The Bottom Line
This is a huge engineering win for Intel, but the compute cores themselves are the weakest aspect of this product launch.
The Core Ultra 5 and 9 desktop CPUs may have launched too early because there are significant new platform teething issues.
The gaming performance is lackluster (perhaps due in part to new platform issues).
There will be huge variability in game testing and benchmarking for launch reviews I suspect. This seems to be down to DDR5 memory kits, CUDIMM (or not) and luck-of-the-draw memory training issues. If you pick up one of these CPUs you may have to spend time training memory. It might work fine setting the XMP profile. It might not.
APO, a new software utility meant to help with cores, scheduling and game optimization, is off to a modest, but promising start.
New Core Ultra parts are only marginally more power efficient at out-of-box defaults than their 14th gen counterparts, though tuning is possible. It is possible for an enthusiast to squeeze signifcantly more performance-per-watt out of these CPUs than their prior-generation intel counterparts.
There is less variability in default power settings vs 14th gen motherboard, which is good.
However variability in other default settings across board vendors can lead to significant (multiple % points) benchmark result variation.
It is a tough competitive landscape for Intel; Intel’s Z890 platform generally has better features at a somewhat better price point, but AMD provides significant more overall value.
Windows 24H2, even insider build 2152, still suffers from more than 3% performance loss in a lot of scenarios with virtualization based security. This is unacceptable.
These factors taken together are sort of a ‘death by a thousand papercuts’ scenario for Intel. I am hopeful intel can improve a lot of these rough spots, but the reality is likely that for gaming workloads the Core Ultra 5 is currently generally a worse choice for gaming than 14th gen i5; The older i9 doesn’t have as many features as the Core Ultra 9 285k, and the Ultra 9 is a bit faster in multicore workloads, but single-thread performance is generally not as good.
This design and approach for client CPUs is still absolutely the right move; I love the improvements I’m seeing in the LGA 1851 platform. I am certain this is the ointment for what ails intel, but currently there are a large number of flies in said ointment.
If you are on Intel 11th gen or older, it is time to consider upgrade options. I wouldn’t necessarily rule out this platform based on my experiences, but please have realistic expectations when working on your upgrade.
Review
I think a lot of gamers will be disappointed by today’s CPU launch. In a best-case scenario, the Core Ultra 9 285K is not enough better than its 14th generation counterpart to really be exciting. This is perhaps unexpected as this is a totally new platform, LGA 1851, and does bring some awesome new features – more Gen5 PCIe lanes for storage, much better high-speed DDR5 support including verified-working DDR5-8000 CUDIMMs.
This disappointment from gamers is likely to overshadow the significant engineering wins at the heart of this new platform. This platform represents what intel should have been working on for two, maybe three, cpu generations now.
And while these 15th gen CPUs struggle to beat their 14th gen counterparts in gaming scenarios by more than margin-of-error margins, the CPUs do eke out a win for multicore workloads. The memory situation is a bit give-and-take – the memory bandwidth benchmarks to a record-setting 121-123 gigabytes pers econd, which is an impressive feat for just two dimms on a DDR5 platform, but the memory latency has increased to around 80 nanoseconds.
Further complicating matters for the memory situations, users may experience platform teething issues with consistent memory training and performance across a variety of DDR5 kits. As we swept through various motherboards and system configurations, it was possible to get the memory into a sub-optimal set of timings or other configuration problem that would result in memory latencies at 120 nanoseconds or more, which typically negatively impact high frame rate gaming scenarios the most.
Gaming
Neither the Core Ultra 9 nor the Core Ultra 5 did well with the somewhat obscure gems we love at LevelTechs – Factorio, Stellaris, and Dwarf Fortress. The performance in these games is significantly higher on prior generation Intel products and competing AMD CPUs.
Similarly, our custom benchmark runs with Baldur’s Gate showed up to 43% performance regression over the prior-generation Core i9 counterparts. This represents the worst-case-scenario we say when using DDR5 memory with sub-optimal timings at 6400. The exact same kit of memory on the older Core i9 performed flawlesly.
There are some interesting observations in the overall gaming data, but the bottom line is somewhere between a –5% and +5% performance bracket centered right on the Core i9 14900k. Some games are better than the older core i9; some games are worse. Generally, for gaming workloads, this CPU is somewhere between 5% and 8% worse than a 7800X3D, with particular games being as much as 15% worse.
The problem we alluded to in the video, however, is that there can be significant variation depending on what is ordinarily almost inconsequential configuration and memory differences. There
Compounding matters is the fact that boards we tested from ASRock, MSI, Asus and Gigabyte all had significnantly different out-of-box bios defaults.
Intel’s expectation was that bios settings necessary to use Intel’s APO software would be enabled out of the box but this was only true half the time.
CUDIMM, or not, seemed to have a big impact on frame rate and frame times, which makes like-for-like benchmarking tricky to do.
Windows also played a role in adding to the nose here – Windows 24H2 generally performed better, but a few specific games like Cyberpunk 2077 and Metro: Exodus performed significantly better on the insider version of Windows 11 build 2152.
I experienced an issue where, when Virtualization is disabled in bios, running any easy anti-cheat game would result in a BSOD.
Microsoft issued a bulletin advising updating games to work around this issue, but the specific failure I experienced was spooky in that immediately following this BSOD if one were to boot to Memtest86, then Memtest86 would detect a memory error (or two).
This is alarming because cold-booting the system would result in a system with no memory errors. This was repeatable across two different kits of memory from two different vendors. I suspect, but I am not certain, this means that the Easy AntiCheat BSOD somehow puts the CPU into a state where memory operations become unreliable once whatever condition EasyAntiCheat. Huge, if true. Or maybe two memory kits happened to be flaky in an almost identical way, never reporting any memory errors even with overnight testing, except immediately following an Easy AntiCheat crash . . . .
Multicore
The multicore is pretty solid overall performance; the Geekbench scores are good here. It would have been nice to see more performance uplift in multicore, however. Certainly this platform has a curiously strong memory bandwidth capability, so I believe it has the capability to be monumental.
APO
This is intel’s new software package that targets specific game optimizations mainly around scheduling and power management. I found that in the current state of APO it didn’t do much and that it was necessary to set “high performance” mode in windows to get games to be on par or slightly better than the 14900k benchmark scores.
APO is the right approach, though. Gamers will benefit tremendously from a program giving the operating systems about the best way to schedule threads.
Linux Related
I am experimenting with an update to the Kernel and will post a Linux-specific video later today. To a small extent “Windows Being Weird” holds the new arrow lake CPUs back, but we’re not talking +15% better.
Multicore on Linux is pretty good. This is the best scenario for the Core Ultra 9; The Core Ultra 5 is in an awkward position just because of pricing, features, and prior gen or competitive parts. The Core Ultra 9 would make a good Linux workstation, but the gaming performance on Linux ranges from behind to well behind both Intel’s prior generation and AMD’s current and prior generation competing parts.
