AMD Ryzen 9 9950X vs Intel Core Ultra 9 285K Faceoff — it isn't even close

[TerryLaze]

Sorry to say, but you're not entirely correct either.

1.- The reason why you would prefer a Z board for K SKUs is because you can't overclock the CPU in any other and only K SKUs can be overclocked. This is by design with Intel and AMD has a more relaxed approach, since you can OC most, if not all, their CPUs in B and X series chipsets. This is leaving outside the use of XMP/EXPO from either as well, where Intel limits this in H chipsets and AMD, I think, does in the A series. So, in the context of this article, if you want to make this correction, then you also need to give the additional context of why AMD would be the superior option as it does not lock you onto the X chipsets for CPU overclocking.

Requires implies that it's required.....
If you prefer to do something else then nobody is stopping you.

But yeah I do agree with you that the only reason they do it is to be able to say that they used default settings with the CPU being over provisioned power by 50% ,only a top end overclocking mobo will be able to withstand that.

If you want to overclock your intel CPU in the same degree that amd allows then all you need is a mobo that allows lifting the power limits and that's possible on pretty much any mobo, that's what PBO is and any manual overclocking on amd will be worse than what PBO can do.

3.- So you want to give Intel points for forcing you to buy more expensive RAM so it can close the gap on performance with AMD? This is not even including stability concerns from using XMP and anecdotal evidence of "I have DDR5-12000 running with no issues in my system!". JEDEC is what should always be evaluated against, which Tom's has always done. XMP and EXPO are forms of overclocking and that is never guaranteed. Not even going to talk about CUDIMM and their price and availability. On the other hand, why not give credit to AMD for supporting "proper" ECC with their CPUs unlike Intel? This is something well documented all Zen CPUs (so far) have proper ECC support and it's up to the motherboard vendors if they want to include RDIMMs, but Intel doesn't even have that option for consumers.

He was talking about native memory frequency, the 6400 he said in the post is the official number and not xmp/expo.

 

[gethen]

Sorry to say, but you're not entirely correct either.

1.- The reason why you would prefer a Z board for K SKUs is because you can't overclock the CPU in any other and only K SKUs can be overclocked. This is by design with Intel and AMD has a more relaxed approach, since you can OC most, if not all, their CPUs in B and X series chipsets. This is leaving outside the use of XMP/EXPO from either as well, where Intel limits this in H chipsets and AMD, I think, does in the A series. So, in the context of this article, if you want to make this correction, then you also need to give the additional context of why AMD would be the superior option as it does not lock you onto the X chipsets for CPU overclocking.
2.- Nothing to say and you're correct there.
3.- So you want to give Intel points for forcing you to buy more expensive RAM so it can close the gap on performance with AMD? This is not even including stability concerns from using XMP and anecdotal evidence of "I have DDR5-12000 running with no issues in my system!". JEDEC is what should always be evaluated against, which Tom's has always done. XMP and EXPO are forms of overclocking and that is never guaranteed. Not even going to talk about CUDIMM and their price and availability. On the other hand, why not give credit to AMD for supporting "proper" ECC with their CPUs unlike Intel? This is something well documented all Zen CPUs (so far) have proper ECC support and it's up to the motherboard vendors if they want to include RDIMMs, but Intel doesn't even have that option for consumers.
4.- Too subjective, but it's a point to consider. I won't read the article spoting "language bias" to contradict or support your assertion.
5.- If you expand the test suit, then AMD would actually take more wins than what you think. "It doesn't matter if you win by an inch or a mile, winning's winning". To quote someone, heh. Still, anything over 3% is definitely outside of margin of error / small variance and can be considered a "proper" win. You can make that analysis individually per test and then tally up. You don't need averages.
6.- New to how publications work? As long as the title is not factually incorrect (which is liable for a lawsuit), then it's fair game. I personally do not like it, but that's a "we" problem unless you can get "click bait titles" into regulation as an offence and force it to be "outlawed". You latter part is your own conclusion and interpretation of the data.

Also, this is not a review.

Regards.

Thank you for the considered and detailed feedback, much appreciated! Keeping it short...
1. I hear you, a good and valid point made, with my only response being that the article on this issue made no reference to overclocking, it simply stated "...the new LGA1851 socket exclusively with Z890 chipset...", although I see they have updated it to include the references to B860 and H810 options now.
3. Interesting point, and I'd only challenge you on your use of the word "force" about Intel and using CUDIMM. I may be wrong, but I think it's actually the JEDEC standard that requires DDR5 DIMM memory running at 6400MHz or higher to include a CKD? Not Intel. This is an industry standard AMD chooses not to adopt, and they're otherwise are a very active contributor to JEDEC standards. Prices for new technology are valways higher to begin with. I'm quite sure we'll all be using CUDIMM in 3-5 years time, bought at reasonable prices by then.

My last point to declare my own bias... I'm writing this to you on a machine with an Ultra 7 265K CPU and Radeon 9070XT 16Gb GPU, and this combo is a little rocket for my own work (not much gaming). I just wish AMD would hurry up and release ROCm for 9000 series graphics cards for Windows and Linux, they are well behind NVIDIA/CUDA for inference stacks.

 

[Paul Alcorn]

<snip>

Thanks for the feedback.

The article has always mentioned that the 285K supports the other two chipsets in the specifications section (Original version here). The part you quoted is from the pricing section, where the other chipsets were not mentioned because they don't make sense for this chip (it would be unwise to buy an overclocking chip and not an overclocking-capable motherboard). However, I have now added a mention of the other chipsets in the pricing area, too, as others may only skim the article and miss the first mention. We tend to avoid this due to repeating ourselves over and over in an article.

Thanks for spotting the memory capacity mistake; that is corrected.

In regard to the perception that the commentary exposes bias, please be sure to quote the full context of a statement in the article when referencing it, lest it might also be misinterpreted by others:

" In A Plague Tale: Requiem, the AMD Ryzen 9 9950X dominates, averaging a 10.6% lead over its Intel counterpart. In Cyberpunk 2077, Intel reverses the trend, averaging a 9.9% better result than the Ryzen." [Emphasis added]

My comprehension of this statement is that the 9950X 'dominates' in A Plague Tale, but Intel reverses that and dominates in Cyberpunk 2077. That's the intent/meaning of "Intel reverses."

In terms of whether or not the title supports the results, I have commented on that matter here:

It isn't even close.

We appreciate the spirited feedback and the spot on the memory error. Thanks!

 

[-Fran-]

He was talking about native memory frequency, the 6400 he said in the post is the official number and not xmp/expo.

Yes, I know Intel now supports JEDEC DDR5-6400 with the corresponding timings, but my point was that around high transfer rates and, in particular, CUDIMM making zero sense for JEDEC at 6400 MT/s, since I'd argue 100% of kits out there can do it (new ones; segmentation be damned), but they just don't allow it. Remember CUDIMMs have their own clock generator so they don't depend on the signal clock (IMC) and trace integrity (caveats apply).

AMD, for instance, can somewhat reliably go up to 7200MT/s and it's well documented. Performance doesn't vary wildly because of the IF clocks still being a bottleneck for the clusters (CCDs), so for Zen CPUs, the IF clocks and the effective ratio matters way more than pure RAM transfer rates (UCLK, MCLK and FCLK being 2:2:3 as the "golden" ratio, with some exceptions). This is more than likely why CUDIMM has no official support with AMD: it makes little to no sense with the current IMC (or I/O die process), but they've stated "you can try" without being commital, but I get it: official support or bust. No idea if AMD would encourage motherboard vendors to QVL CUDIMMs; who knows, perhaps some have done it?

Buildzoid has great videos on the topic and it's quite comprehensive for "fine tuning" almost any Zen CPU. AM4 has coupled UCLK and MCLK; AM5 with Zen4 decoupled them, but you still want to do 1:1 with them as much as possible. The Intel equivalent is the Gear 1 config being 1:1 MCLK and UCLK and Gear 2 being 1:2 MCLK and UCLK. No idea how impactful it is with their uArchs over time, but I think the idea is to always try to use Gear 1. Which reminds me! XMP in B chipsets comes with the caveat you can't change the Gear setting, as that is a CPU OC feature, so that caveat also exists.

And what about RDIMMs for this class of CPU that you can argue qualify as a "work CPU"? What do you think about that in particular?

Thank you for the considered and detailed feedback, much appreciated! Keeping it short...
1. I hear you, a good and valid point made, with my only response being that the article on this issue made no reference to overclocking, it simply stated "...the new LGA1851 socket exclusively with Z890 chipset...", although I see they have updated it to include the references to B860 and H810 options now.
3. Interesting point, and I'd only challenge you on your use of the word "force" about Intel and using CUDIMM. I may be wrong, but I think it's actually the JEDEC standard that requires DDR5 DIMM memory running at 6400MHz or higher to include a CKD? Not Intel. This is an industry standard AMD chooses not to adopt, and they're otherwise are a very active contributor to JEDEC standards. Prices for new technology are valways higher to begin with. I'm quite sure we'll all be using CUDIMM in 3-5 years time, bought at reasonable prices by then.

My last point to declare my own bias... I'm writing this to you on a machine with an Ultra 7 265K CPU and Radeon 9070XT 16Gb GPU, and this combo is a little rocket for my own work (not much gaming). I just wish AMD would hurry up and release ROCm for 9000 series graphics cards for Windows and Linux, they are well behind NVIDIA/CUDA for inference stacks.

No worries. It's always good to have scrutiny to both read about it and have a proper discussion. So, that being said:

1.- There is mention to Overclocking and even has it's own evaluation where Intel wins. With the new provided information, do you think that is correct? Just having "headroom" nowadays is rather poor in terms of criteria to determine "best overclocker", but I agree with the veredict, because more chipsets supporting equal level of OC as premium boards is a value proposition and not a technical superior one, but it is intertwined, I'd say. Having "better RAM OC" is a rather moot point when all that extra bandwidth is not utilised by all apps. Similar vein to VCache's caveat. For Intel it does give it that extra oomph it needs, but at a great extra cost which AMD doesn't really need as long as you know how to set up the RAM (read above's reply to Terry). My 9950X3D is running 6400MT/s CL32 with a 2:2:3 ratio (IF at 2133Mhz, which is stupid hard to get stable, but mine is!) and my overall platform performance is several points upwards from any "stock" and even "PBO'd" 9950X or X3D. I'm sure most Intel enthusiasts that tweak can say the same, but you need to compare baselines. Again: overclocking is never guaranteed. You need to compare baselines. Both AMD and Intel are super scummy for """"sugesting"""" (sending fast memory) for their CPU reviews. I'm glad Toms is still maintaining the line and do proper "baseline performance" analysis, which is what most people getting OEM PCs would see (it's much more representative).

3.- I used "force" in that context to explain that if you use JEDEC, at best you'll be close, but not surpass consistently the 9950X. Paul has already provided the links to the original articles and you can check their original review piece of the 285K and see how it stacks with the JEDEC clocks vs AMD: https://web.archive.org/web/2025072...nents/cpus/intel-core-ultra-9-285k-cpu-review. TL;DR: 9950X still wins in productivity by a good margin. But the eternal caveat in productivity apps is: make sure your important workloads perform the best in the target CPU. I use software that actually leverages AVX512, so AMD just destroys Intel here (word aptly and intentionally used, yes), because Intel just shot themselves on the foot there, rather spectacularly. Cutting one of the ISA elements they've been pushing for so many generations was really dumb to me. They just gave AMD that niche. And as to why AMD is not fussed about CUDIMM, well, I gave that answer to Terry as well (kind of). TL;DR: their IMC can't really make good use of super high clocks and see better performance than lower clocked kits due to the ratios going to the crapper.

Regards.

 

[yzonker]

Thanks for the feedback.

Yes, performance in various categories can be close, but for the purposes of a faceoff, we declare winners in various categories, and then those are tallied to declare an overall winner.

The AMD chip won four categories outright versus the 285K winning one.
If we count the tie, the AMD chip has five categories compared to the Intel chips' two.

I think it is fair to say that both a 4-to-1 and a 5-to-2 advantage would be classified as not being close.

Give me a break. That's just a lame excuse for your clickbait headline. What if the results were this,

Category 1: 1% win AMD
Category 2: 0.5% win AMD
Category 3: 0.1% win AMD
Category 4: 0.3% win Intel

So even though that is well within run to run variance, which probably some of those categories in the article are, that would still constitute "not even close"!??!?!

You need to retract that headline. It's pure BS.