Intro
I recently decided to upgrade my personal workstation that was based on an AMD Ryzen 9 5950x CPU. When the 5950x launched it was top of the line, but since then there has been a lot of competition in the CPU space and I knew from working with newer generations of AMD CPUs that I could perform a noticeable upgrade to my machine. Unfortunately, being on the AM4 socket meant the 5950x was already top spec and so there were no “drop-in” upgrades available to me, which meant that I either had to upgrade to AM5 for the Ryzen 9000 series or socket LGA1851 for the latest Arrow Lake processors from Intel.
LGA1851 launched after AM5 launched and frankly the launch was underwhelming. What is more, Intel’s value proposition made their processors hard to justify, except for some niche workloads where they might have been on par with AMD or pulling slightly ahead. However, as I am sure many readers will have heard, Intel has been struggling as of late and have been failing to sell as many units (as a percentage of total market sales) as they had in their glory days. However, they are taking action and one of those actions was to cut prices across the board for their LGA1851 line-up. With the exception of perhaps their flagship, the Core Ultra 9 285k, which represents binned silicone from their production line. Meaning, the best silicone yield that is capable of sustaining a higher IPC (CPU frequency) across more cores. Additionally, they also added a performance profile, which essentially is a overclocking profile sanctioned by Intel. I.e., it does not void your warranty.
This seemed appealing to me, in part because their overclocking profile significantly increases the performance of their memory controller, and my workloads tend to be memory (RAM) intensive, but also because it makes it easier to recommend to others, since it’s both easier to implement and gives you some guarantees should something go wrong. Here is an overview over all the performance benefits from Intel’s 200S Boost Overclocking Profile:
Intel 200S Boost Settings
| Default Specs for Intel Core Ultra 200S -K Processors | 200S Boost Overclocking Profile for Intel Core Ultra 200S -K Processors | |
|---|---|---|
| Fabric (aka SoC Tile/NGU) | 2.6 GHz | Up to 3.2 GHz VccSA ≤ 1.20V |
| Die-to-Die (aka “D2D”) | 2.1 GHz | Up to 3.2 GHz VccSA ≤ 1.20V |
| DDR5 Memory UDIMM/CUDIMM 1 DIMM per channel (aka “1DPC”) | 6400 MT/s (3200 MHz) | Up to 8000 MT/s (4000 MHz) VDD2 ≤ 1.4V and VccSA ≤ 1.20V Module VDDQ and VDD ≤ 1.4V |
CPU
The Intel Core Ultra 7 265k is the second fastest SKU in Intel’s current Arrow Lake line-up (when counting all 285 and 265 variants as one SKU each). It has a total of 20 cores, 8 of which are “performance cores” and 12 of which are “efficiency cores”. This CPU has no multi-threading, so 20 cores = 20 threads. This, alongside with the asymmetrical architecture (performance and efficiency cores) is one of the bigger differences between AMD and Intel at this juncture. The Core Ultra 9 285k is the better CPU, naturally, it has 8 performance cores and 16 efficiency cores (with an uplift in CPU frequency), but will cost more than twice as much as the 265k. At that price point, it is up against some stiff competition from AMD. Both from the ~$100 cheaper 9950x and ~$100 more expensive 9950x3d.
CPU Cooler
Since we are essentially running an overclocking profile with this build, I wanted to make sure that cooling wasn’t an issue. The 360mm radiator of the Arctic Liquid Freezer III PRO has ample cooling power, with some cooling headroom. That being said, you could also air cool this build with a good air cooler like the Noctua NH-D15. The Noctua NH-D15 will perform similarly to a 240mm liquid cooling solution. I ran the NH-D15 for some time in my 5950x build and had no issues.
Motherboard
In order to run Intel’s 200S Boost profile, you need both compatible RAM and a compatible motherboard. You can find an overview over all the compatible hardware options on Intel’s 200S Boost Profile page. The Gigabyte Z890 Eagle Wifi 7 is one of the cheaper options from their list, but it is appropriate for my use case. Depending on availability and personal requirements, you may prefer to choose another board from their list. All of these approved partner boards are high-end and will work well in this configuration. Just make sure to update to the latest BIOS for the profile compatibility.
Memory
My choice of RAM is the Corsair Vengeance DDR5 CUDIMM 48GB (2 x 24GB) DDR5 8000MHz C38 Dual Channel Kit. Same as with motherboards, there is an approved list of RAM kits for this build. The memory controller is overclocked in the 200S Boost profile, and you will need RAM capable of running at 8000 MT/s and CL 38 timings. The requirements are for there to be 1 DIMM per channel. The DIMM can either be UDIMM or CUDIMM, I went with CUDIMM here. The “C” stands for “clocked” and indicates that the modules come with a clock driver to help maintain stability of the RAM at such high speeds. The caveat here is that at the time of writing there aren’t large kits of RAM that can run in this profile on this platform. 48GB is the upper limit for now. For my use case this is okay, since I am using this as a personal computer and am mostly interested in running smaller workloads as quickly as possible. For larger workloads that require my machine to run for several hours and/or days, I will use a server.
Storage
I’ve selected the WD Black SN 8100 2 TB based on it’s “4k mixed” performance, that is it’s ability to read and write small files quickly. This is supposed to represent a real world use case of reading and writing to something like an SQL database. Of course, it will also perform very well in other read/write related tasks, since it is a PCIe Gen 5 drive. The drive is also available in 4 TB.
Video Card
The Core Ultra 7 265k is capable on it’s own of driving multiple high resolution displays for productivity. With this motherboard, you should be able to drive 2x4k displays without issue, using the iGPU of the 265k. That being said, a dedicated GPU in a modern workstation has the additional benefit of facilitating local LLM inference. For this build I am using AMD’s new RDNA4 RX9070, however there are many options for various budgets and use cases. In my estimation, NVIDIA cards are “safer” due to their support of CUDA for LLM inference and training, and currently reasonable performance on both Windows and Linux. That being said, going with the theme of this build I have chosen to go with the slightly less intuitive option of an AMD graphics card. There are some benefits to this for Linux forward users, primarily due to AMD’s better support for open source drivers, which in theory should mean better support on Linux in terms of relative performance to Windows.
Case
The Lian Li Lancool 217 ATX Mid Tower Case is one of the best at it’s price point. It comes preinstalled with 4 fans and has very good out of the box air flow properties, for superior air cooling performance. PC cases can be somewhat personal and you may prefer a different design. Whatever you choose, pay attention to buying a case with good airflow, so that air can reach critical components, such as the CPU and GPU. Also consider that you may need to add additional fans, depending on what the case comes with and what your objectives are. For more information on this topic, I recommend checking out Noctua’s guide on setting up PC case airflow.
Power Supply (PSU)
SeaSonic might be the easiest PSU brand to recommend. High quality and the go-to brand for overclockers. This 750W 80+ gold unit has enough power for this build, with some additional headroom, should you choose to upgrade at a later point. One caveat here is that if you know that you will be using a very powerful GPU, like an RTX 5090, you will need a more powerful PSU. In that case I would look for a 1000W option like the Corsiar HX1000i.
Benchmarks
I performed a simple benchmark to gauge the performance difference between my old 5950x system and new 265k system. For the benchmark I used MonkerSolver and ran a 6-handed NLHE game tree. Under Windows 11, I was able to observe a performance uplift of ~+68% on the 265k. I also ran the same workload with the 265k on Cachy OS Linux. There I observed a ~+79% uplift.
Notes: Uplift in this context means increase in iterations per second. The 5950x was running stock and with 64GB of DDR4-3200 RAM. For the 6-handed NLHE game tree, the settings were:
Game tree:
‘Auto config’ = ‘default’
‘Add bets and raises according to custom action’ = ‘example’
Abstraction:
Buckets: 30/30/30
Texture: P/M/S
Conclusion
All in all, I am very happy with the performance increase I got for the 265k over the 5950x. My original purchase price for the 5950x with tax was ~$826 and for the 265k it was ~$320. So, considering I was able to spend so much less, whilst gaining ~+68-79% after 4 years, implies to me that the current state of the market is favourable to consumers. Who knows how much this CPU would have cost if Intel were still the dominant market player they once were. It could easily have cost double. I would definitely recommend a build like this to someone looking for a general purpose computer that is also good at small to medium simulation workloads, local database work and any other productivity task. Paired with a decent GPU, you can also add decent gaming performance and AI inference to the list.
Note: When I started this build there was uncertainty around whether Intel would launch any more processors for this socket. At the time of writing there seem to be rumours that there will be an Arrow Lake refresh. This would mean the potential to upgrade to a Core Ultra 9 385k at a later date. If this were the case, I’d expect a further performance increase of 20-30% over the 265k. However, I also expect steep competition from AMD at that point, so a socket switch may be preferable.
The approximate cost of this build was $1,800.
Please let me know if you have any questions or thoughts in the comments below.
Resources:
https://www.intel.com/content/www/us/en/products/sku/241063/intel-core-ultra-7-processor-265k-30m-cache-up-to-5-50-ghz/specifications.html
https://www.intel.com/content/www/us/en/content-details/852792/intel-xmp-3-0-memory-for-intel-core-processors-datasheet.html
https://faqs.noctua.at/en/support/solutions/articles/101000515603
https://gamersnexus.net/cases/lian-li-lancool-217-case-review-thermal-benchmarks-noise-cable-management
Notice: As an Amazon Assosciate I earn from qualifying purchases.