Is mainstream desktop CPU development "completed"?

[witeken]

It shouldnt be any news that desktop is the last to a node. After server, mobile, tablet etc.

Dont expect to see Cannonlake for desktop either, but rather a direct go to Icelake.

Yeah, I wonder if that's a decisions made by BK.

 

[shady28]

If you don't like to discuss IPC that's fine. Besides looking at frequency, number of cores, and special instruction sets, I also like to investigate IPC. Especially with legacy code using a benchmark which has actually been applied to processors going back to 1989. It's just another metric to analyze Intel's progress. This one isolates one particular aspect of performance. And using this metric IPC advances over the last few generations has been less than 10%.

When you move outside of this outdated bench which fits in the cache you do get a better picture of overall core and memory subsystem performance. The link you provided doesn't equalize clocks, which as I stated is fine if you don't care for that data.

A while back I did a pretty extensive analysis of Anand's data when he tested Haswell against many older cores.

http://forums.anandtech.com/showthread.php?t=2324672&highlight=haswell

The summary is here:
Increase from Sandy Bridge to Haswell – 16.4%
Increase from Nehalem to Sandy Bridge – 14.6%
Increase from Conroe to Nehalem – 22.3%
Increase from Pentium 4 to Conroe – 50.7%

As you can see even going tock-to-tock only Conroe to Nehalem got past 20%. Definitely higher than 10% but for that number I was specifically referring to CPUmark99.

Keep in mind these results are with clocks equalized.

We shall see if Haswell to Skylake can crack 10% IPC improvement with CPUmark99. I'll be shocked if it does. And I'll be just as shocked if it can crack 20% IPC increase on actual apps not using instructions only in Skylake (cherry picked apps).

Right, I'm not really saying IPC is invalid, it's part of a 'deep dive' to find out specific performance characteristics. In the same way a consumer probably wouldn't really care about Performance/Watt, but they'll care about side effects like noise, heat, power consumption, and battery life.

The numbers you quote for IPC for example, show a 33% increase in IPC from Nehalem to Haswell. The indexed benchmarks I pointed out showed a 62% increase in raw performance and a 33% decrease in power consumption.

Now if we dive deeper into those benchmarks that show a 62% increase, we find that it is largely a result of more "heterogeneous compute" characteristics showing up in SB and Haswell. This shows up bigtime in Flash and Adobe media creation tools, where Haswell can be 3x faster than Nehalem.

I think that's not terribly important to *most* people, but in some instances it does matter. Remedy (incident creation system for IT departments) for example can make use of flash, and I've found that old Core 2/quad and Nehalem based systems can be very painful using the web based version compared to newer IB / Haswell systems.

Point being that, while general purpose compute may have only gone up by 32% as you state, we cannot completely discount the more specialized extensions nor can we discount the effect of process tech changes affecting clock speeds.

But even if we use the 62% over 5 years, it's a snails pace. At the 62% increase, it will take 15 years to achieve the 400% increase we got in 2 1/2 years in the 80s going from 8088/8086 to 80286.

 

[TidusZ]

Processors are not 32% faster than they were in january 2011, they are basically the same speed after overclocking both, not sure where you get 32% from.

 

[shady28]

Processors are not 32% faster than they were in january 2011, they are basically the same speed after overclocking both, not sure where you get 32% from.

Read posts 196 and 198.

 

[Hulk]

Right, I'm not really saying IPC is invalid, it's part of a 'deep dive' to find out specific performance characteristics. In the same way a consumer probably wouldn't really care about Performance/Watt, but they'll care about side effects like noise, heat, power consumption, and battery life.

The numbers you quote for IPC for example, show a 33% increase in IPC from Nehalem to Haswell. The indexed benchmarks I pointed out showed a 62% increase in raw performance and a 33% decrease in power consumption.

Now if we dive deeper into those benchmarks that show a 62% increase, we find that it is largely a result of more "heterogeneous compute" characteristics showing up in SB and Haswell. This shows up bigtime in Flash and Adobe media creation tools, where Haswell can be 3x faster than Nehalem.

I think that's not terribly important to *most* people, but in some instances it does matter. Remedy (incident creation system for IT departments) for example can make use of flash, and I've found that old Core 2/quad and Nehalem based systems can be very painful using the web based version compared to newer IB / Haswell systems.

Point being that, while general purpose compute may have only gone up by 32% as you state, we cannot completely discount the more specialized extensions nor can we discount the effect of process tech changes affecting clock speeds.

But even if we use the 62% over 5 years, it's a snails pace. At the 62% increase, it will take 15 years to achieve the 400% increase we got in 2 1/2 years in the 80s going from 8088/8086 to 80286.

Agreed.

The rate of performance increases has definitely slowed over the last few generations, more so when you look at an outdated benchmark like CPUmark99 that doesn't take special instructions, multicores, and the memory subsystem into account.

 

[Lepton87]

Processors are not 32% faster than they were in january 2011, they are basically the same speed after overclocking both, not sure where you get 32% from.

It's not that bad, we are about 10% ahead with average clocking chips and then we have a core count increase for that 1000$ CPU.

For example I had an avarage clocking SB and now I have an average clocking HW.

SB could do 4.78GHz on AIR my HW can do 4.3GHz on AIR. So my ST performance increased a bit. It didn't stagnate as you are saying. Of course you can pick a cherry sample sandy that did 5.2GHz and say it was even faster. But then again you can find a HW-e that can do about 5GHz.

Regular HW CPUs clock better like 4.5-4.6.

But it's a really pathetic increase, about 10% after 5 years. PATHETIC.

 

[tential]

Hasn't 14nm been a dud for intel anyway?
Why not wait til the next node and see what intel has for that first before saying performance increases are dead? I think intel had a lot higher hopes for 14nm than what actually happened.

 

[Lepton87]

Hasn't 14nm been a dud for intel anyway?
Why not wait til the next node and see what intel has for that first before saying performance increases are dead? I think intel had a lot higher hopes for 14nm than what actually happened.

We don't know yet how well it overclocks but everything seems to indicate that not that well but we can't be sure unless we check who knows maybe it clocks past 5Ghz but it needs a lot of juice at high clocks maybe more than HW?

 

[mikk]

inf extracted from Intels newest 10.0.27 chipset driver. I mentioned some time ago Kabylake and here it is. I'm still unsure if this is the next Tock after Skylake.

Spoiler

Sky Lake
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_1911
;Sky Lake-E
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2014
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2015
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2016
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2020
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2021
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_2024
;KabyLake
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2A1
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2A3
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2A6
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2A8
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2A9
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2AA
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2AB
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2C7
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2E8
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2E9
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2EA
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2EB
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2EE
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2F7
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A2F9
;Cannon Lake-LP
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DA1
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DA3
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DA6
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DA8
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DA9
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DAA
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_9DAB
;Cannon Lake-H
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A321
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A323
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A326
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A328
;Broadwell-DE
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_0C50
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_0C51
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_0C52
;Lewisburg
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_A1A4
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_1F16
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_1F18
;Knights Landing
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_7801
;Coleto Creek
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_23B0
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_23B1
%IntelDeviceDesc%=Needs_NO_DRV, , PCI\VEN_8086&DEV_23E0