14900 KS - the fastest Intel processor ever just launched

[BoggledBeagle]

the word "fastest" without any specification implies that it's the fastest in everything

No, it does not, people do generalise and simplify when they speak. If somebody would want to say that something is "fastest in everything", it is such an extraordinary thing that he would explicitly say that and even stress that.

BTW I added the 14600K result just to point out, that the strong single thread performance is not just a result of this extremely clocked CPU, but a general characteristics of even a cheap reasonably clocked mainstream CPU of the same kind (I believe that 14600K uses the same chip as 14900KS).

 

[nomdeplume]

I'm not saying that you're wrong, but some criticism actually has some ground. Latest gen i7s and i9s are extremely inefficient. Stating this obvious fact, while unnecessary, doesn't constitute AMD fanboyism.

i7 were this inefficient powerhouse for especially tough work. When they turned into a status symbol prebuilt Platinum class purchase suddenly we needed i9. Now that i things are so ubiquitous in Western society they are rebranding and will probably add the next tier of SKU equating to i11.

Stating this obvious fact, while unnecessary, doesn't stifle blind consumerism. Blind because it is for the best those who will associate with the fastest/best/elitist are not seeing a very broad picture. Toxicity bumping up against walls in the dark is unavoidable side effect. Not hard to separate pleasant hobbyists into a category that never reaches fanatic improvisations. A telling sign is being just as interested in building up and testing a 15 year old system for new possibilities as seeing how the newest performs.

 

[dgianstefani]

No, it does not, people do generalise and simplify when they speak. If somebody would want to say that something is "fastest in everything", it is such an extraordinary thing that he would explicitly say that and even stress that.

BTW I added the 14600K result just to point out, that the strong single thread performance is not just a result of this extremely clocked cpu, but a general charasteristics of even a cheap reasonable clocked mainstream CPU of the same kind (I believe that 14600K uses the same chip as 14900KS).

Plus synthetic single thread tests directly predict performance in games. The fact the X3D chips do so well in singleplayer games is because of the cache, which helps some games with bloated code (e.g. Tarkov, the game I play which is why I have an X3D), and not others. E.g Counter strike or other online multiplayer games where X3D are slower due to lower clocks than normal Zen.

I agree with you that the qualifier "in everything" would be used, and the general term "Fastest" is correct lingually, because, with few exceptions, it's fastest in the vast majority.

 

[AusWolf]

Plus synthetic single thread tests directly predict performance in games. The fact the X3D chips do so well in singleplayer games is because of the cache, which helps some games with bloated code (e.g. Tarkov, the game I play which is why I have an X3D), and not others. E.g Counter strike or other online multiplayer games where X3D are slower due to lower clocks than normal Zen.

They do not. Cinebench doesn't give a damn about cache, while most games do. Sure, some games don't, but they're the exception, not the rule. That's why I'm saying that no CPU is the be-all-end-all of gaming. Every scenario is different, and if you wanted to crown a winner, you could be crowning a multitude of them depending on your use case. If you play CS, then buy a 14900K(S), by all means.

Blaming game code isn't helping, either. If if was really a case of a handful of badly optimised games, while 95% of them favoured clock speed, then X3D wouldn't have a reason to exist, and wouldn't win on the average values in every single review.

I agree with you that the qualifier "in everything" would be used, and the general term "Fastest" is correct lingually, because, with few exceptions, it's fastest in the vast majority.

The Bugatti Chiron is the fastest car - what does that sentence make you think of? Straight line, race track, off road, etc? Honestly, it's a vague, empty, hype sentence that can mean anything and everything. It only conveys any information to those who believe in the presence of "9" and "X" in product names. E-penis measurement and online bragging rights. Nothing more.

 

[dgianstefani]

Maybe, but regardless of what bench you use to measure singlethreaded performance. Cinebench, CPU-Z, web browsing, superpi etc. The results are consistent.

A large cache is helpful because it can be used preferentially over RAM, which in Zen's case, is a higher latency transaction than it is on Intel. As is the core to IO latency, or core to core, for the multi CCD chips, whether this platform generation, or previous ones. Gaming, and other real time tasks, like low latency, and they like high ST performance. Whether that low latency is achieved through faster frequencies, or by fitting things in cache rather than RAM, doesn't really matter. Except you can't fit everything in a ~100 MB cache. Hence why I have to correct people who say RAM speed doesn't really matter on X3D, because it does. I've tested it.

Good luck getting below 60 ns latency in AIDA on Zen though, I've got 57 ns but that took lots of effort.

The point is, ST is always relevant, a large cache is sometimes relevant. Like running out of RAM capacity, it makes no difference whether you have 32 GB or 64 GB of identically clocked RAM for performance, if you're only using ~20 GB of it.

In my experience, online games scale better with frequency, not cache size. Might be something to do with their unpredictability compared to singleplayer games.

A funny thing is the non X3D chips actually have better AIDA RAM latency than the X3D ones, because they clock higher.

I got 55 ns on my 5950X and 58 ns on 5800 X3D.

Raptor Lake with a good memory controller you can go lower than ~45 ns.

 

[AusWolf]

Maybe, but regardless of what bench you use to measure singlethreaded performance. Cinebench, CPU-Z, web browsing, superpi etc. The results are consistent.

Yes - it's consistent across apps that don't use much cache. That's why the 14900K(S) is the productivity king. With a large number of cores at very high clock speeds, it kills everything in apps that need exactly that.

But that's not the case with (most) games that are in general a lot more sensitive for cache size.

Hence why I have to correct people who say RAM speed doesn't really matter on X3D, because it does. I've tested it.

That also depends on your use case. Maybe it does with a 4090, but for me with a 7800 XT, switching between 4800 and 6000 MHz doesn't yield a tenth of an FPS in anything I've tested so far.

The point is, ST is always relevant, a large cache is sometimes relevant.

That's right. So if you've got the money, and you're not afraid of your CPU burning a hole through your motherboard (or you're going with a large AIO or custom loop), then I'm sure the 14900K is awesome. The 7800X3D isn't a winner in my books because it's absolutely faster (it sometimes is, but that's besides the point), but because it performs incredibly well in games relative to its price and power consumption.

 

[rv8000]

Maybe, but regardless of what bench you use to measure singlethreaded performance. Cinebench, CPU-Z, web browsing, superpi etc. The results are consistent.

A large cache is helpful because it can be used preferentially over RAM, which in Zen's case, is a higher latency transaction than it is on Intel. As is the core to IO latency, or core to core, for the multi CCD chips, whether this platform generation, or previous ones. Gaming, and other real time tasks, like low latency, and they like high ST performance. Whether that low latency is achieved through faster frequencies, or by fitting things in cache rather than RAM, doesn't really matter. Except you can't fit everything in a ~100 MB cache. Hence why I have to correct people who say RAM speed doesn't really matter on X3D, because it does. I've tested it.

Good luck getting below 60 ns latency in AIDA on Zen though, I've got 57 ns but that took lots of effort.

The point is, ST is always relevant, a large cache is sometimes relevant. Like running out of RAM capacity, it makes no difference whether you have 32 GB or 64 GB of identically clocked RAM for performance, if you're only using ~20 GB of it.

In my experience, online games scale better with frequency, not cache size. Might be something to do with their unpredictability compared to singleplayer games.

A funny thing is the non X3D chips actually have better AIDA RAM latency than the X3D ones, because they clock higher.

I got 55 ns on my 5950X and 58 ns on 5800 X3D.

Sub 60 on non X3D zen 4 parts is relatively easy to achieve with minimal tuning (IF, c30-c32, and tightened primaries/trfc). Normal result is around 56-58ns on minimum effort tunes/testing, more golden chips/significant tuning can yield 52-53ns.

Having both a 7900X3D (6400/2133) and 14700k (8000), the intel setup actually has significantly worse averages/minimums in cpu limited areas in WoW and GW2 by 10-15 fps.

While certainly off topic, there is a tremendous amount of performance to gain by tuning memory/IF on zen 4 to the point where a low clocking 7900X3D can achieve ~127 points in ST CB24, but such testing is beyond the scope of general CPU reviews. Id wager to guess 7950X could score somewhere in the low to mid 130 range, which is where my 14700k @6ghz w/8000 lands.

 

[dgianstefani]

Sub 60 on non X3D zen 4 parts is relatively easy to achieve with minimal tuning (IF, c30-c32, and tightened primaries/trfc). Normal result is around 56-58ns on minimum effort tunes/testing, more golden chips/significant tuning can yield 52-53ns.

Having both a 7900X3D (6400/2133) and 14700k (8000), the intel setup actually has significantly worse averages/minimums in cpu limited areas in WoW and GW2 by 10-15 fps.

While certainly off topic, there is a tremendous amount of performance to gain by tuning memory/IF on zen 4 to the point where a low clocking 7900X3D can achieve ~127 points in ST CB24, but such testing is beyond the scope of general CPU reviews. Id wager to guess 7950X could score somewhere in the low to mid 130 range, which is where my 14700k @6ghz w/8000 lands.

I mean, you can take a look in the AIDA thread. The vast majority of people are running 60 ns+.

If you're lucky you can get lower than that, but I certainly don't think it's guaranteed.

From the 7800 X3D review.

 

[rv8000]

I mean, you can take a look in the AIDA thread. The vast majority of people are running 60 ns+.

If you're lucky you can get lower than that, but I certainly don't think it's guaranteed.

From the 7800 X3D review.

View attachment 339544

Read more slowly this time.

 

[dgianstefani]

Read more slowly this time.

Nothing wrong with my reading comprehension thanks. I already acknowledged that you can get lower than 60 ns on X3D in the post before yours. I just don't think it's accurate to assume that's the typical case.

A stock 14900KS has 4 ns lower RAM latency than 7800 X3D with identical RAM, and can go to lower than ~45 ns, whereas a maxed out Zen 4 part might get low 50s if you're lucky.

I actually had 55 ns before I decided to go for a BCLK OC instead, which affects RAM/IF too, but 150 MHz extra CPU frequency is more important than 2ns of RAM latency.

 

[rv8000]

Nothing wrong with my reading comprehension thanks. I already acknowledged that you can get lower than 60 ns on X3D in the post before yours. I just don't think it's accurate to assume that's the typical case.

A stock 14900KS has 9 ns lower latency than 7800 X3D with the same RAM, and can go to lower than ~45 ns, whereas a maxed out Zen 4 part might get low 50s if you're lucky.

There was no comment whatsoever claiming what an intel part can do in terms of latency. You incorrectly stated it’s extremely difficult to get below 60ns on Zen 4 which is flat out wrong, not to mention can be done on both non 3dvcache parts and 3dvcache parts.

Showing a screenshot of loose c36 ddr5 6000 proves nothing of the sort.

 

[dgianstefani]

There's also the bandwidth issue, 14900KS is getting 96/84/95 GB/s with identical RAM to the 7800X3D, which gets 60/80/60 GB/s.

There was no comment whatsoever claiming what an intel part can do in terms of latency. You incorrectly stated it’s extremely difficult to get below 60ns on Zen 4 which is flat out wrong, not to mention can be done on both non 3dvcache parts and 3dvcache parts.

Showing a screenshot of loose c36 ddr5 6000 proves nothing of the sort.

What it "proves" is that even at identical RAM timings and settings, Intel chips are tighter in latency and faster in bandwidth. That's the point. A 1:1 RAM comparison, and why we bench CPUs with otherwise identical parts.

If you want to talk tuned setups, Intel can get 120/120/120 GB/s +, still with 5-10 ns lower latency than the fastest Zen.

I do take your point though that it's perfectly possible to get lower than 60 ns on Zen, though I'm not sure how many people are.

The dual CCD Zen chips would get better write bandwidth due to two connections to the IO die, one from each CCD, but that doesn't help with gaming.

 

[AusWolf]

Can I butt in? What does a few ns tighter RAM latency give you? 0.3 FPS in your favourite game? Or 4 more points in Cinebench perhaps?

I have a mid-range GPU which is not affected by system RAM speed at all, so consider this a genuine question.

 

[rv8000]

There's also the bandwidth issue, 14900KS is getting 96/84/95 GB/s with identical RAM to the 7800X3D, which gets 60/80/60 GB/s.

What it "proves" is that even at identical RAM timings and settings, Intel chips are tighter in latency and faster in bandwidth. That's the point. A 1:1 RAM comparison, and why we bench CPUs with otherwise identical parts.

If you want to talk tuned setups, Intel will get 100/100/100 GB/s +, with 5-10 ns lower latency than the fastest Zen.

I do take your point though that it's perfectly possible to get lower than 60 ns on Zen, though I'm not sure how many people are.

The dual CCD Zen chips would get better write bandwidth due to two connections to the IO die, one from each CCD, but that doesn't help with gaming.

Again, completely irrelevant to bring up intel, I made no comment on it’s latency. Why do you insist on bringing this up?

The 7800X3D is the only Zen4 part in which it may be difficult to get below 60 as it’s the lowest frequency part and cannot take advantage of a second, higher clocking CCD (7900/7950X3D) - this in no way represents Zen 4 being “difficult” to achieve sub 60ns latency in Aida.

Having spent hundreds of hours with multiple kits on AM5, and assisting many other people on OCN, its not difficult to achieve. Id say the average is around 56-58ns.

 

[dgianstefani]

Can I butt in? What does a few ns tighter RAM latency give you? 0.3 FPS in your favourite game? Or 4 more points in Cinebench perhaps?

I have a mid-range GPU which is not affected by system RAM speed at all, so consider this a genuine question.

With a mid range GPU faster RAM in latency or bandwidth will typically only give you better 1%/0.1% lows, but if your game is CPU limited, like Tarkov for instance, you'll see better averages too.

With a high end GPU, it makes it so you are CPU limited less often, so you get more usage out of GPU and higher average frame rates too.

Upping resolution is same effect, only minimum FPS will increase. Look at TPU RAM reviews for more info.

Here's a couple examples where you get pretty good FPS increases from a quick tune in a CPU limited scenario, this is a CPU at static clock settings too (P-Core @ 5.5 GHz, E-Core @ 4.3 GHz (locked)), so it's just the RAM that's being tested.

Again, completely irrelevant to bring up intel, I made no comment on it’s latency. Why do you insist on bringing this up?

The 7800X3D is the only Zen4 part in which it may be difficult to get below 60 as it’s the lowest frequency part and cannot take advantage of a second, higher clocking CCD (7900/7950X3D) - this in no way represents Zen 4 being “difficult” to achieve sub 60ns latency in Aida.

Having spent hundreds of hours with multiple kits on AM5, and assisting many other people on OCN, its not difficult to achieve. Id say the average is around 56-58ns.

Intel is the topic of the thread...

The only reason X3D is even being brought up is because it's a competitor, and the different approaches to high FPS are interesting. Or did you want to look at only one side of the data?

If non X3D CCDs are necessary to get lower latency due to CPU clocks being higher, that's hardly helpful to gaming is it, since the X3D chiplet is what matters.

There was no comment whatsoever claiming what an intel part can do in terms of latency. You incorrectly stated it’s extremely difficult to get below 60ns on Zen 4 which is flat out wrong, not to mention can be done on both non 3dvcache parts and 3dvcache parts.

Showing a screenshot of loose c36 ddr5 6000 proves nothing of the sort.

I said "if you're lucky". Please don't misquote me.

Edited a few times to add detail, I was at the gym on phone typing between sets earlier, so it's a bit hard to add screenshots etc.

 

[nomdeplume]

Again, completely irrelevant to bring up intel,

Intel is the topic of the thread...

If I might since we are losing focus. Also since numerous members have been less than recognized for their potential to actually comment on this processor.

Curious how many of you will follow the OC3D video route of delidding and going with a specialty AIO/waterblock?
Especially since it diverts into brands that aren't (site sponsors) popular on TPU.

 

[BoggledBeagle]

The only reason X3D is even being brought up is because it's a competitor

I would not say this. AMD 3D chips are class of their own. Intel DOES NOT HAVE a product that can directly compete with these CPUs.

 

[dgianstefani]

If I might since we are losing focus. Also since numerous members have been less than recognized for their potential to actually comment on this processor.

Curious how many of you will follow the OC3D video route of delidding and going with a specialty AIO/waterblock?
Especially since it diverts into brands that aren't (site sponsors) popular on TPU.

I would be very interested to see a maxed out 14900KS tune under a custom loop from one of the members who've just bought one or have one on the way. Der8auer has just released a $99 direct die copper CNC waterblock which looks great.

 

[FoulOnWhite]

I would be very interested to see a maxed out 14900KS tune under a custom loop from one of the members who've just bought one or have one on the way. Der8auer has just released a $99 direct die copper CNC waterblock which looks great.

View attachment 339574

View attachment 339575

Nice.

That core 1 is still a mighty good block though

 

[Bagerklestyne]

When I wrote "the fastest", I meant single thread performance - this is the fundamental measure of the CPU performance. In multithread you can add performance simply by adding cores, you can make the CPU as fat as you wish...
If I take a single thread Cinebench performance as an example from the review on this site, 14900KS does 2410 points and the best AMD CPU 2037 points. That is almost 20% difference.
BTW even the cheap 14600K does 2072 points.

Without trying to sound pedantic/semantic I feel like single thread performance is A metric of cpu performance, not THE metric.

Perfect example being you cite a cinebench single thread score, but there's virtually no one on the planet who'd find that a useful metric as they're going to multi-thread it.

I agree it's a metric, however it's not the only metric, and for some, not a relevant one as it doesn't reflect something like gaming and some application performance where things like cache size play into the framerates. Others who use cinebench style applications are looking for the multithreaded score as that's how long the 'job' takes.

Unpopular opinion, but I think a single core cinebench score should be relegated to how many mips a cpu can do.

I think there are 3 metrics of a cpu
ipc at a given clock speed (where cinebench might be relevant but isn't necessarily ideal due to variances in instruction set support for a given task)
cache size (and the impact it provides, sometimes to great effect, other time to no useful increase)
clock speed at single thread and multi thread loads in conjunction with ipc (no good having a super fast single thread if multi thread is woeful and you're looking for multithreaded performance)

with a potential 4th metric being the power required to complete a given workload (where efficiency is a consideration)

I am tired so this might not make sense

 

[dgianstefani]

Without trying to sound pedantic/semantic I feel like single thread performance is A metric of cpu performance, not THE metric.

Perfect example being you cite a cinebench single thread score, but there's virtually no one on the planet who'd find that a useful metric as they're going to multi-thread it.

I agree it's a metric, however it's not the only metric, and for some, not a relevant one as it doesn't reflect something like gaming and some application performance where things like cache size play into the framerates. Others who use cinebench style applications are looking for the multithreaded score as that's how long the 'job' takes.

Unpopular opinion, but I think a single core cinebench score should be relegated to how many mips a cpu can do.

I think there are 3 metrics of a cpu
ipc at a given clock speed (where cinebench might be relevant but isn't necessarily ideal due to variances in instruction set support for a given task)
cache size (and the impact it provides, sometimes to great effect, other time to no useful increase)
clock speed at single thread and multi thread loads in conjunction with ipc (no good having a super fast single thread if multi thread is woeful and you're looking for multithreaded performance)

with a potential 4th metric being the power required to complete a given workload (where efficiency is a consideration)

I am tired so this might not make sense

??????

What do you think multithread is if not a multiple of single threaded performance depending on the number of cores?

In my view ST is basically the only metric that is always relevant no matter what you are doing. If you a doing a single threaded task, lightly threaded, massively threaded, gaming or otherwise, per core performance is literally always a relevant factor.

Even with cache size, e.g. 12900K vs 13700K (similar core counts and architecture etc, different cache and frequency), it's still just something that affects ST/MT, or how reliant on fast RAM the system is (cached is better).

IPC, clock speed etc, these are all simply factors that play into the single threaded performance metric (and therefore the MT performance metric, assuming you aren't power or thermally limited).

You're right about cache being somewhat disconnected from ST metrics though, and it sometimes helping, sometimes not.

I suspect this is because most ST benchmark datasets fit fully within the cache, even if it's small, or don't fit at all, therefore whether you have, say a 1 MB dataset or a 10 MB dataset, it doesn't make any difference because pretty much all modern CPUs have at least 30 MB of cache.

 

[Bagerklestyne]

??????

What do you think multithread is if not a multiple of single threaded performance depending on the number of cores?

In my view ST is basically the only metric that is always relevant no matter what you are doing. If you a doing a single threaded task, lightly threaded, massively threaded, gaming or otherwise, per core performance is literally always a relevant factor.

Even with cache size, e.g. 12900K vs 13700K (similar core counts and architecture etc, different cache and frequency), it's still just something that affects ST/MT, or how reliant on fast RAM the system is (cached is better).

IPC, clock speed etc, these are all simply factors that play into the single threaded performance metric (and therefore the MT performance metric, assuming you aren't power or thermally limited).

You're right about cache being somewhat disconnected from ST metrics though, and it sometimes helping, sometimes not.

I suspect this is because most ST benchmark datasets fit fully within the cache, even if it's small, or don't fit at all, therefore whether you have, say a 1 MB dataset or a 10 MB dataset, it doesn't make any difference because pretty much all modern CPUs have at least 30 MB of cache.

Multi threaded core performance is not a direct multiplier though, you don't jump from 1 core to 16 and still get the same clock speed and some chips hold higher multicore boost clocks compared to single core. The 14900KS runs 6.2 on primary threads but it drops into the 5's as you load more cores up.

Point here being is multithreaded is not directly correlated to single core boost. Otherwise X amount of cinebench on a single core would always translate to the same amount of multi thread performance which it doesn't. You don't get say 2000 pts on single core then guarantee you'll get 10000 out of every 16 core cpu.

eg - 5950x (1651 single - 25869 multi - 15.66x faster)
7950x (2037 single - 38036 multi - 18.67x faster) These are TPU's numbers from here

I still maintain that ipc and clock while playing a part are independant of a pure single threaded score. What if the IPC improvements didn't directly correlate to useful improvements for executing cinebench, a given cpu might be faster executing a different task (game ? code compile) but still have an identical cinebench score at the same frequency.

When architecture was far more simplistic that it is now, I'd agree that ipc/clock tied together but with different value now due to multiple supported instruction sets.

Correct me if I read it wrong (3rd line "In my view...") per core performance is relevant, however, in gaming a crappy single thread cinebench score doesn't necessarily equate to crap gaming. Hence it's a yard stick that's not correct.

You can't say 2400 ST cinebench with this gpu gives you 200 fps in this game regardless of the cpu that's given the cinebench score. the 14900K has a ST cinebench of 2400 it's 30+% faster than the 7800X3D but doesn't give you 30% more framerate, or equally, clock the 14900 down to give you a 7800X3D cinebench score and it's gaming performance wouldn't be as competitive.

Point I am angling at here in this waffle is single threaded isn't the be all for a metric of performance, it's relevant but needs context and should be taken into consideration with the other factors I mentioned.

I hope I'm not dragging this off topic too far.

/waffle

 

[dgianstefani]

5950x (1651 single - 25869 multi - 15.66x faster)
7950x (2037 single - 38036 multi - 18.67x faster) These are TPU's numbers from here

Seems almost perfect scaling to me. 1-16x.

 

[Solaris17]

Also since numerous members have been less than recognized for their potential to actually comment on this processor.

Curious how many of you will follow the OC3D video route of delidding and going with a specialty AIO/waterblock?

TBH no one currently keeping this thread alive is interested in what owners think or tested. I'm not participating because of that, and even if I wanted to I would probably just trade a post or two with @Outback Bronze or @Super Firm Tofu in some other thread. Maybe like we did on page 8. It's also likely why they are not posting. Nothing about the current discussion indicates any kind of openmindedness or technical curiosity.

I'd almost rather post in the news thread; and I generally only rank those slightly higher than a cesspit.


Less that I didnt even watch the vid, nor tbh do I even care to.

 

[freeagent]

How silly do you think it would be to run an i9 on a little itx Strix B660?

Probably pretty silly..