Intel Skylake / Kaby Lake

[Anarchist Mae]

Clearly if you compile 3 times a day for 10 seconds each time, then saving yourself 4 seconds is not a case in which you need these chips. You would be much better off with a cheap chip and few cores.

I was speaking of situations in which you actually need this power. Think simulations that take hours/days/weeks to run with the CPU at 100% usage all the time. This is where you really need HEDT. And it is a very common use case (aerospace simulations, fuel efficiency of cars, structural support of high-rises, nuclear reaction simulations, defense calculations, biological structure simulations, chemical production calculations, design of CPU coolers, etc).

This is a more extreme example, but there are uses where people took the time to benchmark 18,432 cores running a single calculation for a wheel design:
http://www.ansys.com/Solutions/Solu...se-17/external-flow-over-a-formula-1-race-car
Think of the cost saving of needing 15% fewer computers in that simulation. Also due to Amdahl's law, they wouldn't have dropped from 80% efficient to 68% efficient and thus probably need closer to 26% fewer computers if only they had chips that were 15% faster.

Interesting, thanks. So the HEDT chips would be better than server chips for this kind of work, unless you need a specific feature of those server chips? Because the server chips wouldn't clock as high, so you'd need more for the same work.

I wonder where that leaves things like AMDs Project 47 which has lots of GPUs to go with those CPUs.

 

[dullard]

Interesting, thanks. So the HEDT chips would be better than server chips for this kind of work, unless you need a specific feature of those server chips? Because the server chips wouldn't clock as high, so you'd need more for the same work.

I wonder where that leaves things like AMDs Project 47 which has lots of GPUs to go with those CPUs.

The server vs. HEDT result depends on your application. Do you need 224 cores? If so, a Xeon 8180 in an 8-socket motherboard can do it all in one computer without complex and expensive high speed data links between computers. Do you need just 32 cores? Then a single 1950X Threadripper might be the best choice. It depends on memory needs, OEM support (can you actually buy it), power consumption, etc. Too many reasons to list in a simple forum post. But a small CPU price difference rarely is one of the main reasons to determine the computer you buy in this line of work.

CPU vs GPU depends on the calculation type. Again, that is too complex for me to try to describe in a forum. I'm sure there are plenty of articles on that.

 

[wildhorse2k]

How are there no compromises? If you determined that you need about 18 cores in your VM host, then there are other Intel based offerings which give you

• higher memory capacity,
• more memory bandwidth,
• ECC memory,
• more I/O,
• lower lifetime cost by better performance/Watt (unless you underclock and undervolt the XE),
• lower purchase cost if you consider the 2nd hand market,
• or several of the above.

There is a single feature of 7980XE which cannot be found in competing Intel processors, which is somewhat higher clocks during low and medium processor utilization. But even this feature is naturally tied to a compromise: Middling performance/Watt.

You make a compromise if you accept that not all your needs are met. I don't need more than 128GB memory, in fact 64GB is enough (8x8GB). DDR 3600 and above in 4 channel configuration give a lot of memory bandwidth. Xeons need 6 channel IMC due to usage of low speed ECC memory. I don't need ECC memory or more IO. I never had a requirement for performance/watt.

 

[Bouowmx]

A question: do you think that a slightly more proper comparison to AMD Ryzen Threadripper 1950X is the Intel Core i9-7940X (14 cores, 4.3 GHz), rather than 7900X? The extra 400 USD in MSRP may look bad, but performance/price may be absorbed in the cost of the rest of the components that don't add to performance (ex. case, power supply, storage), cheaper motherboard (ASRock Fatal1ty X299 Gaming K6 vs ASRock X399 Taichi), and regional pricing (in some areas).

 

[LTC8K6]

A question: do you think that a slightly more proper comparison to AMD Ryzen Threadripper 1950X is the Intel Core i9-7940X (14 cores, 4.3 GHz), rather than 7900X? The extra 400 USD in MSRP may look bad, but performance/price may be absorbed in the cost of the rest of the components that don't add to performance (ex. case, power supply, storage), cheaper motherboard (ASRock Fatal1ty X299 Gaming K6 vs ASRock X399 Taichi), and regional pricing (in some areas).

I think the 7940X is a bit of a sweet spot because of it's clocks and number of cores. It has a higher base clock than the chip above and below it.

 

[wildhorse2k]

Do not buy cheap motherboards for high core count Skylake-X. Read http://www.overclock.net/t/1632665/intel-x299-socket-2066-vrm-thread and https://www.hardwareluxx.de/community/f12/lga-2066-mainboard-vrm-liste-1167715.html . Buy motherboard with 8+4 pin connectors or 8+8.

 

[tamz_msc]

The problem is because you are the one who is not caring about the value for money.

Lets look at ALL of the numbers. I'll take your post using your exact values. Suppose a $5700 computer was 15% faster than a $5000 computer. Yes, you are correct that it cost $700 more to purchase.

But, suppose that computer is running software that costs $10000 per year (which is actually fairly cheap, specialized software often costs $20,000+ per computer per year). The amount of work that the $5000 computer does in 1 year, the $5700 computer could do in 0.87 years (15% faster works out to needing 13% less time). Thus, you only need to pay $8696 for the software to do the same amount of work. The $700 faster processor saves you $1304/year.

Even more importantly the engineers running that software have a salary + benefits that are probably close to $100,000 / year. That engineer that took 1 year to complete a study with the $5000 computer only needs 0.87 years with the $5700 computer. Again, you save $13,043 in engineer time.

That ONE TIME purchase of $700 saves the company $1304+$13,043 = $14,348 PER YEAR.

It gets bigger when you look at teams. Look at just your 20 workstation cost: 20 * $5000 workstations cost $100,000 and the software costs ~$20 * $10000 / year = $200,000 / year. But that same amount of work could be done with just 17 workstations with ~15% faster processor. Thus you need 17 * $5700 = $96,900 in computer costs. It actually is CHEAPER to buy the team the faster processors. And considering the software, you save $30,000/year. That is even if you don't consider employee costs.

No one in business that needs this type of power should ever think that a mere $700 more for a 15% boost is even worthy of consideration.

Except your logic fails when you are outside the industry you're acquainted with. When your workplace is publicly funded, every purchase is through tenders, use of open source software is the norm, and you have compulsory auditing by federal agencies every year, 700$ makes all the difference.

EDIT: Your logic is hugely flawed at the basic level. You make the usual salary/wage fallacy - the engineer may finish the job in 0.87 years but he's not sitting idle for the rest 0.13 years. You still have to pay him his stipulated salary which is still 100K/year. Same with the software - you negotiated with the software-maker for 10K/year, so you still have to pay that amount even if you complete the job before the end of the time-frame.

One project is done - you move on to the next.

 

[Shivansps]

Except your logic fails when you are outside the industry you're acquainted with. When your workplace is publicly funded, every purchase is through tenders, use of open source software is the norm, and you have compulsory auditing by federal agencies every year, 700$ makes all the difference.

Only in the case of $700 for the same performance, if $700 more is for 15% more perf you cant really be sure at all.

 

[Bouowmx]

I watched der8auer's videos about Intel X299 a bit ago (VRM temperature too high, EPS cable current too high, and more), and in the aftermath, there was lots of information being thrown around. So, I kind of haven't followed up.

Is ASRock Fatal1ty X299 Gaming K6 make a reasonable motherboard, or is it too cheap? My use case, in particular, is "overclocking" is switching on Multi-core enhancement, and applications do not use AVX. Or can wait and see about the OC Formula variant.

 

[tamz_msc]

Only in the case of $700 for the same performance, if $700 more is for 15% more perf you cant really be sure at all.

If that 700$ is due to your vendor quoting higher prices because Intel charges more from them then it doesn't matter if it's 15% faster. You have a restrictions on the funds that you're allowed to spend and your output isn't going to be measured in dollar-value.

 

[raghu78]

I think the 7940X is a bit of a sweet spot because of it's clocks and number of cores. It has a higher base clock than the chip above and below it.

Those base clocks mean nothing given that for most desktop workloads (non AVX) the all core boost clocks are going to be way higher.

7980XE - all core turbo 3.4 Ghz
7960- all core turbo 3.6 Ghz
7940 - all core turbo 3.8 Ghz
7920 - all core turbo 3.8 Ghz

http://www.pcgamer.com/full-details-for-intels-core-i9-processor-lineup/

But the 7940x would probably beat 7920x, 7960x and 7980xe in terms of perf/$ in multi threaded workloads.

 

[Arachnotronic]

Those base clocks mean nothing given that for most desktop workloads (non AVX) the all core boost clocks are going to be way higher.

7980XE - all core turbo 3.4 Ghz
7960- all core turbo 3.6 Ghz
7940 - all core turbo 3.8 Ghz
7920 - all core turbo 3.8 Ghz

http://www.pcgamer.com/full-details-for-intels-core-i9-processor-lineup/

But the 7940x would probably beat 7920x, 7960x and 7980xe in terms of perf/$ in multi threaded workloads.

Weren't you the one telling us 3.4GHz all core turbo was "wishful thinking" for the 7980XE?

 

[raghu78]

Weren't you the one telling us 3.4GHz all core turbo was "wishful thinking" for the 7980XE?

I said > 3.4 Ghz definitely was wishful thinking. The = was probably achievable imo. Seems you were super optimistic as usual with regards to anything Intel. Its going to be fun to see your super optimism for Intel over the next 2 years.

 

[TheF34RChannel]

Now we need the most important spec that is missing,

Spoiler

PRICE

And actual worldwide availability!

 

[dullard]

Those base clocks mean nothing given that for most desktop workloads (non AVX) the all core boost clocks are going to be way higher.

7980XE - all core turbo 3.4 Ghz
7960- all core turbo 3.6 Ghz
7940 - all core turbo 3.8 Ghz
7920 - all core turbo 3.8 Ghz

http://www.pcgamer.com/full-details-for-intels-core-i9-processor-lineup/

But the 7940x would probably beat 7920x, 7960x and 7980xe in terms of perf/$ in multi threaded workloads.

Base clocks do matter, although maybe not for typical enthusiasts who have a beast of a cooling system and plenty of space around their computer for airflow. Start running in a work environment where you may cram computers together, start running them 100% non-stop, start using OEM boxes, etc and you will run at base more often than a typical enthusiast.

But, the 7940X still does occupy quite a sweet spot in Intel's lineup. Not just because of the base clock, but also the turbos are the same or faster than all of the chips above and below it at every number of cores used (with the exception of the 4 core chips which aren't much of a competitor to this 14 core chip).

Also Intel's pricing also puts the 7940X in a sweet spot. While HEDT users are less price sensitive than say Celeron users, the 7940X is the best value of the Intel HEDT chips. For example, these latest chips are nearly linear in pricing ($132.68/core)*(# Cores) - $408.86 with a coefficient of determination of 0.9936. The 7940X sits significantly below that trendline (the 7900X sits the most above the trendline as the worst value).

 

[Arachnotronic]

I said > 3.4 Ghz definitely was wishful thinking. The = was probably achievable imo. Seems you were super optimistic as usual with regards to anything Intel. Its going to be fun to see your super optimism for Intel over the next 2 years.

This is what you wrote:

Xeon Gold 6150 has a better base clock at 2.7 Ghz than 7980XE at 2.6 Ghz. Both these chips have 165w TDP. So in terms of binning the Xeon is getting the better chips and their all core turbo is 3.4 Ghz. So claiming that 7980XE will clock >= 3.4 Ghz all core is wishful rather than based on logic.

Clearly an equals sign in there.

I said 3.4ghz minimum and that was correct, it was 3.4ghz.

 

[coercitiv]

I said 3.4ghz minimum and that was correct, it was 3.4ghz.

Credit where it's due!

PS: although I find it extremely amusing all this happens over a 100Mhz difference

 

[dullard]

PS: although I find it extremely amusing all this happens over a 100Mhz difference

We only get chip launches a couple times per year. We need something to debate.

 

[TahoeDust]

Jesus! The last few pages in this thread have sucked.

 

[coercitiv]

We only get chip launches a couple times per year. We need something to debate.

Oh, definitely. As I said, credit where it's due!

 

[raghu78]

This is what you wrote:

Clearly an equals sign in there.

I said 3.4ghz minimum and that was correct, it was 3.4ghz.

You did not say =3.4 Ghz only as the Xeons max all core boost (which had better base clocks due to better binning). That would have been based on logic. You were wishing its more than 3.4 Ghz which turned out false. btw your super optimism will be worth watching with Intel's train wreck at 10nm.

 

[nvgpu]

 

[Arachnotronic]

You did not say =3.4 Ghz only as the Xeons max all core boost (which had better base clocks due to better binning). That would have been based on logic. You were wishing its more than 3.4 Ghz which turned out false. btw your super optimism will be worth watching with Intel's train wreck at 10nm.

That's not what I said, check out my original post below

FUD. It'll be >= 3.4GHz. Look at the Xeons, which are less aggressively binned.

.

Gold 6150 does 3.4GHz all-core turbo on 18 cores in a 165W TDP, and this is not a consumer optimized/binned part.

So minimum will be 3.4GHz.

And further to my point, the 7980XE was indeed binned a lot more aggressively than the Gold 6150, with higher turbo speeds at virtually all active core levels other than 18. Exactly what you would expect from a consumer-focused part.

 

[TheF34RChannel]

They always occupy a shared 1st place among my favourite reviewers so thanks. A good and logical breakdown of things, in my opinion.

 

[Arachnotronic]

They always occupy a shared 1st place among my favourite reviewers so thanks. A good and logical breakdown of things, in my opinion.

They made a slight error. 6-core turbo is 4.3GHz, but one-core turbo is 4.7GHz, so "enhanced turbo" for 8700K would be 4.7GHz, not 4.3GHz