News Intel Core Ultra 9 285K beats the 14900K by 13% in leaked Cinebench R23 multi-core benchmark — Ryzen 9 9950X still leads the pack by 4%

[Admin]

Intel's Core Ultra 9 285K leads the i9-14900K in multi core performance despite lacking Hyperthreading but might lose to the Ryzen 9 9950X.

Intel Core Ultra 9 285K beats the 14900K by 13% in leaked Cinebench R23 multi-core benchmark — Ryzen 9 9950X still leads the pack by 4% : Read more

 

[bit_user]

A screenshot of the CPU's benchmark results has leaked via a hardware enthusiast on Facebook, where the processor is seen sipping up to 250W of power but that is offset by its performance.

Ugh. The only other place I see using this construction is WCCFTech. They do not seem to understand that a sip is a small drink, because they habitually talk about how much power something "sips", no matter how much it is. It's basically the smallest amount you can drink of something. It's distinguished for other drinking words, like "drinks", "gulps", "guzzles", "chugs", specifically by its smallness!

Yes, it can be used as an ironic reference to an unusually large amount, but 250W isn't (any longer) unusually large power consumption, for a desktop CPU. It's also not clear the author is being ironic. If you do this habitually, then it loses its potential for irony, as well as its connotation for being small.

There are so many other descriptive words you could use: gulps, guzzles, chugs, drinks, burns, dissipates, consumes, ...and probably half a dozen more, if I'd bother to pop open a thesaurus.

This puts it 13% faster than the Core i9-14900K; an impressive outcome as Arrow Lake skips out on Hyperthreading.

Yeah, but rendering is heavy on floating-point, which is where Skymont made huge strides. Hyperthreading rarely ever adds more than 20-30%, at best, yet if you look at the claimed floating-point improvements in Skymont, Intel claims a 68% IPC improvement vs. Crestmont, which is already a little better than Gracemont.

​

So, if you consider 0.68 * 16 works out to the equivalent of 10.88 additional Gracemont cores (assuming everything scaled linearly, which it obviously dosen't. Then, consider the worst-case impact of no hyperthreading, which works out to about 23.1% less performance from the 8 P-cores. So, it's basically as if you're trading the IPC of 2 P-cores for an additional 10 E-cores. Obviously, that's a win.

It is pertinent to mention that Cinebench is a synthetic benchmark

It benchmarks an actual production renderer, which is incorporated into Cinema 4D. The point of it is to enable 3D artists to measure how fast different machines will do rendering of their creations. That makes it about as synthetic as a canned benchmark of a game. Both are meant to predict how fast the app will run, but they're pre-scripted to streamline the job of making comparisons.

It's a pretty good proxy for anything fp-heavy that's highly-multithreaded.

 

[TheHerald]

Ugh. The only other place I see using this construction is WCCFTech. They do not seem to understand that a sip is a small drink, because they habitually talk about how much power something "sips", no matter how much it is. It's basically the smallest amount you can drink of something. It's distinguished for other drinking words, like "drinks", "gulps", "guzzles", "chugs", specifically by its smallness!

Yes, it can be used as an ironic reference to an unusually large amount, but 250W isn't (any longer) unusually large power consumption, for a desktop CPU. It's also not clear the author is being ironic. If you do this habitually, then it loses its potential for irony, as well as its connotation for being small.

There are so many other descriptive words you could use: gulps, guzzles, chugs, drinks, burns, dissipates, consumes, ...and probably half a dozen more, if I'd bother to pop open a thesaurus.


Yeah, but rendering is heavy on floating-point, which is where Skymont made huge strides. Hyperthreading rarely ever adds more than 20-30%, at best, yet if you look at the claimed floating-point improvements in Skymont, Intel claims a 68% IPC improvement vs. Crestmont, which is already a little better than Gracemont.

​

​

So, if you consider 0.68 * 16 works out to the equivalent of 10.88 additional Gracemont cores (assuming everything scaled linearly, which it obviously dosen't. Then, consider the worst-case impact of no hyperthreading, which works out to about 23.1% less performance from the 8 P-cores. So, it's basically as if you're trading the IPC of 2 P-cores for an additional 10 E-cores. Obviously, that's a win.


It benchmarks an actual production renderer, which is incorporated into Cinema 4D. The point of it is to enable 3D artists to measure how fast different machines will do rendering of their creations. That makes it about as synthetic as a canned benchmark of a game. Both are meant to predict how fast the app will run, but they're pre-scripted to streamline the job of making comparisons.

It's a pretty good proxy for anything fp-heavy that's highly-multithreaded.

HT off is around a 12-14% drop in performance on 14900k in CBR23 and 24 respectively.

 

[TheHerald]

If there are real numbers, 285k is ~= 14900k. It doesn't seem like it beats it. Got the same score on a stock 14900k roughly at the same power draw

 

[helper800]

Seems good enough for the desktop sector to keep Intel from croaking or being consumed whole by another company for another year or so. I cannot wait for benchmarks.

 

[bit_user]

HT off is around a 12-14% drop in performance on 14900k in CBR23 and 24 respectively.

Thanks. That's a useful data point. I was trying to be generous to HT, so I figured I was probably overshooting.

BTW, what's your take on that supposed peak voltage of 1.336 V? Does that suggest undervolting, or is it typical?

​

 

[bit_user]

Seems good enough for the desktop sector to keep Intel from croaking or being consumed whole by another company for another year or so.

I think Intel needs to do more than just cranking out another CPU. What they need is to repeat the sort of wins they got in Gen 12 or Gen 13. They need something that will compel lots of people to upgrade and to pick them vs. other options. The latter point is specifically relevant, considering some of their recent reputational damage.

In theory, it seems like it shouldn't be that hard to deliver big gains, considering Arrow Lake is two full node jumps beyond Raptor Lake.

 

[TheHerald]

Thanks. That's a useful data point. I was trying to be generous to HT, so I figured I was probably overshooting.

BTW, what's your take on that supposed peak voltage of 1.336 V? Does that suggest undervolting, or is it typical?

​

​

If it was 14th gen that would be UV. Since HT is off it is possible that he can do 5.5 at 1.336. But you can't tell much from the screenshot, I don't think the cpu was pulling 1.336v during the cbr run, if it did the power draw would have been a lot higher, probably exceeding 300 watts.

There is no way he is doing 45k at 250w without undervolting anyways. Plus unless he shows vid and vcore you cannot trust the power draw numbers. If you don't configure AC / DC lls properly your power reporting will be off. I can have hwinfo showing 50 watts while pulling 300. Vid and vcore have to be matching both under load and at idle in order for the power draw to be accurate.

 

[TheHerald]

I think Intel needs to do more than just cranking out another CPU. What they need is to repeat the sort of wins they got in Gen 12 or Gen 13. They need something that will compel lots of people to upgrade and to pick them vs. other options. The latter point is specifically relevant, considering some of their recent reputational damage.

In theory, it seems like it shouldn't be that hard to deliver big gains, considering Arrow Lake is two full node jumps beyond Raptor Lake.

The only reason to consider arrow over existing cpus is the removal of HT. I always hated it. Just makes some app behave more erratically (mostly games) and increases heat concentration. It's tempting me, but the performance uplifts over rpl are approaching 0, so I'll probably skip.

 

[bit_user]

The only reason to consider arrow over existing cpus is the removal of HT. I always hated it. Just makes some app behave more erratically (mostly games) and increases heat concentration.

I maintain this is essentially a software problem. However, it's one that can only be changed on a completely different timescale than the current pace of CPU releases. So, I understand that we need to deal with software as it is, not as we wish it to be.

 

[helper800]

The only reason to consider arrow over existing cpus is the removal of HT. I always hated it. Just makes some app behave more erratically (mostly games) and increases heat concentration. It's tempting me, but the performance uplifts over rpl are approaching 0, so I'll probably skip.

I maintain this is essentially a software problem. However, it's one that can only be changed on a completely different timescale than the current pace of CPU releases. So, I understand that we need to deal with software as it is, not as we wish it to be.

I thought there were major performance increases in single threaded over RL with AL even though MT performance is roughly -10 to+5 only?

 

[TheHerald]

I thought there were major performance increases in single threaded over RL with AL even though MT performance is roughly -10 to+5 only?

Major, no. 10-15% IPC which might drop to 5-8% based on the lower clockspeeds. I mean these are meaningless differences anyways. My 14900k was 18% faster than my 12900k in ST, never noticed any difference

 

[Ogotai]

untill its shown that intels cpus are fixed, current and future, no one i know is even looking at intel right now.. maybe in 3 years.. they are all going amd for their upgrade this time...


sorry, but the trust with intel, isnt there..

 

[helper800]

Major, no. 10-15% IPC which might drop to 5-8% based on the lower clockspeeds. I mean these are meaningless differences anyways. My 14900k was 18% faster than my 12900k in ST, never noticed any difference

Well to be fair I do not recall you using your PCs for more production type tasks where doing something 10-20% faster means you can do 10-20% more revenue.

 

[bit_user]

I mean these are meaningless differences anyways. My 14900k was 18% faster than my 12900k in ST, never noticed any difference

I think it's not relevant for most users just to look at gen-on-gen performance differences, since most people will be doing a multi-gen upgrade. In that context, some decent double-digit improvements over multiple consecutive generations will really stack up!

 

[TheHerald]

Well to be fair I do not recall you using your PCs for more production type tasks where doing something 10-20% faster means you can do 10-20% more revenue.

Production tasks will take advantage of MT , not that much ST.

I think it's not relevant for most users just to look at gen-on-gen performance differences, since most people will be doing a multi-gen upgrade. In that context, some decent double-digit improvements over multiple consecutive generations will really stack up!

Yes, but both zen 5 and arrowlake are basically replacing 2 year old cpus. And they look kinda meh.

 

[helper800]

Production tasks will take advantage of MT , not that much ST.

There are plenty of ST production type tasks. Also when we say ST or MT we mean a performance bias for one or another. For instance there are a huge amount of tasks that can only use a portion of a high core count CPUs' cores. Personally, I call tasks more ST if they cannot take advantage of more cores than we can throw at it and MT if it can.

 

[TheHerald]

OC3D released the reviews accidentally, MT performance seems on par with the 9950x, power draw is much lower than the 14900k but still 40w higher than the 9950x. Overall pretty decent, and it looks easy to cool.

Gaming performance on the other hand seems awful, the 14900k is still at the top, the 285k barely beats alderlake


MT and power draw

Spoiler

Gaming

Spoiler

The i5 looks insane though, 50% faster than the 9600x with same power draw.

 

[helper800]

OC3D released the reviews accidentally, MT performance seems on par with the 9950x, power draw is much lower than the 14900k but still 40w higher than the 9950x. Overall pretty decent, and it looks easy to cool.

Gaming performance on the other hand seems awful, the 14900k is still at the top, the 285k barely beats alderlake


MT and power draw

Spoiler

Gaming

Spoiler

The i5 looks insane though, 50% faster than the 9600x with same power draw.

If everything shown by you is true it would seem there might be some optimization issues or buggy BIOS for gaming performance. I would not expect a 245k to beat or equal a 9600x in ST and MT but then lose or be equal to it in gaming, that's just weird. Maybe unoptimized RAM for the new architecture? They skipped 2 process nodes so I would have expected a bit more. Seems that we need to wait for more data.

 

[bit_user]

There are plenty of ST production type tasks. Also when we say ST or MT we mean a performance bias for one or another. For instance there are a huge amount of tasks that can only use a portion of a high core count CPUs' cores. Personally, I call tasks more ST if they cannot take advantage of more cores than we can throw at it and MT if it can.

The way I look at it, single-threaded is strictly looking at single-core performance. We hope that it scales up well, but it's mainly something you do for the sake of micro-architecture analysis and those few things out there which are truly single-threaded.

Beyond that, I draw a distinction between lightly-threaded and heavily-threaded. The former scales very poorly with core count, the latter very well. Most desktop uses fall in one of these two categories, including web browsing.

Web hasn't really been single-threaded in a really long time. The profusion of cores in phone and tablet SoCs created a big incentive for browsers to become heavily-threaded, even if most of the time is spent in just a few threads.

 

[bit_user]

Gaming performance on the other hand seems awful, the 14900k is still at the top, the 285k barely beats alderlake

This is in line with why I think they're doing Bartlett Lake.

 

[bit_user]

that's just weird. Maybe unoptimized RAM for the new architecture? They skipped 2 process nodes so I would have expected a bit more. Seems that we need to wait for more data.

Could be Meteor's chiplet architecture, that seemed to prioritize efficiency on low-intensity tasks above all else.

 

[helper800]

The way I look at it, single-threaded is strictly looking at single-core performance. We hope that it scales up well, but it's mainly something you do for the sake of micro-architecture analysis and those few things out there which are truly single-threaded.

Beyond that, I draw a distinction between lightly-threaded and heavily-threaded. The former scales very poorly with core count, the latter very well. Most desktop uses fall in one of these two categories, including web browsing.

Web hasn't really been single-threaded in a really long time. The profusion of cores in phone and tablet SoCs created a big incentive for browsers to become heavily-threaded, even if most of the time is spent in just a few threads.

I would be a bit more careful with your specific language. Single threaded a lot of times is a misnomer. Let's' say I have a program that at the code level can only address a single core/thread at a time for all of its work. Modern CPU behavior often makes such a program appear to be able to use more than one thread at a time because the CPU will hot potato the task through its cores to make sure its on the highest possible boosting core at any given time. From an observer's perspective using hardware monitoring programs, this will show utilization on more than one core at any given time because of how fast the CPU will swap the work over to the next highest boosting core even though we know that at a code level the program can only address one core. This is all to say that knowing whether or not a program is truly single threaded cannot always be determined by looking at what the the CPU is doing with monitoring programs. To truly know you would have to look at the specific programs code.

Please correct me if I am wrong.

 

[TheHerald]

If everything shown by you is true it would seem there might be some optimization issues or buggy BIOS for gaming performance. I would not expect a 245k to beat or equal a 9600x in ST and MT but then lose or be equal to it in gaming, that's just weird. Maybe unoptimized RAM for the new architecture? They skipped 2 process nodes so I would have expected a bit more. Seems that we need to wait for more data.

Yeah, I think the numbers are way too low for it not to be buggy bios etc, but who the heck knows. Gaming performance not withstanding, the chips look amazing, but gaming performance is a joke. Not that it's bad per se but it's behind the 13600k (lol)