Intel Skylake / Kaby Lake

[hawtdawg]

Even SB and IB CPU's benefit massively from increased bandwidth on newer games.

http://www.eurogamer.net/articles/d...it-finally-time-to-upgrade-your-core-i5-2500k

you gain the same performance from upping memory speeds as you do from OCing the processor in several cases, even on older intel CPU's

 

[imported_jjj]

So, in one hand, quad channel RAM (increased bandwidth) doesn't matter because of past experience (pre-Skylake). Completely disregarding the fact the Skylake was the first Intel CPU to memory bandwidth starved, and saw large gains in performance with greater memory bandwidth. On the other hand, you state Ryzen requires the as much memory bandwidth as possible because it is different from past AMD CPUs.

So have you seen information/test results regarding how Skylake-X (X299) handles quad channel memory that the rest of us have not? I generally do not like to assume, but I think it is a safe bet that the test results will differ from Haswell/Broadwell simply from the fact that Skylake was much different.

After reading all your posts, it is clear that you are an AMD fanboy and will do your best to continuously move the goalposts to fit your agenda.

You are twisting his statements.
In the second quote he states that with dual channel more BW is good when you got 8 cores while the second part on his statement is about latency not bandwidth. AMD's interconnect is linked to memory clocks so latency decreases as you clock the DRAM higher and this is why DRAM oC is more rewarding than on Intel. It's not just inter-CCX that has lower latency though, memory latency scales better too.
Maybe also notable that AMD has higher latency but higher bandwidth than Intel and it's safe to assume that remains true with Skylake X. AMD hits very close to theoretical bandwidth so you can't really do better.
The number of memory channels is not about just bandwidth. Cost ( both die size and system costs) and power are highly important and then there is the actual need for more channels.

AMD with 8 cores and 4GHz is fine with the DRAM at 3200 ,especially in consumer.Intel with higher IPC, not that much higher though when you factor in SMT, it's not really limited in consumer with 8 cores bellow 4GHz either. With more cores and/or higher clocks they would start to be limited in some scenarios with current DDR4 speeds-prices are high and above 3200 they really get too high.
Intel with quad channel for 12-18 cores is really starving them though, we'll see how high they clock them as the TDP will be limiting but there is no way out really.
I'll assume you understand what the theoretical limit is for bandwidth and that it's a hard limit , you can't do better than that at a given memory speed.
Maybe if DDR4 4400 becomes more affordable (slim chance as DRAM are still rising but might stabilize towards the end of the year), it gets better but we'll have to see how Skylake X scales with memory or if scaling stops well bellow those speeds.

 

[itsmydamnation]

So, in one hand, quad channel RAM (increased bandwidth) doesn't matter because of past experience (pre-Skylake). Completely disregarding the fact the Skylake was the first Intel CPU to memory bandwidth starved, and saw large gains in performance with greater memory bandwidth. On the other hand, you state Ryzen requires the as much memory bandwidth as possible because it is different from past AMD CPUs.

Hardly, separate out LATENCY from more bandwidth, comparing 2133 15-15-15 to 3200 15-15-15 is a 55% reduction in latency. The problem is review sites have done a crap job at testing because they never compare different memory speeds at the same Latency (ns).

There is always going to be corner cases but generally quad cores aren't memory bottleneck with dual channel. Think what would be the point in 22 core Xeon's.

Now Zen liking memory bandwidth has nothing to do with the memory bandwidth and everything to do with the internal data fabric running at a speed that is derived from the memory clock. Before 3200 you see perf gains from either lower latency or higher clock but once past 3200 extra bandwidth doesn't really increase performance only lower latency does and in the worse case it can do it quite a bit.

 

[Edrick]

Hardly, separate out LATENCY from more bandwidth, comparing 2133 15-15-15 to 3200 15-15-15 is a 55% reduction in latency. The problem is review sites have done a crap job at testing because they never compare different memory speeds at the same Latency (ns).

Agreed that review sites have done a crap done up to this point.

There is always going to be corner cases but generally quad cores aren't memory bottleneck with dual channel. Think what would be the point in 22 core Xeon's.

We are talking about Skylake-X vs. Threadripper. My point is that Skylake-X will be memory bottlenecked with dual channel and we most likely will see noticeable gains from quad channel. Hell, there is a reason Intel is going 6 channel for their high end Xeons. Simply discounting the benefit of quad channel on a CPU that we have not seen any tests on as of yet is shortsighted at best.

Now I may eventually be proven wrong when the NDA is lifted. But making claims to discredit Intel X299 platform without the facts in hand shows his bias.

 

[tamz_msc]

We are talking about Skylake-X vs. Threadripper. My point is that Skylake-X will be memory bottlenecked with dual channel and we most likely will see noticeable gains from quad channel. Hell, there is a reason Intel is going 6 channel for their high end Xeons. Simply discounting the benefit of quad channel on a CPU that we have not seen any tests on as of yet is shortsighted at best.

Now I may eventually be proven wrong when the NDA is lifted. But making claims to discredit Intel X299 platform without the facts in hand shows his bias.

Every point I made is applicable when comparing X99 to the existing Ryzen X370 platform, which was the context of the posts.

Unless your applications are regularly running out of memory bandwidth, which means that it's I/O bound, 50GB/s aida64 read bandwidth with 3200 MHz+ at CL14 is plenty for most people.

Running a SQL server, multiple VMs? Quad channel might be useful. Other than those, what else?

 

[Zucker2k]

Every point I made is applicable when comparing X99 to the existing Ryzen X370 platform, which was the context of the posts.

Unless your applications are regularly running out of memory bandwidth, which means that it's I/O bound, 50GB/s aida64 read bandwidth with 3200 MHz+ at CL14 is plenty for most people.

Running a SQL server, multiple VMs? Quad channel might be useful. Other than those, what else?

Those aren't enough? How about all multi-tasking scenarios you can think of, for starters? Like gaming while a crap load of other apps are running in the background without breaking a sweat? Which happens to be what certain fanboys always bring up when it comes to Ryzen vs the Z270 platform.

Insulting other members is not allowed.
Markfw
Anandtech Moderator

 

[tamz_msc]

Those aren't enough? How about all multi-tasking scenarios you can think of, for starters? Like gaming while a crap load of other apps are running in the background without breaking a sweat? Which happens to be what certain fanboys always bring up when it comes to Ryzen vs the Z270 platform.

That has got more to do with the amount of memory than the number of channels. A 32GB quad-channel kit on a 6900K is rarely going to be faster than a dual-rank 32GB dual channel kit on a 1800X.

 

[Edrick]

Every point I made is applicable when comparing X99 to the existing Ryzen X370 platform, which was the context of the posts.

Unless your applications are regularly running out of memory bandwidth, which means that it's I/O bound, 50GB/s aida64 read bandwidth with 3200 MHz+ at CL14 is plenty for most people.

Running a SQL server, multiple VMs? Quad channel might be useful. Other than those, what else?

Not sure why you are comparing X99 and X370 (especially in a Skylake thread). However, you were replying to a X299 (Skylake-X) post when you discounted quad channel memory (and AVX-512, and additional PCIe lanes). And until we see actually numbers, all you have are assumptions which you are using to paint a certain picture.

Again, I could be wrong and we may see zero difference with quad channel memory. But the only fact here is that no one knows for certain yet (except people under NDA).

Also, you are correct in that for many applications, people may not "see" a difference. And in your own interests, that may also be true. But many of us here are not just gamers, and leverage technologies that will benefit from increased bandwidth on a daily basis.

Its great that you are an AMD fan. But please stop clouding the issues here in the Skylake thread.

 

[krumme]

AVX2 has proven very useful in handbrake and would probably get adopted more in content creation going forward. It's advantages can't be ignored for too long now.

Yes H265 in handbrake is a very nice tool for using that avx2 ability. The problem as i can tell is an oc 1700 vs a oc 6800 is about the same perf for h265 encoding. For similar price as the 1700 the 7700 is slower. What is it worth then?

It illustrates the basic problem that sklx will also face vs tr for that type of workload that uses avx2. A highend 16c tr will probably be faster than a 12c sklx.

Thats cherry picking a load where core line is strong and its barely on par for the price. But take a normal load like blender/cb and its just a mess. Granted there is performance situations where is faster for the price but its literally like 1:1000 for productivity.

And as i can tell Avx2 have been here for years. Its where it is and will be for the forseable future. The consoles got avx. Dont know if that help for avx adoption in gaming. It is going dog slow unfortunately. One reason is perhaps like the one explained by virtual larry. Why develop for something that is not widely used?

 

[Edrick]

And as i can tell Avx2 have been here for years. Its where it is and will be for the forseable future. The consoles got avx. Dont know if that help for avx adoption in gaming. It is going dog slow unfortunately. One reason is perhaps like the one explained by virtual larry. Why develop for something that is not widely used?

Exactly. Developers are not leveraging it because not everyone has a CPU which can take advantage of it. At least that is one major reason.

Eventually, AVX2 and such will be more widespread. Slowly.

 

[vissarix]

You can tell Skylake X might be a huge success when the entire Amd army is on this thread desperately trying to downplay it

 

[tamz_msc]

Not sure why you are comparing X99 and X370 (especially in a Skylake thread). However, you were replying to a X299 (Skylake-X) post when you discounted quad channel memory (and AVX-512, and additional PCIe lanes). And until we see actually numbers, all you have are assumptions which you are using to paint a certain picture.

Again, I could be wrong and we may see zero difference with quad channel memory. But the only fact here is that no one knows for certain yet (except people under NDA).

Also, you are correct in that for many applications, people may not "see" a difference. And in your own interests, that may also be true. But many of us here are not just gamers, and leverage technologies that will benefit from increased bandwidth on a daily basis.

Its great that you are an AMD fan. But please stop clouding the issues here in the Skylake thread.

Like I said before, quad-channel memory is useful to only a small subset of the target audience of X299/X99, and AVX512 is useful for an even smaller subset.

You are certain that your use case means those features would be put to use? No problem, but there are lots of people who have absolutely no use of said features and constantly harp on their presence making Intel HEDT the superior platform by default.

Products don't exist in a vacuum, and with the options that are available now, it no longer makes sense to talk about rigid distinctions like HEDT and mainstream.

 

[coercitiv]

Again, I could be wrong and we may see zero difference with quad channel memory. But the only fact here is that no one knows for certain yet (except people under NDA).

If SKL-X vs. SKL-S shows a considerable memory bottleneck on the latter, what do you make of CFL 6c/12t running dual channel only?

 

[beginner99]

Content creation is where Sky-X is going to stretch it's legs the most, in my opinion, with it's high low-thread turbo, many cores, and quadchannel, and AVX-512, it's got content creation covered.

Fast RAM on Ryzen is because inter-ccx communication depends on memory speed and it's a limitation. So faster RAM = faster infinity fabric. In soem edge causes the higher bandwidth might help too but it's mostly the inter-ccx speed.

* Lacks AVX-512

Who the fuck cares about AVX-512? If you do, there is no discussion needed and you should probably buy Xeon Phi anyway in that case. AVX is very old already and barley used in consumer software. We can talk about it when games use AVX521 which will probbaly take another 2 decades.

 

[ManyThreads]

Coffee lake delayed until February of next year?

https://www.golem.de/news/cpu-roadm...-coffee-lake-im-februar-2018-1706-128176.html

 

[Sweepr]

Coffee lake delayed until February of next year?

https://www.golem.de/news/cpu-roadm...-coffee-lake-im-februar-2018-1706-128176.html

Fake, for the reasons already explained here.

 

[Karnak]

You can tell Skylake X might be a huge success when the entire Amd army is on this thread desperately trying to downplay it

There is nothing to downplay.

- The 7800X is DOA with CFL-S launching in August.
- The 7820X is still too expensive. Only 28 lanes. HEDT, lol.
- You have to pay at least $999 to get 44 lanes.
- Not soldered.

And watching Linus' video that's not even all.
If the rumors about Threadripper pricing are true... oh boy Intel. $849 16C/32T vs. $999 10C/20T.
Not to mention that all Threadipper SKUs will offer 64 lanes.

 

[Edrick]

If SKL-X vs. SKL-S shows a considerable memory bottleneck on the latter, what do you make of CFL 6c/12t running dual channel only?

I personally think that CFL 6/12 is going to be bottlenecked using dual channel. We will have to wait and see, but that it my prediction.

 

[Zucker2k]

There is nothing to downplay.

- The 7800X is DOA with CFL-S launching in August.
- The 7820X is still too expensive. Only 28 lanes. HEDT, lol.
- You have to pay at least $999 to get 44 lanes.
- Not soldered.

And watching Linus' video that's not even all.
If the rumors about Threadripper pricing are true... oh boy Intel. $849 16C/32T vs. $999 10C/20T.
Not to mention that all Threadipper SKUs will offer 64 lanes.

For the sake of competition and choice, let's just hope TR arrives squeaky clean as possible. Nothing of those inter CCX and memory subsystem bugs that plagued the Ryzen launch.

 

[Zucker2k]

I personally think that CFL 6/12 is going to be bottlenecked using dual channel. We will have to wait and see, but that it my prediction.

I think Intel's reworking of the cache system would alleviate any foreseen issues in this area.

 

[Zucker2k]

Deleted.

 

[Edrick]

Who the fuck cares about AVX-512? If you do, there is no discussion needed and you should probably buy Xeon Phi anyway in that case. AVX is very old already and barley used in consumer software. We can talk about it when games use AVX521 which will probbaly take another 2 decades.

There are those of us who do. The entire CPU universe does not revolve around gaming.

 

[Edrick]

I think Intel's reworking of the cache system would alleviate any foreseen issues in this area.

Are you certain that the new Skylake-X cache system will make its way down to CFL?

 

[Zucker2k]

Are you certain that the new Skylake-X cache system will make its way down to CFL?

Most certainly. That 30% faster than Kabylake Intel is quoting won't happen without tweaking/reworking the cache and memory subsystem.

 

[tamz_msc]

There are those of us who do. The entire CPU universe does not revolve around gaming.

I want to hear from these people who do, because I want to ask them why they haven't moved on to Xeon Phi or GPGPU, like the rest of the HPC world have.