Der8auer hits 5.0GHz on i9-7900X with AIO water cooling

[79wjd]

17 minutes ago, Red Hardware said:

Hi

It's probably a cherry-picked processor with lowest TDP and best bins that they could find . It's unlikely that others would be able to get 4.8 or 5 without LN2 out of it.

And it doesn't change any fact about their horrible X299 platform and pricing .

Some people mention that Intel has better IPC than AMD so it has a better single core performance, the problem is that anybody who is going to buy one of these and use X299 or X399 is looking into multi core/thread workloads that can benefit from all of these extra cores so talking about single core performance doesn't make sense .

There are tasks that still scale with both single core performance and more cores. So if both are weighted equally (which probably is the case for at least one use case) then an 8 core skylake X would be comparable to a 10c threadripper and a 12c skylake X would be comparable to a 15c threadripper (in theory).

 

Also, I think the big point with this thread isn't the achievable clockspeeds as much as it's the fact that skylake X being not soldered doesn't appear to affect it as people feared it would.

[Flavio hc 16]

a 10 core run a 4.8 ghz yeeeeee!

Doesn't matter that it is on a 280mm rad and it is running at 85-90 degrees, even the one delidded is running in the high eighties.

Not talking about the fact that it may as well be a golden chip

[done12many2]

15 minutes ago, Red Hardware said:

Hi

It's probably a cherry-picked processor with lowest TDP and best bins that they could find . It's unlikely that others would be able to get 4.8 or 5 without LN2 out of it.

And it doesn't change any fact about their horrible X299 platform and pricing .

Some people mention that Intel has better IPC than AMD so it has a better single core performance, the problem is that anybody who is going to buy one of these and use X299 or X399 is looking into multi core/thread workloads that can benefit from all of these extra cores so talking about single core performance doesn't make sense .

 

HI, 

 

You seem to have trouble with details. He stated it was cherry picked in the video. 

 

Like I mentioned earlier, 90% of the people posting here won't use either AMD's or Intel's HEDT platforms, but we've all got our opinions on what they're capable of. 

 

I for one use HEDT, but I still care about single-threaded performance. 

 

How about when Threadripper releases and you finish your build, come back and report your findings?

 

I'll do the same. 

[leadeater]

36 minutes ago, done12many2 said:

That latency though.  

 

I've been following Jpmboy on OCN in the Ryzen memory thread and latency is pretty bad.  Bandwidth seems okay, but the latency on Ryzen's dual channel is way higher then the Intel's current dual channel.  It's even higher then the latency on my x99 quad channel.

 

Does latency not impact performance on Zen as it does Intel?

Right now the issue is the latency, memory bandwidth itself is fine or is if you can queue up enough and keep it that way.

 

There is also the concept of throughput that needs to be remembered, think of bandwidth as the technical specification or theoretical limit and throughput as the actual amount you are able to achieve,

 

Throughput is directly related to latency, higher the latency the lower the throughput (without large queue depth).

 

Throughput = data size / latency (this is true for every data link i.e. network or memory).

 

This is why adding memory channels is not that effective unless you are starving multiple cores of memory bandwidth, like those old dot matrix printers. They are slow as shit but you can print multiple copies at once by layering pages, pages per minute increases but the actual printer speed is the same.

[Red Hardware]

17 minutes ago, done12many2 said:

 

HI, 

 

You seem to have trouble with details. He stated it was cherry picked in the video. 

 

Like I mentioned earlier, 90% of the people posting here won't use either AMD's or Intel's HEDT platforms, but we've all got our opinions on what they're capable of. 

 

I for one use HEDT, but I still care about single-threaded performance. 

 

How about when Threadripper releases and you finish your build, come back and report your findings?

 

I'll do the same. 

Hi

I don't have "trouble with details" . I didn't watch the full video so that's why I said that.

And I know that most of us are not going to use X299/X399 . 

I just said my opinion on it as others and I don't know why you are getting mad at it ?

Quote

How about when Threadripper releases and you finish your build, come back and report your findings?

I'm not going to buy Threadripper or anything HEDT(X299/X399) . I'm going for a Ryzen 7 with 8 cores/ 16 threads and it's more than what I need.

I just said ,my idea about Intel's HEDT platform ; as you said most of the people are here to do the same.

So what's the problem ???

 

Have a nice day

[done12many2]

1 minute ago, Red Hardware said:

Hi

I don't have "trouble with details" . I didn't watched the full video so that's why I said that.

And I know that most of us are not going to use X299/X399 . 

I just said my opinion on it as others and I don't know why you are getting mad at it ?

I'm not going to buy Threadripper or anything HEDT(X299/X399) . I'm going for a Ryzen 7 with 8 cores/ 16 threads and it's more than what I need.

I just said ,my idea about Intel's HEDT platform ; as you said most of the people are here to do the same.

So what's the problem ???

 

Have a nice day

 

The problem is, through your own self admission, you're commenting on stuff you didn't bother to watch and don't intend to purchase.  Even if I'm not interested in something, I'll read or watch whatever it's about before I end up saying something.  

 

So if you couldn't take the time to watch what you're commenting on do you expect people to take your opinion seriously?  If they can't take your opinion seriously is talking about something just for the sake of talking about it worth your time?  To each their own.

[Dackzy]

let us just say that this is true for them all, it is still a 1000$ CPU, while AMDs 10c offering is prob going to be between 500 and 600$ (at least if that 850$ for the 16c is correct). That is a 400-500$ difference, which is a rather big difference in price, plus then we have that 44 lanes vs 64 lanes, which might not be a difference that is worth anything for a lot of people. 

Is that extra IPC and higher clocks really worth the extra 400-500$? I personally don't think it is, for that money I could buy a GTX 1070 or 1080. I would personally rather put the 400-500$ towards a better GPU and maybe a nice 500GB NVMe SSD.

[Rune]

This makes me sad my 6850k can only hit 4.9

[MageTank]

33 minutes ago, leadeater said:

Right now the issue is the latency, memory bandwidth itself is fine or is if you can queue up enough and keep it that way.

 

There is also the concept of throughput that needs to be remembered, think of bandwidth as the technical specification or theoretical limit and throughput as the actual amount you are able to achieve,

 

Throughput is directly related to latency, higher the latency the lower the throughput (without large queue depth).

 

Throughput = data size / latency (this is true for every data link i.e. network or memory).

 

This is why adding memory channels is not that effective unless you are starving multiple cores of memory bandwidth, like those old dot matrix printers. They are slow as shit but you can print multiple copies at once by layering pages, pages per minute increases but the actual printer speed is the same.

I tried to touch upon that in my memory guide. Bandwidth only ever matters if you do not have enough of it. Once you have enough to saturate whatever it is you are doing, more won't help. Latency on the other hand, ALWAYS scales. Meaning, you will always see an improvement (even if you can't feel it per say) in your results. It's especially true when comparing synthetic benches, of low latency vs high latency.

 

It's why I tell people "once you have enough bandwidth, focus entirely on latency". Modern dual channel DDR4 is plenty of bandwidth in my eyes. Especially once you hit speeds of 3000mhz, then you have copious amounts (unless AVX is brought into the equation, but there is an entirely different set of obstacles to overcome for that, so i'll leave it at that). 

 

There is a lot that I do not understand about Ryzen, and I certainly do not claim to be an expert on that subject, but I imagine with the way it's Infinity Fabric works, both are still important, but latency has to still play a huge role in that. We won't know until we get some validated tests with the new AGESA firmware and unlocked registries, but I'd wager some sizable gains in performance if we can get that latency down to tolerable levels. My first goal is sub-50ns (a far cry from the 65ns I am currently hovering at) and from there, sub 45ns. That would put it in the realm of "normal" Intel DDR4 results. In order for it to hit "mega tweaked" levels, it would have to be capable of sub-40ns (something even the best IMC's struggle with if your board/kit isn't up to the task). Again, all of these results are within the context of Aida64, your latency will change depending on the problem size and program you use.Various compression levels of your software will also impact the results, as do background tasks and memory utilization.

40 minutes ago, djdwosk97 said:

There are tasks that still scale with both single core performance and more cores. So if both are weighted equally (which probably is the case for at least one use case) then an 8 core skylake X would be comparable to a 10c threadripper and a 12c skylake X would be comparable to a 15c threadripper (in theory).

 

Also, I think the big point with this thread isn't the achievable clockspeeds as much as it's the fact that skylake X being not soldered doesn't appear to affect it as people feared it would.

If that were the big point of this thread, it would say in the title "without a delid". That being said, most of us have been repeating on this forum that the problem was never the thermal compound. The glue itself was always a big issue with causing higher per-core temp variance, and in the case of the glue expanding improperly, extremely high thermals all around. Yes, it's no secret that liquid metal is superior for raw thermals, but normal conventional pastes are still good from a longevity standpoint (assuming it can withstand the harsh thermal cycling) and will still do a great job for the most part. The elevated substrate has almost certainly changed Intel's IHS gluing procedure, and I think we can blame that for this radical change in thermal performance.

 

I would like to remind everyone that I have never bashed Dow Corning paste, and I never will. It does exactly what it was designed to do, and it's yet to be matched in that regard. Shin Etsu's formula has gotten extremely close though, so there is always that alternative for those of you that wish to find a nice durable paste for long-term use on bare dies. EVGA uses Shin Etsu on their GPU's, so there is a little fun fact for you. 

 

We only "bashed" Intel on their decision to use conventional pastes over solder, because there seems to be no logical reason (other than saving cost/rushing products) for them to do so. At the end of the day, solder is superior on a die this large. 

[PCGuy_5960]

48 minutes ago, Red Hardware said:

I don't have "trouble with details" . I didn't watch the full video

It's 3 fucking minutes long. Watch it before you post.

[ravenshrike]

2 hours ago, PCGuy_5960 said:

X299 has 24 chipset lanes, so you get 68 lanes in total

No, you get 44 lanes with and additional 24 shunted over a DMI 3.0 bus along with all your other I/O. Assuming I'm got the specs right, that's about equal to a single PCI-E 3.0 x4 connection. It is not remotely as useful as 24 full up PCI-E 3.0 lanes.

[PCGuy_5960]

1 minute ago, ravenshrike said:

No, you get 44 lanes with and additional 24 shunted over a DMI 3.0 bus along with all your other I/O. Assuming I'm got the specs right, that's about equal to a single PCI-E x4 connection. It is not remotely as useful as 24 full up PCI-E 3.0 lanes.

Yeah, but you still get another 20 or so lanes from the chipset...

[79wjd]

@MageTank I didn't get a chance to watch the video, but the OP focused on the overclocking results relative to delidded results, so I'd still say the focus (or the intended focus) of the thread is that not being soldered is fine. Although yes, there isn't any reason for Intel to have not soldered them.

[Red Hardware]

50 minutes ago, done12many2 said:

 

The problem is, through your own self admission, you're commenting on stuff you didn't bother to watch and don't intend to purchase.  Even if I'm not interested in something, I'll read or watch whatever it's about before I end up saying something.  

 

So if you couldn't take the time to watch what you're commenting on do you expect people to take your opinion seriously?  If they can't take your opinion seriously is talking about something just for the sake of talking about it worth your time?  To each their own.

what ???!!!

Quote

you're commenting on stuff you didn't bother to watch and don't intend to purchase.

Don't intend to purchase !!?!!?? So all of the people here are going to bye this ?????

 

Quote

So if you couldn't take the time to watch what you're commenting on do you expect people to take your opinion seriously?

The whole thing that you are arguing about is that I didn't know that he said he is using a cherry picked processor in his video and ended up figuring it up myself . Is that your problem ??? Okay , sorry 

And I tried to watch the video but i couldn't because it wasn't loading at the time.

And I'm interested in this . That's why I'm here. You should learn to be flexible and talk kindly to people instead of attacking them for no reason

Have a nice day

[leadeater]

5 minutes ago, PCGuy_5960 said:

Yeah, but you still get another 20 or so lanes from the chipset...

It's more about how effectively those lanes can be utilized. Since PCIe is serial more lanes means more devices can be used since you cannot share lanes, however the overall bandwidth is limited to x4 across the DMI interface meaning those 24 lanes are mostly only useful for USB and 1Gb Ethernet and not great for multiple NVMe devices.

[PCGuy_5960]

3 minutes ago, djdwosk97 said:

I'd still say the focus (or the intended focus) of the thread is that not being soldered is fine. Although yes, there isn't any reason for Intel to have not soldered them.

I don't think that if they were soldered they would OC better... Since Broadwell-E which was soldered also got a 5-10c drop if you replaced the solder with liquid metal

[MageTank]

1 minute ago, PCGuy_5960 said:

I don't think that if they were soldered they would OC better... Since Broadwell-E which was soldered also got a 5-10c drop if you replaced the solder with liquid metal

Where did you see that? I saw 3-5C at best from Der8auer himself. The risk certainly wasn't worth the reward, and it certainly wasn't 10C. 

[PCGuy_5960]

Just now, MageTank said:

Where did you see that? I saw 3-5C at best from Der8auer himself. The risk certainly wasn't worth the reward, and it certainly wasn't 10C. 

IIRC, der8auer got a 5-10c drop with Broadwell-E and a 3-5c drop with Haswell-E

[dexT]

Nice, awesome potential really exciting.

 

Just remember though, this guy is one of the best overclockers in the world. He could go in to any of our systems and get a little extra over what the average enthusiast can get.

[done12many2]

5 minutes ago, Red Hardware said:

Have a nice day

 

You too.  Good luck with your Ryzen R7 build and thanks for sharing your thoughts.  

[done12many2]

1 minute ago, dexT said:

Nice, awesome potential really exciting.

 

Just remember though, this guy is one of the best overclockers in the world. He could go in to any of our systems and get a little extra over what the average enthusiast can get.

 

Even he can't get anymore OC out of certain platforms using just an AIO.  

[done12many2]

6 minutes ago, PCGuy_5960 said:

IIRC, der8auer got a 5-10c drop with Broadwell-E and a 3-5c drop with Haswell-E

 

This was all he got with the 6950x delidded.

 

Lack of solder proved to not be that big of a deal at all.  

 

[MageTank]

Just now, PCGuy_5960 said:

IIRC, der8auer got a 5-10c drop with Broadwell-E and a 3-5c drop with Haswell-E

He saw a 6C difference at "max temp" and a 4C difference at "average max temp". Max temps can be deceptive, as CPU sensors themselves are prone to random spikes that slightly skew the results. You can see that by the fact that his hottest core was 80C after delid, while the rest did not exceed 78C. However, pre-delid, his max temp was 86C on multiple cores (2 to be exact) with the next highest being 83C. That being said, his average max temp is exactly in line with what we expected. He even goes on to say it's not worth it. 

 

If you've seen a 10C difference from delidding a soldered CPU, then something went horribly wrong in the soldering process, and it's a miracle the die itself survived.

[Red Hardware]

34 minutes ago, PCGuy_5960 said:

It's 3 fucking minutes long. Watch it before you post.

Hi

It wasn't loading , can you understand that ????????? And I guessed that he is using a cherry picked cpu that turned out right !!! What's your problem ?? Why people are rude here ?

[dexT]

1 minute ago, done12many2 said:

 

Even he can't get anymore OC out of certain platforms using just an AIO.  

True but even both on air he could clock higher than me I image.