Intel Skylake / Kaby Lake

[Edrick]

However, just as Bronze/Silver/Gold 5000 Xeon-SPs have only one FMAC per core

Do you have a link? I am not saying you are wrong, just that I would like to read it. Are you saying the Platinum will have 2 FMAC per core?

If the benchmarks come out, and we find out that the 8C part has AVX-512, runs in excess of 4 GHz, etc. I will eat my words.

Even half speed allows me to develop against it. And since that is my primary function, I will take what I get as long as it is supported. However, 4Ghz AVX-512 would be very nice!

Not having 6 channels will be crippling past 12C if AVX-512 is full speed. If AVX-512 is half-speed, then indeed it will be no big deal.

Many people on these forums claim that the difference between 2 channel to 4 channel is not a big deal on AVX2. Yet, you claim not having 6 with full speed AVX-512 will be crippling. I think we will need to wait for reviews of not only Skylake-X, but also Purley to make that determination.

 

[tamz_msc]

New tech should always offer better value than old tech.

6900K doesn't seem to offer better value than the 5960X, so there's an exception to your rule.

 

[Edrick]

What do YOU want the 7820X for?

I do mostly development with a lot of virtualization and some 3D rendering. Of course I play games as well, but dont need a 7820X for that.

 

[Phynaz]

You went from a 6950X to a 7700K. What use do you suddenly have now for a 7900X?

He said he wanted one, not that needed one. Why do you care?

 

[Edrick]

He said he wanted one, not that needed one. Why do you care?

Sometimes my 'want' turns into a 'need' when it comes to tech

 

[tamz_msc]

He said he wanted one, not that needed one. Why do you care?

Did he want the 6950X originally? Why do you answer for him?

I do mostly development with a lot of virtualization and some 3D rendering. Of course I play games as well, but dont need a 7820X for that.

I actually think that the 7820X is the best value in the LCC lineup, but are you in any way going to be disadvantaged due to fewer PCI-E lanes from the CPU?

 

[Ajay]

Well, cool stuffs and a bummer at the same time. No soldered TIM. That's one of the pluses of former 'E' CPUs. Prices are down, but now I'll have to shell out dollars for delidding*. I suppose many ppl won't care because they don't run overclocked 24x7 for years at a time on the same CPU. I do and I care.


* Hopefully Silicon Lottery will stick with $50 for SKX.

 

[Shlong]

I just upgraded to Ryzen 1700 but if the Threadripper 16 core is around $1200 vs 12 cores of 7920X and also offers more PCI-e lanes (64 vs 44), I'll probably go with AMD again.

 

[Edrick]

I actually think that the 7820X is the best value in the LCC lineup, but are you in any way going to be disadvantaged due to fewer PCI-E lanes from the CPU?

I will run with (2) NVMe SSDs (4 lanes X 2) and a single graphic card (16 lanes). So 28 lanes on the CPU plus 24 on the chipset gives me plenty of room.

I run that same setup today on a Z170X with only 16 lanes on the CPU and 20 from the chipset.

 

[guskline]

Getting the 7980XE is going to create contentment for me. I'll certainly be content with it until they release the next big thing.

AdamK47, you are a man with STYLE!

What Video cards will you run with it?

 

[blue11]

Do you have a link? I am not saying you are wrong, just that I would like to read it. Are you saying the Platinum will have 2 FMAC per core?

It is in my Xeon-SP product catalog. Can't say much more on that. Dual FMAC will also be present on Xeon Gold 6000-series, which begins with 16C (likely $2000).

Even half speed allows me to develop against it. And since that is my primary function, I will take what I get as long as it is supported. However, 4Ghz AVX-512 would be very nice!

I'm not sure about this, since you will need to optimize the code to the actual target, which presumably has dual FMAC (otherwise why bother). For some applications, the additional shuffling and masking modes in AVX-512 could be neat even without additional throughput.

Many people on these forums claim that the difference between 2 channel to 4 channel is not a big deal on AVX2. Yet, you claim not having 6 with full speed AVX-512 will be crippling. I think we will need to wait for reviews of not only Skylake-X, but also Purley to make that determination.

I think those claims are for general application performance. Even dual channel is already a limiting factor for the quad-core SKUs (on JEDEC speeds), yet Broadwell-EP series has 6x as many cores with only twice the bandwidth. With higher DDR4-2666 (or DDR4-3200) speeds, Skylake-X will probably be good for up to 12C in AVX-512. At 18C, probably only LINPACK and machine learning will be able to utilize dual FMACs.

 

[formulav8]

offers more PCI-e lanes

From what I've seen TR's CPU will have more lanes than Intel's. But I'm not sure if AMD has specifically confirmed or not.

 

[Edrick]

It is in my Xeon-SP product catalog. Can't say much more on that. Dual FMAC will also be present on Xeon Gold 6000-series, which begins with 16C (likely $2000).

I'm not sure about this, since you will need to optimize the code to the actual target, which presumably has dual FMAC (otherwise why bother). For some applications, the additional shuffling and masking modes in AVX-512 could be neat even without additional throughput.

I think those claims are for general application performance. Even dual channel is already a limiting factor for the quad-core SKUs (on JEDEC speeds), yet Broadwell-EP series has 6x as many cores with only twice the bandwidth. With higher DDR4-2666 speeds, Skylake-X will probably be good for up to 12C in AVX-512. At 18C, probably only LINPACK and machine learning will be able to utilize dual FMACs.

Well damn. I may have to get myself a Gold Purley after all. 6 channel and dual FMAC's will be nice. And get a dual socket MB and throw a Phi on there just because.

 

[blue11]

Well damn. I may have to get myself a Gold Purley after all. 6 channel and dual FMAC's will be nice. And get a dual socket MB and throw a Phi on there just because.

The frustrating part is the "teraflop CPU" claim which is in that territory where it just might or might not be AVX-512. It's going to be a long wait for the relevant benchmarks.

Also, Xeon Phi is not compatible with Skylake-SP boards. Even though they share a socket, Phi has its own boards (1P only). If you're not averse to HSA, I think KNL is still available as an add-in card, so you could connect one to a Skylake-X workstation.

 

[TheF34RChannel]

Those extra PCI lanes sure make the 10 core tempting. I wonder how far it will lag behind the 7820 in terms of overclock - I guess I'm going to be stalking Silicon Lottery the next few months.

Are we a collective...? I was just musing about the same thing tonight... However not for the PCI lanes (1 GPU and 1-2 M.2's and that's me done) but identical boost clocks with more cores/threads, cache. Major price difference though.

Why soldering is a bad idea..

http://overclocking.guide/the-truth-about-cpu-soldering/

It's from der8auer, not some random Intel fanboi.

Nice read, thanks for the link. Seems more useful to LN2 users than to 'normal enthusiast platform users'.

 

[AdamK47]

AdamK47, you are a man with STYLE!

What Video cards will you run with it?

Same as I have now. Two Titan Xp.

 

[Shlong]

From what I've seen TR's CPU will have more lanes than Intel's. But I'm not sure if AMD has specifically confirmed or not.

I got 64 PCI-e lanes from this video: https://www.youtube.com/watch?v=nynBEGCtg90 (at 1:50).

 

[Edrick]

Also, Xeon Phi is not compatible with Skylake-SP boards. Even though they share a socket, Phi has its own boards (1P only). If you're not averse to HSA, I think KNL is still available as an add-in card, so you could connect one to a Skylake-X workstation.

Everything I read about the Phi (Gen 2) up to this point has been that they are socket compatible so they can be used as "co-processors" in Xeon systems. I mean what the hell is the point otherwise? So only the PCIe versions can be used with Xeon systems?

I believed the only reason they used their own (1P) board thus far is because Intel hasn't release Purley yet. Do you have documentation that back this up?

 

[PaulIntellini]

Skylake-X data sheeet:

http://www.intel.com/content/www/us/en/products/processors/core/6th-gen-x-series-spec-update.html

if the data sheet is correct, then Skylake-X will indeed have FIVR.

Note that they got the cache size wrong in the beginning (copy/paste error from Broadwell-E).

 

[PaulIntellini]

Skylake-X specification update:

http://www.intel.com/content/www/us/en/products/processors/core/6th-gen-x-series-spec-update.html

- U0 stepping
- some "UPI" errata was copied from Xeon SP.

 

[dahorns]

That's no excuse for the 6900K being worse value than the 1700X at present, for example.

Ok? What does that have to do with the pricing of Skylake-X? From a business perspective it wouldn't be smart to drop the price of the 6900k prior to Skylake-X's release as Intel would be creating a competitor for its own product. Best to leave it where it is and let it slowly die out from lack of demand.

Yes, the new part from AMD represents a better value than the 6900k. The 6900k does have several things going for it over the 1700x or 1800x (OC, memory compatibility, PCIE lanes, quad-channel memory) so it isn't totally apples to apples. Even so, I would expect 6900k sales have slumped since Ryzen's release. And I think Intel is ok with that. You should be too.

 

[blue11]

Everything I read about the Phi (Gen 2) up to this point has been that they are socket compatible so they can be used as "co-processors" in Xeon systems. I mean what the hell is the point otherwise? So only the PCIe versions can be used with Xeon systems?

I believed the only reason they used their own (1P) board thus far is because Intel hasn't release Purley yet. Do you have documentation that back this up?

They probably reuse the same socket to reduce design costs. After all, Phi and SP share similar external interfaces (6 memory channels). I can't find a document that explicitly states Phi won't work in 2P Xeon boards, but the two chips have very different platform capabilities (e.g. OmniPath) and BIOS configurations (e.g. MCDRAM cache modes), so I don't see how they could be used heterogeneously in 2P configuration.

The reason Knights Landing is socketed is to enable self-hosting systems (avoid buying a CPU) and higher density (blades). The "coprocessor" solution for KNL refers to the PCI-E add-in card.

EDIT: Actually, Xeon Phi doesn't have QPI, so I don't see how this could possibly work.

 

[Atari2600]

If you don't like my posts then please put me on ignore.

@Phynaz
Consider it done.

 

[AdamK47]

You guys keep going on with this whole AMD vs Intel thing in the same pointless endeavor happening ever since the Internet existed. No conclusion is ever met. Nobody ever wins. It's the whole endless loop of wasted text over and over again.

What I say next may come as a shock, but some people actually buy hardware instead of just talk about it. Shocking, right?

What I'd like to know is when exactly are the CPUs and motherboards going to be available for purchase? I hope the Rampage VI Extreme releases at the same time. What cooling options are there at release?

 

[Arachnotronic]

You went from a 6950X to a 7700K. What use do you suddenly have now for a 7900X?

I buy hardware because I like hardware, I make no pretense that I buy this stuff because I "need" it.

It may come as a shock to you, but hobbyists don't buy things necessarily because they need them

Much cheaper hobby than smoking, gambling, or any of those other things that I don't do!