Who's buying Skylake-X? (You may now change your vote)

[Ajay]

@blue11 How many HEDT buyer do you think don't have at least an enthusiast grade GFX card? Really, this is a kind of silly comparison. Two different classes of buyer in most cases. Also, you know that in computers performance doesn't typically scale with cost. 'Cutting edge' tech follows a power curve, not a linear one.

 

[blue11]

@blue11 How many HEDT buyer do you think don't have at least an enthusiast grade GFX card? Really, this is a kind of silly comparison. Two different classes of buyer in most cases. Also, you know that in computers performance doesn't typically scale with cost. 'Cutting edge' tech follows a power curve, not a linear one.

For GPU workstations, performance is measured by GPU utilization, not CPU throughput, so it never makes sense to upgrade the CPU. Besides which, if you have a pair of $600 GTX 1080 Ti, the marginal increase in CPU cost will be quite minor.

My point was that the 8C will be perceived as poor value, because on top of the increase in CPU price, it also incurs the platform cost increase, whereas going from 8C to 10C, that platform cost will have already been paid. The key distinction in the hypothetical next generation is that the desktop to baseline HEDT transition only offers 33% performance instead of 50%. You can already see from the previous calculations that going to 10C will actually be better cost efficiency than going to 8C, assuming pricing follows the current model.

Previous lineup:
4C->6C (+50%) + 6C->8C (+33%) = +100%

New lineup:
6C->8C (+33%) + 8C->10C (+25%) = +67%

So we see that going to 10C is closer to the bump in performance that HEDT provided in previous generations.

 

[wildhorse2k]

Skylake-X will likely ship at 3-3.5 GHz stock frequency, just like Zen, but we already know the OC limits for both chips. The thing about the 8-core SKU, assuming it replaces the 6850K at $500-600, and assuming a 6C Coffee Lake at $350-400, is that it will cost a lot more for only 33% maximum performance gain. On top of the actual chip costing ~50% more, the rest of the system becomes more expensive due to a platform change (more expensive boards, twice the DRAM, and likely idle power draw increase).

The platform change causes the system cost to increase by 77% despite performance only increasing by up to 33% (marginal cost-to-performance ratio of 2.3). On the other hand, if the i9-7900X ($1000 est.) were used in place of the i9-7820X, the cost for the HEDT system would rise to $1637, or 144% cost for 67% performance (marginal cost-to-performance ratio of 2.1). Going to i9-7920X ($1700 est.) is a different story of course.

Skylake-X frequencies have already been leaked, I think they are either in this or other topics. 10C was supposed to have 3.3Ghz base frequency. The 10C may OC reasonably well but I expect it will reach TDP limit before reaching frequency limit, just as 12C. So the theoretical top frequency will mostly not matter. The 8C will be best for OC. I don't think price will be the same as for 6850K, I don't think Intel will be that generous. I expect it to be about $150 higher.

Platform change is not only about CPU performance, but also M.2 performance which is not good on Z270 when you need more of them, more RAM, 4 channel memory controller, more PCIe lanes. If someone needs only faster CPU then it's going to be an expensive upgrade.

 

[blue11]

Skylake-X frequencies have already been leaked, I think they are either in this or other topics. 10C was supposed to have 3.3Ghz base frequency. The 10C may OC reasonably well but I expect it will reach TDP limit before reaching frequency limit, just as 12C. So the theoretical top frequency will mostly not matter. The 8C will be best for OC. I don't think price will be the same as for 6850K, I don't think Intel will be that generous. I expect it to be about $150 higher.

If the 8C costs more, it will make it even less attractive of a proposition against 6C Coffee Lake, unless the CFL also costs $500. It is really Coffee Lake pricing that is the more unclear of the two.

There is no reason to expect TDP to limit OC, as it has never done so historically, and certainly does not stop people from overclocking existing HEDT CPUs to the limits of (electrical) stability.

Platform change is not only about CPU performance, but also M.2 performance which is not good on Z270 when you need more of them, more RAM, 4 channel memory controller, more PCIe lanes. If someone needs only faster CPU then it's going to be an expensive upgrade.

These "advantages" are mainly perceived and not actual, although they may be indeed important to HEDT buyers (as status symbols). PCI-E x4 is already fast enough for M.2 SSD, and PCI-E switchs can be used to add more. There is no realistic use case where every SSD in a multi-SSD system is running at maximum sequential bandwidth. Quad-channel memory increases bandwidth, but more cores increases bandwidth consumption, so there is no net gain (yet latency increases, making it actually worse). The need for PCI-E lanes (e.g. because of multi-GPU) is basically imaginary.

 

[pj-]

My 5820k based system is approaching 2 years of age and I am curious to see if anything coming out this year will make me want to upgrade. I run at 4.2ghz since anything higher requires voltages I don't care for, so I'd be looking to get 8 or more cores that can easily hit 4.7ghz+ in order to justify the upgrade to myself.

 

[Edrick]

I can't wait to see how bottle necked a 6C/12T CPU is going to be on a dual channel memory configuration. We already know the 7700k is memory starved and requires some RAM OCing to see maximum performance. I really think Sky-X is going to shine on a quad channel system (much more than previous generations).

 

[LTC8K6]

My 5820k based system is approaching 2 years of age and I am curious to see if anything coming out this year will make me want to upgrade. I run at 4.2ghz since anything higher requires voltages I don't care for, so I'd be looking to get 8 or more cores that can easily hit 4.7ghz+ in order to justify the upgrade to myself.

SL-X is probably considerably faster than Haswell-E.

Those are your logical Intel choices:

Core i9-7820X
8C/16T
11MB L3
28 PCIe Lanes
3.6Ghz Base
4.3Ghz Turbo 2.0
4.5Ghz Turbo 3.0

Core i9-7800X
6C/12T
8.25MB L3
28 PCIe Lanes
3.5Ghz Base
4.0Ghz Turbo 2.0

I think easily getting to 4.7+ is quite a stretch for either and I wouldn't expect it. The base and turbo frequencies are up nicely from HW-E and BW-E.

 

[wildhorse2k]

duplicate

 

[LTC8K6]

I can't wait to see how bottle necked a 6C/12T CPU is going to be on a dual channel memory configuration. We already know the 7700k is memory starved and requires some RAM OCing to see maximum performance. I really think Sky-X is going to shine on a quad channel system (much more than previous generations).

We haven't seen a leak of any 6C/12T desktop chip yet. We have only seen a 6C6T leak.

The 300 series chipset might have a better mem controller, along with higher stock mem speed of 2666.

 

[wildhorse2k]

These "advantages" are mainly perceived and not actual, although they may be indeed important to HEDT buyers (as status symbols). PCI-E x4 is already fast enough for M.2 SSD, and PCI-E switchs can be used to add more. There is no realistic use case where every SSD in a multi-SSD system is running at maximum sequential bandwidth. Quad-channel memory increases bandwidth, but more cores increases bandwidth consumption, so there is no net gain (yet latency increases, making it actually worse). The need for PCI-E lanes (e.g. because of multi-GPU) is basically imaginary.

With HEDT you could use it as server and play games on it at the same time and not notice it. Or have some calculations running. It has big reserves. Z270 architecture is wrong, M.2 doesn't belong to chipset as it can easily overload DMI 3.0 (single M.2 can fully load it). AMD got it right in X370 and X99 got it right too. On X99 you can copy data from one M.2 drive to another at full speed. PCIe switches add latency, not good for gaming. Boards having them are expensive too. They are mostly useful to allow other IO cards to share bandwidth with GPU. Quad channel is still better than the 2 channel Ryzen gets. Usually little practical advantage is seen, as single core cannot benefit from 4 channel memory controller. Only simultaneous usage of all cores can benefit from it and only in certain loads. So benchmarks show little benefit. PCI-E lanes are there not only for GPUs, but also other cards. You could also use independent GPUs for computations at x8 (4x GPU), or even play multiple games at the same time on a single machine.

Most people do not need those features so they go for the cheaper Z270/X370 instead. I would expect an average HEDT customer to go for it for more reasons than just CPU power, so your argument doesn't apply as the gain is bigger than you think. It's not something imaginary in my opinion.

 

[pj-]

SL-X is probably considerably faster than Haswell-E.

Those are your logical Intel choices:

Core i9-7820X
8C/16T
11MB L3
28 PCIe Lanes
3.6Ghz Base
4.3Ghz Turbo 2.0
4.5Ghz Turbo 3.0

Core i9-7800X
6C/12T
8.25MB L3
28 PCIe Lanes
3.5Ghz Base
4.0Ghz Turbo 2.0

I think easily getting to 4.7+ is quite a stretch for either and I wouldn't expect it. The base and turbo frequencies are up nicely from HW-E and BW-E.

I'm fine not upgrading if those speeds aren't attainable. Although, I hope it's not a big stretch considering even my mediocre specimen 5820k had 600mhz of OC headroom on top of the boost clock. In addition, I saw a lot of people getting 4.3-4.4ghz on similar voltages. The 7820k would only need 400mhz OC to get what I'm looking for.

I understand it's more cores and that intel may be clocking more aggressively out of the box, but I would still expect to be able to get something extra out of it.

 

[TheF34RChannel]

That would the joke of 2017, only releasing 6C6T for the mainstream.

That would be a disappointment, and I wouldn't put it past Intel.

It would steer me towards Skylake-X immediately. I also read someone saying you need double the RAM? Depending on what you're using the system for I'd agree with 32GB for workstation use but if it's gaming you're after (edit: without streaming maybe) 16 should be just fine; I know of plenty of people using just that currently.

 

[DrMrLordX]

We already know the 7700k is memory starved and requires some RAM OCing to see maximum performance.

Actually, the real gains on Skylake/Kabylake (if I recall correctly) came from speeds of DDR4-2400 and higher; after that point, you could tell it was latency reduction that improved performance what with the tendency for performance to stay about the same when moving CAS/CL upwards in increments of 1 while also increasing effective memory clockspeed by 266 MHz. Tight timings with DDR4-2133 did not yield proper performance scaling, indicating that 4c/8t Skylake/Kabylake - at least at stock speeds - had a bandwidth threshold somewhere between the speeds of DDR4-2133 and DDR4-2400.

I would be surprised if Coffelake 6c/12t CPUs really required anything more than DDR4-3200.

 

[frozentundra123456]

I was discussing the winner of best gaming CPU between CFL and SKL-X. If you want to throw Ryzen into the mix be my guest.

Wasn't referring to your posts at all. I was replying to the poster who flooded this thread with Ryzen posts and claimed skylake X was going to be rendered irrelevant

 

[TheF34RChannel]

I would be surprised if Coffelake 6c/12t CPUs really required anything more than DDR4-3200.

^ this; if not benchmarking, which seems to favour high frequency rather than lower timings, you seem to be better off with 3200MHz and tighter timings than vice versa (higher frequency and looser timings).

Re-raising the question though; why haven't there been 6C/12T Coffee Lake leaks yet... It'll begin to look suspicious soon... If none pop up in June/July I'll start to think they will have no HT part...

 

[DrMrLordX]

Intel usually leaks early, especially when they have something to prove. We even had Broadwell-C leaks.

 

[jpiniero]

Re-raising the question though; why haven't there been 6C/12T Coffee Lake leaks yet... It'll begin to look suspicious soon... If none pop up in June/July I'll start to think they will have no HT part...

Remember they are rushing this out; so maybe the validation cycle is really compressed.

 

[VirtualLarry]

I think, if they released a 6C/6T CFL at the high end... that it would be mostly irrelevant. I mean, AMD released Thuban, a 6C/6T, how many years ago? Ryzen offering 6C/12T for cheaper, I mean, it's kind of a no-brainer, that leave the customer choosing between Ryzen and SKL-X. Or maybe that's Intel's point? That the CFL 6C/6T SKU is a "dummy" SKU, intended solely to push their customers to SKL-X?

 

[Edrick]

The 300 series chipset might have a better mem controller, along with higher stock mem speed of 2666.

The IMC is in the CPU, not the chipset. Don't see how the 300 series can improve anything.

 

[Edrick]

Anyone else notice that the leaked MBs for dual socket 3647 only have 6 total mem channels? 3 per socket. Are we not getting true 6 channel systems?

 

[blue11]

Anyone else notice that the leaked MBs for dual socket 3647 only have 6 total mem channels? 3 per socket. Are we not getting true 6 channel systems?

To what image are you referring? I found on Google this Gigabyte 2U server from Computex 2016 with 12 DIMM slots per socket for 2DPC. Some workstation boards might only have 6 DIMM slots per socket for space reasons.

 

[coercitiv]

Wasn't referring to your posts at all. I was replying to the poster who flooded this thread with Ryzen posts and claimed skylake X was going to be rendered irrelevant

Quoting my post did the trick though, in the sense of seeing it as some kind of reply.

Anyways, my small and insignificant bet is SKL-X is going to be the one stealing Coffee Lake's thunder, not the other way round. I think we tend to bet too much on things we can more easily understand, such as better production nodes leading to performance improvements, while also minimizing the potential impact of things we have less knowledge about, such as a radical change in the cache system.

A long time forum member pointed out Intel would not be doing this unless they expected a decent increase in performance. I subscribe, 2017 is not done being disruptive

 

[LTC8K6]

In the other thread, SL-X appears to run rings around BW-E in the SiSoft battle...

 

[LTC8K6]

The IMC is in the CPU, not the chipset. Don't see how the 300 series can improve anything.

You got me there...

 

[Edrick]

To what image are you referring? I found on Google this Gigabyte 2U server from Computex 2016 with 12 DIMM slots per socket for 2DPC. Some workstation boards might only have 6 DIMM slots per socket for space reasons.

Here: https://ecauk.com/files/2017/05/Advantech-ASMB-825-Datasheet.pdf

And Here: http://oi1024.photobucket.com/albums/y303/martmail55/DASBjheUAAAqQ-S_zpsxuynq4ct.jpg

 

[blue11]

Here: https://ecauk.com/files/2017/05/Advantech-ASMB-825-Datasheet.pdf

And Here: http://oi1024.photobucket.com/albums/y303/martmail55/DASBjheUAAAqQ-S_zpsxuynq4ct.jpg

It looks like there isn't enough room on that board physically for 6 DIMMs per socket. If you notice, that is an ATX form factor board, which is rather compact for a dual socket server.