Intel Core-X Series: Full Specs Revealed

[Taf the Ghost]

Actually attached the image so it loads.

 

It would appear the lowest SKU of both dies for Skylake-X are the weakest silicon. (Makes sense.) Thus the base clock being low enough to ensure operation.

 

I'm honestly curious what the upper limit on power draw is for the socket, as we should find out with the 18c part.

[Drak3]

45 minutes ago, 3DOSH said:

And for the people putting TR against the highest end ship from Intel, your insane. TR is 1K $ so its competition should be the nearest prices chip from Intel, SO the X7900 or X7920. 

All CPU comparisons are fair. To pin the 1950X against the the 7980XE because they're the top dogs is fair. To pin the 1950X against the 7960X because they've got the same core counts is fair. To pin it against the earliest Pentium that is still on sale, is fair.

[Dylanc1500]

4 minutes ago, Drak3 said:

All CPU comparisons are fair. To pin the 1950X against the the 7980XE because they're the top dogs is fair. To pin the 1950X against the 7960X because they've got the same core counts is fair. To pin it against the earliest Pentium that is still on sale, is fair.

So can I use an 8086? They are all x86.

 

joking.

[Drak3]

Just now, Dylanc1500 said:

So can I use an 8086? They are all x86.

 

joking.

Yes, absolutely. It's a fair comparison.

Not one that is very realistic, but fair.

[Dylanc1500]

Just now, Drak3 said:

Yes, absolutely. It's a fair comparison.

Not one that is very realistic, but fair.

I honestly wish some would see how many magnitudes of time it would take to do a set of work between an 8086 vs 7980xe.

[Drak3]

Just now, Dylanc1500 said:

I honestly wish some would see how many magnitudes of time it would take to do a set of work between an 8086 vs 7980xe.

I'm sure that any task that could run on the 8086 would stop scaling before requiring the power of a single 7980XE core @ 1GHz.

[Techstorm970]

7 hours ago, Fonzie92 said:

More like every single area, LOL.

 

Edit: LMFAO @ Zen2 competitor. No it wouldn't. It'd compete DIRECTLY with Ryzen. It's launching next month.

 

No, Coffee Lake is launching in December or January.

 

https://www.pcgamesn.com/intel/intel-14nm-coffee-lake-release-date

[suits]

I'm really interested to see how they overclock. The base to boost clock gap is freaking huge and makes me wonder if even though they are unlocked you can truly overclock past the boost

[suits]

3 minutes ago, techstorm970 said:

No, Coffee Lake is launching in December or January.

 

https://www.pcgamesn.com/intel/intel-14nm-coffee-lake-release-date

And Zen 2 should be coming around in early 2018. That's the impression AMD gave anyway. We'll see.

[3DOSH]

43 minutes ago, Drak3 said:

All CPU comparisons are fair. To pin the 1950X against the the 7980XE because they're the top dogs is fair. To pin the 1950X against the 7960X because they've got the same core counts is fair. To pin it against the earliest Pentium that is still on sale, is fair.

sorry I meant stupid

[Drak3]

Just now, 3DOSH said:

sorry I meant stupid

It's not a stupid comparison either. Top dog vs top dog.

[3DOSH]

Just now, Drak3 said:

It's not a stupid comparison either. Top dog vs top dog.

What is the point then ????? To see who has the biggest E-penis. With that logic car manufacturers should compare the best LAMBO to the Best Honda 

[Drak3]

Just now, 3DOSH said:

What is the point then ????? To see who has the biggest E-penis. With that logic car manufacturers should compare the best LAMBO to the Best Honda 

To determine which processor is the better choice for different workloads that benefit from 16+ cores.

[3DOSH]

4 minutes ago, Drak3 said:

To determine which processor is the better choice for different workloads that benefit from 16+ cores.

why not EPYC since money is not problem then, that has 32 cores ...

[Drak3]

Just now, 3DOSH said:

why not EPYC since money is not problem then 

Intel has AVX512, and we don't know if EPYC is overclockable.

[3DOSH]

Just now, Drak3 said:

Intel has AVX512, and we don't know if EPYC is overclockable.

Shouldn't matter snice its top dog vs top dog. Your words... 

[Drak3]

Just now, 3DOSH said:

Shouldn't matter snice its top dog vs top dog. Your words... 

When comparing processors, you compare features and performance of said processors, and decide what is best based on use case, and what is best for you based on your use case and budget.

If I benefit from heavily AVX512, every single AMD offering goes down in value significantly, to the point that the 7900X becomes a much better choice than the 1950X.

If I benefit from overclocking and PCIe lane count higher than 44 (but not over 60), then I benefit from Threadripper, not EPYC.

If I benefit from 65+ PCIe lane count, then I benefit from EPYC.

If I benefit from high single core performance, more than any Zen chip can offer, and good multicore performance, I benefit from X299.

 

Each of those are comparisons that are grounded in the real world, and each offer a different answer as to which processor is the best for what I want and/or need.

 

Ignoring if EPYC will even be available for consumer purchase at time of launch.

 

But we seldom get that. We get point blank, poorly thought out, game meme benchmarks from 'reviewers' quick to declare and dismiss processors as being all around good or bad without giving full consideration to the strengths and weaknesses of both.

 

And btw, the 1950X and 7980XE are the top dogs of HEDT. They mix things from the server market and the consumer market. They're not 100% interchangeable with products from either market.

[MageTank]

Just now, Drak3 said:

Intel has AVX512, and we don't know if EPYC is overclockable.

Intel's AVX512 wasn't enough to save them from EPYC's massive amounts of bandwidth, lol. AMD proved that AVX can be brute forced with enough ram bandwidth:

http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/21

 

Even on Intel ICC, it can still lift, lol. 

Quote

For our first shot with this benchmark, we used version 2.10 to be able to compare to our older data set. Version 2.12 seems to make better use of "Intel's compiler vectorization and auto-dispatch has improved performance for Intel processors supporting AVX instructions". So let's try again:

Quote

The older Xeons see a perforance boost of about 25%. The improvement on the new Xeons is a lot lower: about 13-15%. Remarkable is that the new binary is slower on the EPYC 7601: about 4%. That simply begs for more investigation: but the deadline was too close. Nevertheless, three different FP tests all point in the same direction: the Zen FP unit might not have the highest "peak FLOPs" in theory, there is lots of FP code out there that runs best on EPYC.

I gotta say, the FP performance caught me off guard. Normally, AMD wins with Ints, but loses with floats. This time however, that role seems to have reversed, even with AVX being involved (remember, AVX512 can do twice the AVX 2 bit-ops/clock, and 4x AVX 1). 

[Drak3]

Just now, MageTank said:

Intel's AVX512 wasn't enough to save them from EPYC's massive amounts of bandwidth, lol. AMD proved that AVX can be brute forced with enough ram bandwidth:

http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/21

 

Even on Intel ICC, it can still lift

 

I gotta say, the FP performance caught me off guard. Normally, AMD wins with Ints, but loses with floats. This time however, that role seems to have reversed, even with AVX being involved (remember, AVX512 can do twice the AVX 2 bit-ops/clock, and 4x AVX 1). 

Leadeater already showed me that link. I did some math in that thread, and my conclusion: It takes AMD throwing bandwidth at the processor to get AVX performance to a good point, but it's not perfect or really exceptional. Once you get to the last two benchmarks, which are operating with DRAM, the numbers start aligning with core count and the offset. Quite frankly, it takes quite a bit to get the Zen architecture off its ass in AVX performance, and TR doesn't do it.

 

Which is why I personally hope EPYC is overclockable, as it'd stand to be a much better HEDT platform when AVX performance is where it should be, and the PCIe lane count is nice too.

[MageTank]

46 minutes ago, Drak3 said:

Leadeater already showed me that link. I did some math in that thread, and my conclusion: It takes AMD throwing bandwidth at the processor to get AVX performance to a good point, but it's not perfect or really exceptional. Once you get to the last two benchmarks, which are operating with DRAM, the numbers start aligning with core count and the offset. Quite frankly, it takes quite a bit to get the Zen architecture off its ass in AVX performance, and TR doesn't do it.

 

Which is why I personally hope EPYC is overclockable, as it'd stand to be a much better HEDT platform when AVX performance is where it should be, and the PCIe lane count is nice too.

The last bench is with Intel optimized code. The fact that it's brute forcing still held up, is nothing short of impressive in my eyes. Even factoring in core counts, they are still technically at a disadvantage by about 4x the bit-ops/clock. It should also be noted that EPYC was tested at 2400mhz ram, which impacts the IF quite a bit. If AMD can improve upon their memory support, and allow for faster ram (even if it's only 2666mhz over 2400), you will see a slight improvement from that as well, especially considering this is 8 channels we are talking about. Peak theoretical memory bandwidth of DDR4 2400 in Octa-channel is 153.6GB/s. Peak theoretical memory bandwidth of DDR4 2666 in Octa-channel is 170.6GB/s. That's a difference of almost exactly 10% bandwidth. Considering how well AVX scales with bandwidth, I'd say it would be enough to improve the numbers. I should also point out, this is per socket. Having 2 CPU's would double this. 

 

The Zen architecture as a whole, would be far scarier if it had a more potent IMC. I sincerely hope Zen 2.0 delivers exactly in that respect. 

[Drak3]

4 minutes ago, MageTank said:

The Zen architecture as a whole, would be far scarier if it had a more potent IMC. I sincerely hope Zen 2.0 delivers exactly in that respect

Same. It'd be a damn near Intel killer if AMD focused only on the IMC and upping the clocking potential.

[AnonymousGuy]

Has Intel ever published the specs on the LGA pins?  I looked for 15 minutes but couldn't find anything. If I were to estimate each contact to be equivalent to 24awg wire, that's maybe 3A max per contact, or about 600A total for the package @ 2V VCCin?

[DocSwag]

14 hours ago, Darel321 said:

I said it one time and it's still true in my opinion, from the release of core2 series intel stopped developing new tech, yes they perfected the cpu they got to the limits.. but still they didnt make anything ''realy'' new... and now people are starting to realize its true... its just sad that with the billions they gain with their overpriced cpus they just cant make anything new like amd.. and actually they never have...

If you ask me nehalem and Sandy bridge were pretty big changes. At this point there really isn't a big space to innovate in the CPU ipc department, really. What amd did with Ryzen was just catch up with Intel on the ipc department. They adopted a lot of things Intel did ages ago (micro op cache for example, that was introduced on Sandy bridge)

[porina]

13 hours ago, MageTank said:

Intel's AVX512 wasn't enough to save them from EPYC's massive amounts of bandwidth, lol. AMD proved that AVX can be brute forced with enough ram bandwidth:

http://www.anandtech.com/show/11544/intel-skylake-ep-vs-amd-epyc-7000-cpu-battle-of-the-decade/21

 

[snip]

 

I gotta say, the FP performance caught me off guard. Normally, AMD wins with Ints, but loses with floats. This time however, that role seems to have reversed, even with AVX being involved (remember, AVX512 can do twice the AVX 2 bit-ops/clock, and 4x AVX 1). 

I'm only familiar with AVX2 workloads comparable to Prime95 (so could include things like Linpack). Based on my observations to date, for larger working sets (big FFT size) Intel systems can often be ram bandwidth limited. It depends on how much total CPU demand there is to available ram bandwidth, with cache optimisation thrown in as an extra variable. The extra potential of AVX512 may be limited in potential due to this. Roughly speaking I'm working on needing about double the Ryzen cores as Intel ones to equalise, but if ram bandwidth is limiting, look at ram bandwidth!

 

For smaller size workloads then peak performance can be met.

11 hours ago, AnonymousGuy said:

Has Intel ever published the specs on the LGA pins?  I looked for 15 minutes but couldn't find anything. If I were to estimate each contact to be equivalent to 24awg wire, that's maybe 3A max per contact, or about 600A total for the package @ 2V VCCin?

I wonder if you can estimate this. Look up the socket total max current, and divide by the number of power pin pairs.

[MageTank]

1 hour ago, porina said:

I'm only familiar with AVX2 workloads comparable to Prime95 (so could include things like Linpack). Based on my observations to date, for larger working sets (big FFT size) Intel systems can often be ram bandwidth limited. It depends on how much total CPU demand there is to available ram bandwidth, with cache optimisation thrown in as an extra variable. The extra potential of AVX512 may be limited in potential due to this. Roughly speaking I'm working on needing about double the Ryzen cores as Intel ones to equalise, but if ram bandwidth is limiting, look at ram bandwidth!

 

For smaller size workloads then peak performance can be met.

I wonder if you can estimate this. Look up the socket total max current, and divide by the number of power pin pairs.

They most certainly were ram bandwidth limited. They were using hexa-channel memory at 2666 (127.9GB/s x 1 CPU) vs EPYC's octa-channel 2400 (153.6GB/s x 1 CPU).

 

It should also be noted that Skylake Purely's per-core memory bandwidth was extremely weak, at 12GB/s, vs even Broadwell's 18.5GB/s bandwidth per-core. AMD delivered a massive 27GB/s per core, but suffered a limit of 30GB/s for each quad core CCX. No idea how that will impact specific workloads.