Intel’s Arrow Lake has tremendous memory overclocking potential — as I’ll outline below. I was able to set the world memory overclocking frequency record at an insane DDR5-12666 — and some dead-simple tuning tips can help improve performance even for regular builds. However, the new chip design presents plenty of challenges for both standard and professional overclockers alike. Below, I’ll cover the most important learnings on boosting performance with standard cooling and with the more extreme liquid nitrogen.
Tiles come to desktop PCs
Like it or not, regardless of what you call them, tiles, chiplets, or silicon “chonks,” are the future. Having a single, large monolithic chip die is not a sustainable path, and Intel is at the end of what’s possible on refining its existing process and increasing clock speeds like we saw with the 12th- to 14th-Gen cpus.
Enter Arrow Lake. Unlike AMD’s chiplets, which have a big space between them, Intel’s approach is to arrange the tiles directly next to each other. If you look at it cross-eyed, it looks just like a Raptor Lake chip. I’m not an electrical engineer nor a semiconductor architect, but I assume everything is physically close together so that the die-to-die interconnects are short and fast.
Think of asking your partner in the next room to bring you a sandwich versus yelling at your neighbor three houses down to bring you one. Your partner will bring it faster, and they know what you like, so it will taste better also.
This is similar to how latency works, which is one of the things Arrow Lake could improve on. With tiles, even if you have fast interconnects, they just can’t physically offer the performance of on-die interaction. The memory controller being on a different tile than the cores is the main cause of this latency issue. We’ll come back to this when we talk about liquid nitrogen overclocking.
There are some questionable choices, too. Where is my HyperThreading?! This is outrageous!. Hyperthreading, for the uninitiated, essentially doubles your threads and gives a ~15-30% performance gain vs. “real” threads in heavy workloads. Hyperthreading was always essentially free (for the user) performance; there was never a reason to shut it off.
I think the fact that HyperThreading is missing makes this product worse than if it had it. Again, I’m not an architect, but I would assume this choice was based on cost. Adding complexity adds cost to an already expensive, TSMC-made product, and perhaps a choice was made for a little extra margin vs a little extra performance.
I really hope we don’t keep seeing companies trending toward decisions like these from the top down. If hyperthreading was there, I am positive reviews would generally be much more complimentary, and the chip would easily be a more viable contender in the space, and we could put Raptor Lake fully behind us.
The positive? In times like these, where negativity and greed run rampant, it’s often more constructive to find the good, or at least plaster on a half smile and bear some optimism through your gritted teeth. Lucky enough for Intel (and AMD), we live in a nearly polarized bubble. Many people will never buy an AMD chip, and many people will never buy an Intel chip. You can give them facts and teach them that loyalty does not help you in these cases; it does not matter.
Quick and simple tips for enthusiasts
There is plenty to like, but it will take some time with the platform and proper educating.
With Arrow Lake, the quickest and easiest way to improve performance is to overclock the NGU (uncore) and the D2D (die-to-die) multipliers. On a high end motherboard, you don’t even need to adjust the voltages. The NGU default ratio is 26x, but it can easily do 34x+. The D2D default is 21x, and it can easily do 34x+ without touching anything else. This will net you 2-20% in performance gain, depending on the workload! We haven’t even touched the core clocks yet, but we are getting gains. That is positive! Intel…maybe you can raise the defaults of these values for people who are afraid of the BIOS.
For Arrow Lake, we have now been introduced to DLVR (Digital Linear Voltage Regulator). This is Intel’s new power management tech to “improve power delivery, efficiency, and thermal management in various applications.” It’s an efficient brain inside the CPU. One of the best decisions that Intel made on these CPUs was to allow you to work with DLVR, give it offsets, and allow it to adapt in the ways you want it to, OR, if you are below 10C, totally bypass it.
Bravo to Intel for spending extra time validating both ways of controlling the CPU, as bypass mode is necessary for sub-ambient cooling to easily max out the CPU. This gives you basically any amount of control you desire, for better or for worse. DLVR is easily something they could have forced on us, especially after the 13th and 14th-gen degrade-gate, but they chose not to, and I thank you on behalf of extreme overclocking enthusiasts. Boom, Positive!
Memory Overclocking
The Arrow Lake memory controller is next level. It’s so good, in fact, that it is reaching frequencies high enough to start making Gear 4 (memory controller running at ¼ the RAM) a viable option. Previously, in almost every scenario, it was always best to remain in Gear 2 (memory controller running at ½ the RAM) unless you were going for an E-Peen memory validation WR (Arrow Lake has the DDR5 world record by a mile), as the latency hit was so bad that it didn’t make up for by the gain in bandwidth. Now we are seeing benchmarks around DDR5-8600 in Gear 2 and DDR5-9733 in Gear 4 — basically neck and neck. Flexibility in memory choices and using previous-generation memory is fine. Oops, still positive!
The meh?
Big.little-style CPUs will always be more beholden to scheduling and OS optimization than non-hybrid-core-style CPUs. This means having Microsoft refine some things, which takes time. I would not be shocked if Windows sees a performance boost soon via an update. Some users are experiencing performance uplift from disabling the P cores, but there is not a world that makes sense in. In theory, hybrid designs make sense, but until both sides are primarily doing it, I think there will be many bugs, such as in the scheduler and otherwise.
I also find it odd that Intel has modified the hold-down mechanism, even though it was never officially stated to have been a problem with Raptor Lake CPUs. A third-party replacement hold-down frame has become a necessary staple for overclockers and enthusiasts because of the flexing caused by the older stock Intel design.
Now, as you can see below, some variants of the socket use a shim for the hold-down brackets with a couple of mm of white plastic. This will lower the force put on the cpu pushing into the socket. Is it effective? I have to say, when removing the heatsinks, the spread of thermal paste is much more even than in previous generations. Will it outperform a frame? I would suspect probably not.
How to “Overclock”
Having spent so much time tuning Arrow Lake, I can tell you that conventional overclocking is still there, but before we even get that far, let’s tweak everything else. You may be surprised by how much you can milk out of the system before even touching the core frequencies. As I said previously, NGU overclocking and D2D overclocking will give a solid performance gain, and memory overclocking and tightening are also very beneficial on Arrow. As you can see in the chart below, going from the default stock system to a system with memory and NGU/D2D overclocked, I am seeing a 2-33% gain in synthetic benchmarks.
| Header Cell – Column 0 | 7Zip (Pt.) | Cinebench R23 (Pt.) | PYPrime 2B (Sec.) | YCruncher (Sec.) |
|---|---|---|---|---|
| CPU Stock 6400 Memory | 177340 | 42476 | 13.174 | 17.382 |
| CPU Stock 9600 Gear4 | 199348 | 42408 | 9.988 | 15.51 |
| CPU Stock 8800 Gear2 | 204340 | 42490 | 8.956 | 15.506 |
| CPU Stock NGU/D2D 34x 8800 Gear2 | 207423 | 43179 | 8.789 | 14.739 |
| Fixed OC 5.6 4.9 NGU/D2D 34x 8800 Gear2 | 220854 | 45832 | 8.312 | 13.322 |
Once that framework is laid, I move on to the core overclocking, for which I always prefer a fixed clock rate overclock. As you can see, at 5.6 GHz P-cores and 4.9 GHz E-cores, the fixed clock rate beats even the default turbo 5.7 GHz in single-thread benching. Is this a scheduler or code issue? I don’t know. Gaining 8% in Cinebench and 37% in PYPrime 2B is absolutely wild. It’s a great reward for learning the system.
Stock temps for enthusiasts — Liqmax
The Enermax LiqmaxFlo 420mm dominates this Core Ultra 9 285K. At default fan speeds, which make it nearly silent, I see low-80C temperatures. With a full overclock and a tune of the fan profiles, I am only hitting high 70Cs in the Cinebench R23 stress test. With the integrated fan on the block top, I don’t even bother using a fan on the VRMs for my bench system. If your case can handle the size of the cooler, then I would recommend picking one up.
It is a monster. I used to rely on a custom loop, but for all the hassle, I can just strap this thing on in one minute, be within a couple of degrees of a custom loop, and then put it away when I’m not using it. It just makes more sense. The fans are the quietest of any AIO that I have tried, and the fan cabling makes things so much cleaner than standard-style cables.
The System
ASRock is a partner of mine, so I always use their stuff. In this scenario, it’s the Z890 Taichi OC Formula. Instead of the stand-alone OC Formula branding, it carries the Taichi name as well. I would assume this is how it will be going forward. The Taichi is the mid-high tier, Formula is the OC-focused model, and the Aqua is the water-incorporated bling-bling high-dollar flagship. The OC Formula has many OC features, buttons aplenty, and performance to match the name.
The first thing you notice is the Memory Shield metallic sticker. ASRock states, “ASRock’s exclusive Memory OC Shield is a patented feature engineered to reduce electromagnetic interference (EMI) that could affect memory overclocking. By shielding memory modules from EMI noise and optimizing the layout, the Memory OC Shield enhances stability and reliability during high-frequency operations, enabling users to push their systems to the maximum potential.”
To me, this reads like early 2000s HiFi / Monster Cable advertisements or those stickers they make for cell phones that “boost your signal by two bars.” I have no way to measure EMI in my basement, so we will just assume it’s not all fluff. If I get bored enough this winter, maybe I will try max memory overclocking with the sticker and then peel it off; maybe not. At least we are getting some innovation. The power delivery and memory trace routing, as always, are top-notch, and the motherboard won’t be the limiting factor in any of your CPU or memory overclocking endeavors. Sorry, my board is covered in black liquid electrical tape, and I don’t use any heatsinks on it.
I only run G.Skill memory. Their support of the entire OC and enthusiast community is unmatched. If you like overclocking, I can’t think of a better option. They provided their latest and greatest memory, the Trident Z5 CK Series Overclocked DDR5 CU-DIMM with Clock Driver.
What is a CU-DIMM, and what is CK? “The Trident Z5 CK and Trident Z5 CK RGB series are built on the new CU-DIMM standard, which introduces a built-in clock driver (CKD) chip on the memory module. Designed to strengthen signals between the CPU and memory IC chips, the CKD helps in improving stability in high-speed memory operations.” This allows the memory to run crazy speeds on Arrow Lake easily. This set here is DDR5-9600 — Yes, that is not a mistake.
Power Delivery
I’m running the Enermax Revolution DFX 1650W. I know, with its sizable 112.5Amps on the 12V rail, it could probably handle three of these systems at the same time if it had more 24 and 8-pin connections. Having unlimited power and never worrying whether it is an Arrow Lake or Threadripper makes it nice to just have this sitting here.
It’s under $299 for a 1650W Gold PSU, which is wild to me. They now have beautiful individually sleeved cables that don’t kink and are malleable enough to make smooth bends, and they are fully modular. No ugly zip ties. RGB is controllable and can be shut off if you’re a hater.
Cooling — LN2 issues, SOC die cold bug, memory limits on LN2
For the greater part of three months, I have been hammering on Arrow Lake with liquid nitrogen, which is not easy. With tiles usually come cold issues; for Arrow Lake, that cold issue is the SOC tile. It absolutely hates the cold. Around -100C, the system will no longer turn on as, for the sake of chip longevity, the boot-up voltages are super low. Once you are booted up, most of the time, around -160C to -180C seems to be the coldest it will go in most circumstances before it throws errors.
Throw in DLVR vs. Bypass mode, and there are a lot of new things to play with. In general, bypass mode seems to be the best. The thing about bypass mode is that the chip can no longer feed the P-cores, E-cores, and Ring different voltages. They will all be in sync with the highest of the three applied, which you would think is not optimal. In a perfect world, P-cores would like 1.5-1.65V, E-Cores 1.3-1.4V, and the Ring needs little more than stock volts. In bypass mode, all three would get 1.65V, which is pretty spicy.
With these new issues and tile growing pains, liquid nitrogen results are hit or miss. A strong part will do up to 6.9-7 GHz in threaded benchmarks, trailing Raptor Lake’s peak of 7.6 GHz by a decent margin. It does make up for some frequency deficit by having E-Cores that clock higher and are much more efficient.
The missing hyperthreading is a devastating blow to world record and global rankings on HWBot, where Arrow Lake’s 24 cores and 24 threads are in the same category as Raptor Lake’s 24 cores and 32 threads. Regardless, I completed over 30X category gold cups for Core Ultra 9 285k rankings. Not a bad effort.
Now, back to the memory controller. Arrow Lake can pump insane memory frequency on air and liquid nitrogen cooling. I managed to break the DDR5 frequency world record at DDR5-12666! Using the GSKILL Trident and pouring liquid nitrogen on them (well, maybe it’s not that easy), and after trying a couple of different CPUs, it all fell into place. Sorry about the football game in the reflection on the memory DIMM below; it was an accident, I swear.
For the record, the memory wanted to be as cold as possible, but the CPU itself only needed to be around -80C. The board being insulated keeps me relatively safe as far as water and shorting things are concerned. It was really fun benching this.
Experiments with drilling the IHS
In my first day’s experience with Arrow Lake, I realized that the SoC die was the source of the cold bug problem with liquid nitrogen. The SoC die is conveniently located at the bottom of the CPU, so I had the idea of grinding away the heat spreader as much as I could to see if there was any effectual cold tolerance gained. It’s far from a perfect solution, but when all you have is a Dremel and too much ambition, paired with a disregard for the hardware at hand, it was try to improve things or do nothing.
This modification keeps the liquid nitrogen pot from directly touching the portion of the integrated heat spreader (IHS) that has the SoC tile underneath. Naturally, the IHS still touches the tile, and the solder below obviously still connects everything, but we did gain about 50C of extra cold tolerance on the part! I hope we see some innovative OC guys like Der8aur or Elmor run with this and make something that works 100%, as the theory has proven to benefit.
Arrow Lake, on the whole, has read as seemingly pretty negative. While I can’t argue with all of the points, there are a few I would like to make. I can’t fault reviewers for not spending three months finding performance gains on this platform, especially when most users will never even enter their BIOS.
This is Intel’s most advanced mainstream processor by far. There are so many features and options that we are still finding performance gains daily. In fact, ASRock still sends me BIOS updates every day, and they aren’t for bugs; they are for performance and memory tuning. I think there are some things Intel could have stuck their neck out a bit more for and pushed. They could have been a bit more aggressive and have higher default NGU and D2D speeds, but I’m sure there are kid gloves on after recent degrading issues on 13th- and 14th-Gen, so playing it safe might be the call. If you like overclocking and tuning, it really is a fun platform with plenty left on the table for enthusiasts to find.
Writers note: I received the CPU and motherboard from ASRock, Memory from G.Skill, and PSU from Enermax.
🔴 Bonus
If you hold the CPU just right, you may notice what looks to be damage or some other non-uniformity. But don’t worry, that is normal. I was informed that those pertain to the PCIe and DMI. It is not an error or pad damage.