Rendering Performance
In all of our graphs to follow, the bars are arranged from the fastest MT/s speed down, so 8800MT/s is on top, with 7200MT/s on the bottom, and this arrangement is locked in independent of the performance shown; they are not arranged by result, but rather the DDR5 memory speed.
Cinebench 2026
In Cinebench 2026, there are some slight differences, and we once again see the 7600MT/s UDIMM at CL38 remain the top performer at 1% above the 7200MT/s performance and 1% above the 8400MT/s and 8800MT/s CUDIMM performance. In fact, the 8400MT/s and 8800MT/s CUDIMM are just slightly behind the 7200MT/s UDIMM due to the higher timings.
Corona Benchmark
In the Corona 1.3 Benchmark, the 7600MT/s UDIMM at CL38 is again providing the fastest result with the most Rays a Second, and is 2% faster than with the 8800MT/s CUDIMM at CL43. Improving upon the 7200MT/s performance, the 7600MT/s provides 1% more performance.
7-Zip
In 7-Zip’s compression test, we actually saw a pretty good result with the 7600MT/s UDIMM at CL38. It was 2% faster than the 7200MT/s speed, and just as fast compared to the 8400MT/s and 8800MT/s CUDIMM, which matched 7200MT/s performance.
There was one instance where the higher bandwidth of the 8800MT/s CUDIMM at CL42 shone in our testing, and that was here in the 7-Zip Decompression test. The 8800MT/s CUDIMM yielded the highest result at 201.6, which was 1% faster than the lowest result and was the fastest. Otherwise, the 7600MT/s UDIMM at CL38 is right on its heels, and honestly can be considered similar performance.
Blender Benchmark
In Blender, we see much of the same; there is an advantage with the 7600MT/s UDIMM at CL38 over the 7200MT/s UDIMM at CL38 by 1%. However, the 8400MT/s and 8800MT/s CUDIMM at CL40 and CL42 are slower than the 7200MT/s at CL38.
The same pattern is experienced in the Junkshop scene in Blender; the 7600MT/s UDIMM CL38 memory is 1% faster than the 7200MT/s UDIMM CL38. However, 8400MT/s and 8800MT/s CUDIMM CL40 and CL48 memory is slower than 7200MT/s CL38.
In the Classroom scene, the groupings are tighter, but the same pattern remains true.
HandBrake
There is an actual difference in video transcoding time between the memory speeds. The slowest time to render was actually the 8800MT/s at CL42, taking 7.23 minutes to render. The 8400MT/s or 7200MT/s with lower timings was next. However, the 7600MT/s with CL38 took the cake and rendered in the least amount of time at 7.18 minutes.
V-Ray 6 Benchmark
In V-Ray 6, once again, the 7600MT/s UDIMM at CL38 was the fastest and improved upon 7200MT/s by 1.1%, while the 8400MT/s and 8800MT/s CUDIMM with higher timings lagged behind.
Discussion (12 replies)
Join Discussion →Pretty much what I expected to see to be honest.
Try buying some???
Need a loan.....LoL
Would have been interesting to have some "run of the mill" 6000MT/s RAM like alot of folks use (because its affordable) to compare as a baseline......?
Some serious hot deals on Arrow plus combos right now. AIOs, Lego Batman, PSUs, boards. Also, 270K is below MSRP and the second best seller on Amazon today. Been a long time since a Intel CPU cracked the top 5.
Just wanted to point out that actual overclockers may take offense at there being no information available on the values for NGU, D2D and Ring. If those were kept at default then the memory overclocks are almost pointless due to there being a bottleneck in the flow of data between the CPU and RAM sticks. For anything over 8000 MT/s, NGU multiplier needs to be 34, D2D should ideally be 35 or higher and ring from 42 to 46 would be ideal. It's unfortunate that Intel did not make it so that these values are optimized automatically as RAM speed is increased.
My own 245KF was able to do NGU 34, D2D 35 (latest BIOS kept it at 30 despite setting a higher value) and ring 42. Sadly, the CPU is out of the socket of the mobo as I was trying to sell it so any benchmark data I have is inaccessible at the moment unfortunately.
Regarding the RAM timings, the sticks should be able to 7200C34 and 7600C36. The 8000+ timings seem to be ok.
Just did a web search on this, have yet to find a review that did the NGU/D2D/Ring overclocking, ect. Currently have my 265KF at 70.1ns latency. Wondering if 270K can get it any lower, or if I should even waste the $$?
The only low latency Arrow Lake champ right now is 250K/KF.
You may gain ~10 fps plus future APO/iBOT game/application support with 250KF.
There's unfortunately a rather glaring problem with the testing here which is comparing a kit using 16GB modules to one using 24GB. Subtimings are important and almost every single higher capacity module has higher cell refresh timings than an equivalent speed/primary latency lower capacity. That is absolutely the case here which throws off the results and in turn the conclusion. If one is trying to compare speed and primary timings the kit capacities really need to be the same.
What kinda latency are you getting with your setup??
y-cruncher score comparison 6000C42 vs. 8200C38 (NGU is 32 or 34, D2D is above 30 and Ring is 42)
6000C42:
8200C38:
Another multicore scaling benchmark here: https://forums.anandtech.com/threads/arrow-lake-builders-thread.2622775/post-41610100
I haven't really done a lot of optimization as real life kept getting in the way. I did replace a 265K with the 270K Plus and they both performed within margin of error of each other when it came to latency using the same settings. I imagine best case scenario for latency with ARL is going to be based on bins more than SKU.
From what little tweaking I've done NGU seems to make a lot less difference than D2D. Subtimings on memory can also make a pretty big difference. The board I'm using has had some weird issues with DRAM timings so once I found stable numbers I just left it, but this is high up on my list for when I get time.
Skatterbencher puts out my favorite writeups (and also has videos) for CPU tweaking: https://skatterbencher.com/2025/06/18/skatterbencher-87-core-ultra-7-265k-overclocked-to-5700-mhz/
They've done more recent ones with the 270/250 as well.
Ok, upgraded to a 270k, and another new motherboard, since I bent socket pins (Rookie move, first time ever...) Proc clocks almost exactly like my 265KF, 56P core, 49 E core (265KF was 56/50. 41 cache, 33 NGU/D2D, Running 48gb Kingston Renegade 8400 @ 38-49-49-56 timings, same 71.3 ms latency as with 265KF as well. Seems to run cooler, and extra 4 cores. XTU benchmark went up nicely, as well as my 3DMark bench's. Have yet to run PCMark bench yet. All in all, runs well in a dead end platform, should keep me happy for @ least 4 yrs i'd hope.