900X or 5950X and it cannot match the power efficiency of the Ryzen 7 5800X. Like the FX-9590, the Core i9-11900K has been pushed far out of its metaphorical comfort zone to provide maximum performance at maximum TDP. Like the FX-9590, its appeal is limited to the enthusiast faithful rather than the general market.

It seems unlikely that the present incarnation of Rocket Lake is what Intel was hoping for when it decided to backport the Sunny Cove microarchitecture to 14nm. The team successfully brought Ice Lake’s IPC improvements backward a generation, and they did it with only minimal tweaks to top-end clocks, but the power consumption trade-off being made here is ugly.

The Core i5-11600K is winning praise for its price and positioning relative to AMD’s 5600X and we intend to take a gander at both cores in the near future, but the Core i9-11900K we’ve reviewed here today is hard to recommend. Multi-threaded performance against the Core i9-10900K is flattish to slightly down. Game performance uplift is in the mid-single digits. CPUs like the Ryzen 9 5900X outperform the Core i9-11900K in every multi-threaded non-gaming workload.

Neither AM4 nor LGA1200 are expected to be long-lived platforms at this point, but AMD still has an advantage over Intel where upgrades are concerned. Even if Ryzen 5000 was the last AM4 product AMD ever launched — and it probably won’t be — anyone who buys an eight-core Ryzen 7 5800X today has the option to step up to a 16-core Ryzen 9 5950X at some point in the future. In contrast, the Core i9-11900K is the fastest Rocket Lake CPU we’ll ever see.

Rocket Lake (desktop) and Tiger Lake (mobile) will both be succeeded by Alder Lake, Intel’s upcoming hybrid architecture that pairs small, energy-efficient Atom cores with larger big-core chips. For now, we’re presuming that Intel will launch on mobile first in late 2021 and then follow-up with a desktop refresh in 2022. We still think this is the most likely outcome, but the Core i9-11900K’s power efficiency problems make the idea of a fast replacement cycle more plausible.

One thing Rocket Lake does make clear: Intel’s 14nm node is well and truly tapped. There are no more pluses to wring out, no spare efficiency stuck amongst the couch cushions. Intel enthusiasts who are still back on slower CPUs with lower core counts may find the Core i9-11900K’s high IPC and 5GHz+ boost frequencies attractive, provided they’ve got coolers that can handle the heat. Anyone who is not specifically looking for an Intel processor may find they are better served by an AMD Ryzen 5000 CPU, assuming one can be found at MSRP.

Now Read:





Source link