N4CR
Supreme [H]ardness
- Joined
- Oct 17, 2011
- Messages
- 4,947
That is unfortunate.
Perhaps a limitation of scaling the interconnect topology to 8 cores? Or thermal?
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Intel Core i9-9900K 9th Generation CPU Review @ [H]
That is unfortunate.
Perhaps a limitation of scaling the interconnect topology to 8 cores? Or thermal?
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
Hard to say without frequency plots, frametimes, or well any real data at all with that review. Suffice to say that if gaming is the most intense thing you do, the 9700K would likely be indistinguishable.
I'd prefer the smattering of [H] and Techreport comparison reviews before making hard conclusions, for the better data and for different samples.
- Joined
- May 18, 1997
- Messages
- 55,634
Intel said it would be sending the 9600K soon, and we did ask for a 9700K, but not sure on it. Have not been able to catch a 9700K in the wild for sale yet.
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
Thanks. I didn't see that review before. I just cancelled my 9900K pre-order.
I knew it would be a marginal upgrade, but looking at the 1440p/4k benchmarks it's barely even an improvement at all (the 8700K won some tests, even by a fraction of an fps, but still).
Returning to the power consumption:
Intel's datasheet, page 103 specify:
PL1 = 95W,
PL2 = PL1 * 1.25 [=119W]
PL1 Tau = 0.1 to 8 seconds.
These numbers are the same for all upper end 8th generation Core CPUs
There are footnotes on page 98 saying that
8. PL2 can be exceeded [up to PL3, stated elsewhere in the same document] for up to 10ms.
16. PL1*1.25 is the default value for PL2, it might not be the optimum for performance. It can be higher when proper thermal management is used.
Anandtech mentions in their review that PL2 is supposed to be 210W. No reference for this number seems to be given in the review, but I have another source mentioning that the 210W comes from an Intel employee interviewed by Anandtech.
To me it seems like Intel has set their official power ratings to match previous Coffee Lake so that "old" motherboards with 300-series chipset are fully compatible from a technical point of view.
That said, more power (and cooling) is required to hit the maximum boost clocks, especially if those speeds are to be sustained beyond PL1 Tau.
The 210W PL2 is possibly the "optimum performance" value mentioned in footnote 16, what the CPU actually draws when all eight cores are running at 4.6 GHz. There are possibly some Z370 motherboards around that can't provide that much power in a reliable manner.
Intel's datasheet, page 103 specify:
PL1 = 95W,
PL2 = PL1 * 1.25 [=119W]
PL1 Tau = 0.1 to 8 seconds.
These numbers are the same for all upper end 8th generation Core CPUs
There are footnotes on page 98 saying that
8. PL2 can be exceeded [up to PL3, stated elsewhere in the same document] for up to 10ms.
16. PL1*1.25 is the default value for PL2, it might not be the optimum for performance. It can be higher when proper thermal management is used.
Anandtech mentions in their review that PL2 is supposed to be 210W. No reference for this number seems to be given in the review, but I have another source mentioning that the 210W comes from an Intel employee interviewed by Anandtech.
To me it seems like Intel has set their official power ratings to match previous Coffee Lake so that "old" motherboards with 300-series chipset are fully compatible from a technical point of view.
That said, more power (and cooling) is required to hit the maximum boost clocks, especially if those speeds are to be sustained beyond PL1 Tau.
The 210W PL2 is possibly the "optimum performance" value mentioned in footnote 16, what the CPU actually draws when all eight cores are running at 4.6 GHz. There are possibly some Z370 motherboards around that can't provide that much power in a reliable manner.
As someone who bitched about the lack of metal TIM etc... I wonder if my bitching was now justified. Yeah great cpu but get out your wallet and blinders . I’ll stick with my delid 8086k@5ghz on aio at low volts. I’m a “144hz gamer” and see no reason to upgrade. Sad as I really wanted to want this cpu. I usually find any reason to upgrade because I love playing with new bits but this doesn’t even tempt me in the least. I call this a fail. My 2cents and I’m sure others will disagree but with the metal TIM and refined process I was expecting average overclocks at least as good as 8700/8086k levels. Seems the added cores are exposing the process/power limits. It is what it is and I don’t blame intel but I also don’t plan to buy one. I should really have guess this since my [email protected] saturates my current Artic 240 AIO cooler. Hopefully AMD can up the MHz with 7nm zen and get to 4.8ghz plus. I think I'll be disappointed there too lol. Oh well, can always hope!
Last edited:
vmirjamali
Limp Gawd
- Joined
- May 14, 2010
- Messages
- 170
That's kind of sad.. Double the cost if a 2700 with more power as well.. I think amd was holding out the 2800x for this. We shall see.
Thank you for the review.
Going back to percentage mathematics, the % computations for "Time in Seconds - Lower is better" graphs are systematically incorrect.
For an example, consider the Blender test in https://www.hardocp.com/article/2018/10/19/intel_core_i99900k_9th_generation_cpu_review/3, where it reads
The 9900K Stock is ~18% faster than the 2700X and the 9900K OC is 25% faster than our 2700X.
Intel 9900K Stock finished the test in time of 18.08 seconds, whereas 2700X finished the test in time of 22.25 seconds. One can write the following correct statements:
- Intel 9900K Stock took 81.26% of the time of 2700X to finish the test (18.08 / 22.25), or
- Intel 9900K Stock took -18.74% less time than 2700X to finish the test (18.08 / 22.25 - 1).
The figure 18% in quote seems to come from computing 1 - 18.08 / 22.25 = 18.74%. Under this math, for Intel to have been called 100% faster, it would have had to complete the test in 0 seconds (1 - 0 / 22.25 ) = 100%), but that does not make sense. If product A is 100% faster than product B, then it can complete two units of work in the same time as product B completes one unit of work. Conversely, if product A is 100% faster than product B, then it needs to take just half the time (not zero time) for A to complete the same amount of work than product B does. Hence, if Intel 9900K had completed the Blender test in time of 11.125 seconds, it would be twice as fast (by completing the test in half the time of Ryzen's 22.25s time).
The corrected percentage statement is
Intel 9900K Stock is 23.06% (22.25 / 18.08) faster than the 2700X and the 9900K OC is 32.68% (22.25 / 16.77) faster than our 2700X.
HardOCP gets the "faster than" comparisons correct in graphs where higher is better, since those naturally compare amount of work performed in the same amount of time, but in tests where the amount of work is the same, and the amount of time to complete it varies (i.e. lower is better), the math goes wrong.
Going back to percentage mathematics, the % computations for "Time in Seconds - Lower is better" graphs are systematically incorrect.
For an example, consider the Blender test in https://www.hardocp.com/article/2018/10/19/intel_core_i99900k_9th_generation_cpu_review/3, where it reads
The 9900K Stock is ~18% faster than the 2700X and the 9900K OC is 25% faster than our 2700X.
Intel 9900K Stock finished the test in time of 18.08 seconds, whereas 2700X finished the test in time of 22.25 seconds. One can write the following correct statements:
- Intel 9900K Stock took 81.26% of the time of 2700X to finish the test (18.08 / 22.25), or
- Intel 9900K Stock took -18.74% less time than 2700X to finish the test (18.08 / 22.25 - 1).
The figure 18% in quote seems to come from computing 1 - 18.08 / 22.25 = 18.74%. Under this math, for Intel to have been called 100% faster, it would have had to complete the test in 0 seconds (1 - 0 / 22.25 ) = 100%), but that does not make sense. If product A is 100% faster than product B, then it can complete two units of work in the same time as product B completes one unit of work. Conversely, if product A is 100% faster than product B, then it needs to take just half the time (not zero time) for A to complete the same amount of work than product B does. Hence, if Intel 9900K had completed the Blender test in time of 11.125 seconds, it would be twice as fast (by completing the test in half the time of Ryzen's 22.25s time).
The corrected percentage statement is
Intel 9900K Stock is 23.06% (22.25 / 18.08) faster than the 2700X and the 9900K OC is 32.68% (22.25 / 16.77) faster than our 2700X.
HardOCP gets the "faster than" comparisons correct in graphs where higher is better, since those naturally compare amount of work performed in the same amount of time, but in tests where the amount of work is the same, and the amount of time to complete it varies (i.e. lower is better), the math goes wrong.
OrangeKhrush
[H]ard|Gawd
- Joined
- Dec 15, 2016
- Messages
- 1,673
Zen+ is still behind Skylake in terms of both IPC and clockspeeds- AMD has made up a lot of lost Dozer ground, but what they've caught up to (say, Haswell) is still five years behind. Intel getting stuck at 14nm is the main reason they've gotten so close, and good on AMD, but the facts still put a lot of work on their shoulders to 'catch up'.
IPC is the same, Intel has been the same for years now and to alter that will require a new architecture. What intel are doing is running 800-1000mhz faster clockspeeds and motherboard vendors are breaching boost state and TDP policies to ensure higher performance but on a clock vs clock comparison they are about par. They do have better branch predictions and higher throughput than AMD but AMD is catching up fast. The only difference here is that AMD are not shady about it.
FedEx is just messing with me. My 9900K sat 3.5 hours away for a day. 24 hours later, it’s finally moving from Indy to Memphis. But I live in the St. Louis area.
It;s going to take them 3 days to get it to me when it was just 3,5 hours away to begin with.
It;s going to take them 3 days to get it to me when it was just 3,5 hours away to begin with.
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
OrangeKhrush
[H]ard|Gawd
- Joined
- Dec 15, 2016
- Messages
- 1,673
Intels single precision IPC Haswell to Coffee lake is 32 instructions per cycle, Ryzen and Ryzen 2 are the same.
Shady? really do we have to go through this again? of all the shinnanigans of recent weeks on top of years of veiled daggers I think it is hard to look at intel without cringing.
We have run our own testings running intel without boosts and tricks, running it at same clocks as AMD and the results on a clock vs clock is quite a different story.
- Joined
- May 18, 1997
- Messages
- 55,634
The IPC seems to be about the same and most legit reviewers agree.
Steve from GN and now Steve from HW are bringing attention to the default infinite turbo states on all of the motherboards. At the very least, Intel needs to change their TDP for the 9th gens.
Of course, some will say "well Intel can't help what the MB manufactures do." That would be rather naive as it is clear that Intel wants their chips to run on at 4.7 ghz since it seems to be the default on ALL of the motherboards.
Long-hair Steve shows that, without these infinite turbos, the 9900k and 2700x are very close in performance.
IdiotInCharge
NVIDIA SHILL
- Joined
- Jun 13, 2003
- Messages
- 14,675
"Oh no, Intel is letting its CPUs run as fast as possible! Someone stop them!"
This behavior is desirable. Particularly if the cooling solution and power delivery can handle it, which most can; hell, what if it's running in a cooler environment?
The key here is stability, and that has not presented itself to be an issue. You can quibble about the marketing but the performance stands.
This behavior is desirable. Particularly if the cooling solution and power delivery can handle it, which most can; hell, what if it's running in a cooler environment?
The key here is stability, and that has not presented itself to be an issue. You can quibble about the marketing but the performance stands.
- Joined
- May 18, 1997
- Messages
- 55,634
I think this is all effed up due to poor marketing on Intel's part. It has become an issue instead of a feature.
- Joined
- May 18, 1997
- Messages
- 55,634
Nightfire and IdiotInCharge - You cannot stop the personal sniping and bickering, you are no longer allowed to post in this thread. If you do not heed this warning, it will get a month vacation.
OrangeKhrush
[H]ard|Gawd
- Joined
- Dec 15, 2016
- Messages
- 1,673
Actually lying is a massive felony, it is probably the reason why Intel have a ledger book of pending legal actions against them. VW lied about carbon emissions offering high performance on fictitious carbon emissions proven to be "half truths" or blatant lies, the penalty for VW group was very severe. Intel are breaking thermal policies just to hold on to the cup cake, this is the reason they don't ship stock coolers and Intel board vendors have been thrown under the bus by industry figures as to how they are manipulating results by tampering with thermal policies. Lying to cheat results to influence a market is about every contravention of international treaties regarding anti competition practices. The problem is more are wiser to this and it hasn't stopped AMD gaining traction.
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
Intel has a TDP, and specifies how the chip should operate to remain within that window. And then, since it is an unlocked chip, they allow you to actually run it however you want, with what is obviously going to increase the power draw.
The alternative is a locked down chip, which I really don't think any of us want, right?
My Ryzen runs a lot faster because I changed it from "safe stock TDP" as well. It's sucking down power like a cougar hitting mojitos at happy hour, but I asked it to, and it is delivering some scary performance in return. It could of course sip white wine politely, but I didn't buy it for that.
This is the same story as my Intel box, which is also changed from stock settings, and of course also consuming more power in the process.
The stock TDP means F all in both cases.
This is a win for the consumer, I am not clear why we'd view it as anything other than that.
Last edited:
kirbyrj
Fully [H]
- Joined
- Feb 1, 2005
- Messages
- 30,693
I think the real problem from what I've seen is not so much that you are able to go beyond the normal TDP range. It's that motherboard vendors are presetting it to override the TDP values to boost performance (in theory) to separate themselves from other manufacturers and not being clear that this is being done. I'm not sure this is really Intel's fault per se unless they are actively encouraging it. There are several articles dealing with this issue. A "stock" 9900k limited to a 95W TDP over sustained loads performs at "only" 71% of the performance of one with an unlimited TDP imposed by the motherboard according to one site (Anandtech or Techspot I think I read that).
when new games come out down the road that require more cpu muscle the i9 9900k is going to take a bigger lead against the 2700x and 8770k that is not being shown at this moment.
I bought the 9900k so I don't have to upgrade down the road, it actually saves me money.
bf1 mp in a full server is unplayable for me with a 4 core cpu, with 40-50 fps drops within seconds, heavy movement, stutters, etc. with a 6 core cpu like the 8700 the game pretty much wakes up inside a full bf1 mp server.
so the 2 extra cores in bf1 mp full server makes a huge difference, in the future the same faith will come to 8700 cpu where 6 cores just wont be enough. as for the 2700x, it will fair better than than the 8700 in terms of stability of the game code, but the avg fps in games will drop considerably due to weaker cores and slower memory.
and I think this is where the investment in 9900k will pay off big time.
I bought the 9900k so I don't have to upgrade down the road, it actually saves me money.
bf1 mp in a full server is unplayable for me with a 4 core cpu, with 40-50 fps drops within seconds, heavy movement, stutters, etc. with a 6 core cpu like the 8700 the game pretty much wakes up inside a full bf1 mp server.
so the 2 extra cores in bf1 mp full server makes a huge difference, in the future the same faith will come to 8700 cpu where 6 cores just wont be enough. as for the 2700x, it will fair better than than the 8700 in terms of stability of the game code, but the avg fps in games will drop considerably due to weaker cores and slower memory.
and I think this is where the investment in 9900k will pay off big time.
Should we overclock CPUs for every comparison and label them as stock in the results? It's a feature and a win for the consumer, after all.
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
No, you run them and label them as typical or observed, since that is what they are.
If there is an enthusiast crowd interested in 95W results, you could provide those, but that’s not of high interest to most folks frequenting sites like this one. It may be a hit on [Z]eroOCP though...
As long as 0.01% are happy, reviewers can continue to label OCed results as stock?No, you run them and label them as typical or observed, since that is what they are.
If there is an enthusiast crowd interested in 95W results, you could provide those, but that’s not of high interest to most folks frequenting sites like this one. It may be a hit on [Z]eroOCP though...
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
No, that isn't at all what I'm saying.
I'd like things labeled for what they are. If you're making a point of testing stock (fully within TDP and specs), just label it that way. Kyle does this in every graph in the article we're discussing. That's what should be done. If a review site isn't doing that - dump them, they probably don't know what they're doing.
Now, the other issue is "what should MB manufacturers do"? I have zero problem with easy OC settings being present and even the default in enthusiast boards. That's what those boards are for. Notably, business boards do NOT do this. Both are doing what their market desires. The buyer is going to get what they want most easily.
I would make a nod to your point that I'd like all boards to have a simple and highly visible option to return to stock performance / settings at any time. Ideally, with a toggle, and once in "Stock" mode, you couldn't set anything beyond those levels until you toggled out to "Performance" mode or whatever we'd call it.
Now I'd expect dang near nobody except reviewers to use it, but hey, it would be there and labeled so everyone is clear on what is going on.
OnceSetThisCannotChange
Limp Gawd
- Joined
- Sep 15, 2017
- Messages
- 249
HW Unboxed Steve just did the test and 95W 9900k is a bit more efficient than 2700x with similar performance for productivity, outside AVX(CB score about the same), while maintaining the edge in gaming.
kirbyrj
Fully [H]
- Joined
- Feb 1, 2005
- Messages
- 30,693
Exactly, an OEM or a business machine where "stock" settings are normal shows that the average consumer is throwing away $200 buying Intel over AMD for virtually the same performance.
OrangeKhrush
[H]ard|Gawd
- Joined
- Dec 15, 2016
- Messages
- 1,673
You are obfuscating it with what you want to believe it is created for. If someone sells you a car say BMW and it is pandering 500BHP and you find it to only be the case when detuned and running in thinner air that is a misrepresentation. TDP is industrial specs, it is safety specs and if you purport to say it does x with y TDP and it is a lie then you are committing and act of misrepresentation.
I can turn my airbags off because I want the thrill of it, doesn't make it any less a requirement by law to have them engaged. The only reason to circumvent it was to run out of clock policy boost times to fabricate bench results as hardware unboxed showed the 9900K running locked to 95w loses serious amounts of performance in certain benches and looks a lot less appealing.
OutOfPhase
Supreme [H]ardness
- Joined
- May 11, 2005
- Messages
- 6,579
A point I think you missed is that while I like customized defaults (including beyond 'stock' performance on enthusiast boards), I also want it indicated what that represents. I am quite literally asking that we not obfuscate anything.
If the motherboard makers are actually saying they are running with a 95W TDP and exceed it in statistically significant ways then sure, call them out for misrepresenting. If they aren't saying that, don't assume they are.
- Joined
- May 18, 1997
- Messages
- 55,634
What is funny is that it has been this way for years and years. Nothing new to see here, but seems like some folks need something to be enraged about. On this week's review, we will show you both however. Fun stuff.
- Joined
- May 18, 1997
- Messages
- 55,634
Almost like stock settings should be discussed in very specific terms....just like we did.
https://www.hardocp.com/article/2018/10/19/intel_core_i99900k_9th_generation_cpu_review/5
- Joined
- May 18, 1997
- Messages
- 55,634
Yep. As we have for years. Given our audience retarding the performance is not exactly our MO.
And know this, this IS an Intel "stock" setting as P1 and P2 power levels are defined by Intel. TDP for P1 and P2 are different. This is really a marketing failure on Intel's part, and has been for some time.
Intel says the warranty is invalidated when the processor is "operated outside of Intel’s publicly available specifications". So does this mean that if I buy a particular z-series motherboard from... *checks data* ...any board vendor that my warranty is invalidated as soon as I power up the system? (Yes, even Asus is violating publically disclosed specs, see the default Tau values.)
This is academic since there's no way for Intel to know about this for an RMA claim (unless hidden fuses are intentionally blown to log such violations), but it's kind of odd. Intel should really be requiring board vendors to stick to their specs by default. Then, as soon as the user enters the BIOS, a prompt with a disclaimer should be shown placing liability with the end user. Right now, there is no such warning.
This is academic since there's no way for Intel to know about this for an RMA claim (unless hidden fuses are intentionally blown to log such violations), but it's kind of odd. Intel should really be requiring board vendors to stick to their specs by default. Then, as soon as the user enters the BIOS, a prompt with a disclaimer should be shown placing liability with the end user. Right now, there is no such warning.
It ties into P1 and Tau (duration). Since Tau is an exponential moving average parameter that works together with P1 and P2, setting P2 to a really high value means the power limit threshold is harder if not impossible to reach, allowing for higher sustained clocks. If Tau was at Intel spec and PL2 was at the proper value of 1.25*95W=119W, the processor would downclock to PL1 levels (95W) when 119W had been exceeded for Tau = 1 second (assuming constant load since it's an exponential moving average).
It looks like this is a common cheat since even my ThinkPad has set Tau to 28 seconds, just like my two Asus based PCs. This is long enough to complete the basic benchmarks that laptop/pc reviewers typically run. Max Tau is supposed to be 8 seconds according to Intel's data sheets, going back several generations.
Oh yeah, it's neat being able to quickly verify these things by simply running Intel's own XTU software side-by-side with your favorite stress test like Prime95 26.6 (non-AVX) or Cinebench. XTU shows the PL1/PL2/Tau values set by your BIOS, even locked down ones like the ThinkPad.
- Joined
- May 18, 1997
- Messages
- 55,634
OK.
drutman
Gawd
- Joined
- Jan 4, 2016
- Messages
- 617
Amazon had one for 799 USD who in their right mind would pay this for a 1151 CPU?
I got a 5820K new for $279 at our local microcenter.
I got a 5820K new for $279 at our local microcenter.
- Joined
- May 18, 1997
- Messages
- 55,634
Dunno. The new Intel PR guy does not seem to be put off by the truth as much as the last one.
Still trying to get CPUs. Asked again today.