Intel Skylake / Kaby Lake

[Dave2150]

As the others pointed out, it doesn't matter how many PCIe lanes comes of the PCH. You'll still be limited by the PCIe 3.0 x4 "DMI 3.0" interface between CPU and PCH. Which can already be saturated by a decent PCIe SSD.

When I heard that, all my enthusiasm for regular Skylake went right out the window. Honestly how much cost and effort would it take to add a couple, or four, general purpose PCIe lanes from the CPU? Especially since they're changing socket anyway. AMD has been doing this since Llano...

DMI 3.0 is 4 GB/s.

As far as I'm aware, the only currently released consumer PCI-E NVMe SSD is the Intel 750 series, which caps out at 2GB/s from what I saw in reviews.

How is DMI 3.0 on Skylake not adequate?

I plan to upgrade to Skylake and will be installing a PCI-E nvme ssd at the same time, though if Skylake z109 can't fully utilize it, I'll have to jump on x99 instead I guess.

 

[ShintaiDK]

DMI 3.0 is 4 GB/s.

As far as I'm aware, the only currently released consumer PCI-E NVMe SSD is the Intel 750 series, which caps out at 2GB/s from what I saw in reviews.

How is DMI 3.0 on Skylake not adequate?

I plan to upgrade to Skylake and will be installing a PCI-E nvme ssd at the same time, though if Skylake z109 can't fully utilize it, I'll have to jump on x99 instead I guess.

DMI 3.0 is 2GB/sec each way. Just as a PCIe 3.0 x4.

 

[Sweepr]

Next items of interest

1. Throttling profile

crossing fingers..

Broadwell and Cherry Trail can maintain max turbo clocks for longer periods than its predecessors, I expect the same from Skylake.

 

[Insert_Nickname]

DMI 3.0 is 4 GB/s.

DMI 3.0 is 2GB/sec each way. Just as a PCIe 3.0 x4.

No, it is 4GB/s both ways simultaneously. However, peak bandwidth can rarely be achieved in practice.

http://www.anandtech.com/show/2973/6gbps-sata-performance-amd-890gx-vs-intel-x58-p55/2

As far as I'm aware, the only currently released consumer PCI-E NVMe SSD is the Intel 750 series, which caps out at 2GB/s from what I saw in reviews.

How is DMI 3.0 on Skylake not adequate?

I plan to upgrade to Skylake and will be installing a PCI-E nvme ssd at the same time, though if Skylake z109 can't fully utilize it, I'll have to jump on x99 instead I guess.

Think of it this way; you can have a (hypothetical currently) PCIe 3.0 x4 SSD connected to the PCH, no problem. But there will be no bandwidth left for anything else if its going full-throttle. The DMI link between CPU/PCH also introduces a small amount of latency. This might not matter if your workload isn't particularly disk I/O dependent. A PCIe SSD is still blazing fast. I just hate bottlenecks, and who's to say we won't see SSDs with some form of DRAM cache that can saturate the link in the future? I mean, look at just how fast SATA3 was saturated sequentially.

If you're thinking of moving up to the HEDT, like me, at least wait until Skylake-E. You get PCIe 4.0 as a (rumoured) bonus.

 

[ShintaiDK]

No, it is 4GB/s both ways simultaneously. However, peak bandwidth can rarely be achieved in practice.

Sorry, for some reason I keep thinking in PCIe 2.0 speeds for 3.0 >.<

 

[Dave2150]

No, it is 4GB/s both ways simultaneously. However, peak bandwidth can rarely be achieved in practice.

http://www.anandtech.com/show/2973/6gbps-sata-performance-amd-890gx-vs-intel-x58-p55/2



Think of it this way; you can have a (hypothetical currently) PCIe 3.0 x4 SSD connected to the PCH, no problem. But there will be no bandwidth left for anything else if its going full-throttle. The DMI link between CPU/PCH also introduces a small amount of latency. This might not matter if your workload isn't particularly disk I/O dependent. A PCIe SSD is still blazing fast. I just hate bottlenecks, and who's to say we won't see SSDs with some form of DRAM cache that can saturate the link in the future? I mean, look at just how fast SATA3 was saturated sequentially.

If you're thinking of moving up to the HEDT, like me, at least wait until Skylake-E. You get PCIe 4.0 as a (rumoured) bonus.

Thanks for the reply.

I'd like to wait for Skylake-E, though I've read that we'll be waiting until late 2016 for this to be released. I don't think I want to wait that long.

Either way I'll wait for official Skylake reviews, hopefully there will be in-depth tests of nvme drives and how this extra latency of the dmi 3.0 interface will affect performance.

 

[cytg111]

Broadwell and Cherry Trail can maintain max turbo clocks for longer periods than its predecessors, I expect the same from Skylake.

Sounds promising

 

[Insert_Nickname]

I'd like to wait for Skylake-E, though I've read that we'll be waiting until late 2016 for this to be released. I don't think I want to wait that long.

You can of course just use a few of the lines from the CPU. That means cutting bandwidth to the GPU, and requires a Z-series chipset. But it doesn't matter much in practice, unless you're running some serious GPU compute. The performance hit is on the order of 1-2% for gaming.

Or wait and see if something like the Asrock Z97 Extreme9 is released for Skylake. It uses a PLX chip to provide maximum bandwidth to both a PCIe SSD and graphics card. The downside is that its still limited by the 16 lanes coming of the CPU, but at least both devices have peak bandwidth when they need it. Its properly unlikely both will need that at the same time.

In short there are plenty of possibilities. One just has to find them.

Either way I'll wait for official Skylake reviews, hopefully there will be in-depth tests of nvme drives and how this extra latency of the dmi 3.0 interface will affect performance.

Some good reads:

http://www.tomshardware.com/reviews/samsung-xp941-z97-pci-express,3826.html
http://www.tomshardware.com/reviews/ssd-dc-s3500-raid-performance,3613.html

 

[Sweepr]

Year/rear improvements were pretty good in the last 2 years. Devil's Canyon was 12-13% faster than Core i7 4770K and hopefully Skylake brings another ~15% or so this year. Haswell-E made 6C/12T chips more accessible too, sooner or later Intell will push hexa-cores to the mainstream platform (after Cannonlake?).

 

[Enigmoid]

No, it is 4GB/s both ways simultaneously. However, peak bandwidth can rarely be achieved in practice.

http://www.anandtech.com/show/2973/6gbps-sata-performance-amd-890gx-vs-intel-x58-p55/2



Think of it this way; you can have a (hypothetical currently) PCIe 3.0 x4 SSD connected to the PCH, no problem. But there will be no bandwidth left for anything else if its going full-throttle. The DMI link between CPU/PCH also introduces a small amount of latency. This might not matter if your workload isn't particularly disk I/O dependent. A PCIe SSD is still blazing fast. I just hate bottlenecks, and who's to say we won't see SSDs with some form of DRAM cache that can saturate the link in the future? I mean, look at just how fast SATA3 was saturated sequentially.

If you're thinking of moving up to the HEDT, like me, at least wait until Skylake-E. You get PCIe 4.0 as a (rumoured) bonus.

Very true but seriously? How often do you expect it to be going full throttle?

If you buy a 1 TB SSD and it manages 4 GB/sec (which nothing on the market currently does) you still will, doing continuous writes, fill up the drive completely in <250 seconds.

Not to mention that 4 GB sec is going to require a ton of CPU power to really do anything useful with the data. Unless its brute force transfers I really don't see the problem in real world usage.

 

[Insert_Nickname]

Very true but seriously? How often do you expect it to be going full throttle?

If you buy a 1 TB SSD and it manages 4 GB/sec (which nothing on the market currently does) you still will, doing continuous writes, fill up the drive completely in <250 seconds.

True, but again I hate being bottlenecked. Real or theoretical. I don't see anyone complaining about faster memory for example. Especially since this would be so easy for Intel to solve. Just add 4 general purpose PCIe lanes from the CPU (16>20 or 22/24 for some fast USB3.1/Thunderbolt combined with SSD). Would likely save a bit of power too, since fewer lanes are required from the PCH, and the CPU is manufactured on a smaller process node. As I wrote in a previous post AMD has been doing this for years.

Its actually bad enough I might have to take 8 lanes from my 3770/GPU to get full advantage of a PCIe SSD. But the trade-of should be well worth it...

 

[Fjodor2001]

Not to mention that 4 GB sec is going to require a ton of CPU power to really do anything useful with the data. Unless its brute force transfers I really don't see the problem in real world usage.

Burst loading 16 GB VMware machine images (i.e. "virtual RAM" being persisted on disk) in 4 seconds instead of 32 seconds as on SATA3?

 

[therealnickdanger]

Its actually bad enough I might have to take 8 lanes from my 3770/GPU to get full advantage of a PCIe SSD. But the trade-of should be well worth it...

Don't forget Gen 9 graphics! You won't even miss your GPU!

 

[IntelUser2000]

Yea, about those gen 9 graphics. Hope they bring some real improvement. Gen 8 was supposed to, but except for more shaders, it seems the architechture did almost nothing.

I cannot believe that Gen 9 is going to be substantially better than what Gen 8 brought.

Gen 8 was 20% over Gen 7.5. Gen 9 is probably going to be the same over Gen 8, but with same EUs. Considering how even Intel graphics seems so bandwidth bound nowadays, and with Broadwell GT3e coming at the same time as Skylake GT2, its logical to conclude Intel will incline people to Broadwell GT3e for graphics, and Skylake GT2 for CPU.

There's only about 60% difference between GT2 and GT3e at the same generation/TDP, so Skylake's gain must be much less than 60%.

 

[Insert_Nickname]

Don't forget Gen 9 graphics! You won't even miss your GPU!

I do have this strange feeling I'll miss my 970 quite a lot for 1440p gaming...

The Gen9 graphics do look very interesting from an HTPC perspective, mind you. We shall see soon enough.

 

[IntelUser2000]

Nice, 20% over haswell. not bad not bad at all. At 20% I'd be hard pressed to choose between a hexcore and a quadcore+20%st .. I would problary go with the quad!

It won't be 20%/clock. It's based on the false assumption that Core i7 4790K can reach 4.4GHz with all 4 cores, while 6700K can only do 4.2GHz. In reality 4790K can only do 4.2GHz with 4 cores so its 15%.

 

[Enigmoid]

Burst loading 16 GB VMware machine images (i.e. "virtual RAM" being persisted on disk) in 4 seconds instead of 32 seconds as on SATA3?

Won't you hit CPU bottlenecks there?

 

[AtenRa]

It won't be 20%/clock. It's based on the false assumption that Core i7 4790K can reach 4.4GHz with all 4 cores, while 6700K can only do 4.2GHz. In reality 4790K can only do 4.2GHz with 4 cores so its 15%.

And this is from Cinebench MT, which should be the most favorable of the benchmarks. On average it should be lower.

 

[Sweepr]

And this is from Cinebench MT, which should be the most favorable of the benchmarks. On average it should be lower.

Cinebench and 3DMark CPU. You can't say it will be lower or higher than that ES yet. Either way, Skylake already looks like a winner from a performance per core POV.

Also 4-core Turbo might be a tad lower than Core i7 4790K.

 

[IntelUser2000]

And this is from Cinebench MT, which should be the most favorable of the benchmarks. On average it should be lower.

That's actually about the gain from 975EE to 2600K. If it can maintain those averages it should be pretty good.

It's true that the gain isn't spectacular. If they marketing wanted to artificially inflate the gains though, they could have decided not to release the 4790K just like they decided not to reach mainstream Westmere quads.

Also, its at very high clocks, so the gains should be better on the lower SKUs and power use. At 35W we are looking at 30-40%, and 65W, 20-30%. I am looking forward to see how the SKL-Y and SKL-U does at the end of the year.

 

[ehume]

It won't be 20%/clock. It's based on the false assumption that Core i7 4790K can reach 4.4GHz with all 4 cores, while 6700K can only do 4.2GHz. In reality 4790K can only do 4.2GHz with 4 cores so its 15%.

True, if you go by specs. My GA-Z97X mb starts the 4790K at 4.4GHz, as the "optimal" default.

 

[Sweepr]

I cannot believe that Gen 9 is going to be substantially better than what Gen 8 brought.

Gen 8 was 20% over Gen 7.5. Gen 9 is probably going to be the same over Gen 8, but with same EUs. Considering how even Intel graphics seems so bandwidth bound nowadays, and with Broadwell GT3e coming at the same time as Skylake GT2, its logical to conclude Intel will incline people to Broadwell GT3e for graphics, and Skylake GT2 for CPU.

There's only about 60% difference between GT2 and GT3e at the same generation/TDP, so Skylake's gain must be much less than 60%.

Well even if we assume 20% better performance per EU compared to Broadwell we would be looking at 40-50% faster graphics than desktop Haswell. Considering AMD will only launch Kaveri Refresh in non-BGA form for desktops this year that should close the gap considerably.
Also Intel will likely use the same IGP in cheaper models like Core i3 and DDR4 should help alleviate the BW bottleneck.

 

[inf64]

Why would you get a K variant with Haswell, but not with Skylake?

Because I don't need to OC Skylake if it is indeed ~15-20% faster than Haswell. This should mean that 3.3-3.9Ghz i5 Skylake is as fast as ~3.8-4.5Ghz Haswell i5 (OCed 4670K). Why pay more for K Skylake when it performs as well as OCed Haswell?

 

[IntelUser2000]

Well even if we assume 20% better performance per EU compared to Broadwell we would be looking at 40-50% faster graphics than desktop Haswell. Considering AMD will only launch Kaveri Refresh in non-BGA form for desktops this year that should close the gap considerably.
Also Intel will likely use the same IGP in cheaper models like Core i3 and DDR4 should help alleviate the BW bottleneck.

You know how low performing lower end iGPUs like GT2s are. 30-40% is not enough to put 15 fps to playable range. You need 2x. 3x. Something like that. It's a piss poor gain for a GPU.

I find it ridiculous how Intel goes from 22nm to 14nm the "biggest process gain evah" and you get 20% while Nvidia changes architecture and gets 2x perf/watt. So its really different way of achieving the same goal and the latter does far better.

 

[Sweepr]

You know how low performing lower end iGPUs like GT2s are. 30-40% is not enough to put 15 fps to playable range. You need 2x. 3x. Something like that. It's a piss poor gain for a GPU.

I find it ridiculous how Intel goes from 22nm to 14nm the "biggest process gain evah" and you get 20% while Nvidia changes architecture and gets 2x perf/watt. So its really different way of achieving the same goal and the latter does far better.

Well I'm not expecting playable framerates at the most demanding games, I'm talking about improving their position relative to AMD's APUs. I personally couldn't care less about desktop iGPUs but I know some people do.
20% better than Broadwell was your prediction, not mine, could be less or more than that (hopefully the latter).