Has the wikipedia article written accurately?

This section doesn't seem to be cited.

https://en.wikipedia.org/wiki/Cannonlake

Due to low 10nm yields, Cannonlake will be limited to 15 Watt U and 5.2 Watt Y system-on-chip parts with GT2. Higher power mobile and desktop platforms will receive an update in the form of a 2nd 14nm process refinement, Coffee Lake, that is said to share Cannonlake's architectural refinements.

Click to expand...

Coffee Lake will exist in lieu of Desktop Cannonlake?

muon said:

Has the wikipedia article written accurately?

This section doesn't seem to be cited.

https://en.wikipedia.org/wiki/Cannonlake



Coffee Lake will exist in lieu of Desktop Cannonlake?

Click to expand...

Essentially yes.

Intel has said Coffee Lake will come after Kaby Lake, and be essentially a 14nm version of Cannonlake. Also supposedly it'll go up to 6 cores.

Whether or not desktop gets 'true' 10nm Cannon Lake is less clear. But 14nm 6 core Coffee Lake (which is Cannon Lake arch) is meant to be confirmed.

What a disaster. The whole point of a die shrink is that you can increase the clockspeeds on a mature architecture.

Well done to those that bought Skylake.

Broadwell, Skylake, Kaby Lake, Coffee Lake. All desktop 14 nm parts.

Coffee Lake is the most pointless money grab ever.

"14nm version of Cannonlake"

So.... Skylake?

muon said:

What a disaster. The whole point of a die shrink is that you can increase the clockspeeds on a mature architecture.

Well done to those that bought Skylake.

Click to expand...

Clockspeeds (particularly max overclocked speeds) haven't really gone up now for a while.... Witness bw-e often being outpaced oc'd by hw-e chips that clock better

Orangey said:

Coffee Lake is the most pointless money grab ever.

"14nm version of Cannonlake"

So.... Skylake?

Click to expand...

Well it's mean't to have refinements, so maybe 10% more IPC than Skylake (Kaby Lake supposedly has 5% more, and we'll find out soon).

So 6-core chip with 10% more IPC than Skylake, and 15% more IPC than Broadwell-E, likely also clocking better than Broadwell-E. This sounds like a decent product, also assuming the 6-core sells for ~£300.

These IPC numbers from the past half-decade are pure fiction.

Please explain what you mean by that Orangey?

Orangey said:

These IPC numbers from the past half-decade are pure fiction.

Click to expand...

What? Are you trying to say Skylake has the same IPC as Sandy Bridge?

Skylake's IPC is roughly 60% higher than Bloomfield/Westmere, which came 7 years earlier. Obviously it depends on the workload though, this is just an appromximate average.

So that's less than 10% improvement per year. Would be interested to see how that compares to previous 7 year periods if such a table/graph exists.

DragonQ said:

Skylake's IPC is roughly 60% higher than Bloomfield/Westmere, which came 7 years earlier. Obviously it depends on the workload though, this is just an appromximate average.

So that's less than 10% improvement per year. Would be interested to see how that compares to previous 7 year periods if such a table/graph exists.

Click to expand...

It's a bit less than that probably. Clock for clock an i5-6600 is about 43% faster than an i5-760 in cinbench, x264, and 7zip.

There's only 5 years between these chips (Q3'15 and Q3'10), so a 7.4% improvement year on year (compounded).

Edit: actually you said bloomfield/westmere, so we could compare an i7-920 and an i7-6700 (ark). Surprisingly the mean efficiency difference is only +38%, over 6.75 years that's 8.8% per year. (The median is probably more reliable, +49%, because those 3d particle movement benchmarks drag the average right down.)

muon said:

What a disaster. The whole point of a die shrink is that you can increase the clockspeeds on a mature architecture.

Well done to those that bought Skylake.

Click to expand...

I thought the point was efficiency.

Anandtech also did an in-depth look into the IPC increases since Sandy Bridge.

Summary: http://images.anandtech.com/doci/9483/01 - Gains over Sandy_575px.png

From article: http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

So in the absolute worst case, it still has a 5.5% improvement. And when new instruction sets are used, and other arch refinements taken advantage of the improvement is just shy of a whopping 70%.

Adding extra RAM bandwidth from DDR4 also nets you a little more IPC.

So the overall average, according to Anandtech is 24.8% increase in IPC from Sandy to Skylake, which a maximum increase of ~70% if a task is strongly tailored to the improvements.

If Kaby Lake adds another ~5% we're then up to ~30% average IPC over Sandy Bridge. This may not be earth-shattering, but it's certainly a decent increase for people on Sandy/Ivy.

AllBodies said:

So in the absolute worst case, it still has a 5.5% improvement. And when new instruction sets are used, and other arch refinements taken advantage of the improvement is just shy of a whopping 70%.

Click to expand...

A bit misleading. It looks like only Dolphin Benchmark (a Gamecube/Wii emulator?) had a really substantial jump, and even then it was mainly between IB and Haswell, almost certainly due to AVX -> AVX 2.

joeyjojo said:

A bit misleading. It looks like only Dolphin Benchmark (a Gamecube/Wii emulator?) had a really substantial jump, and even then it was mainly between IB and Haswell, almost certainly due to AVX -> AVX 2.

Click to expand...

Perhaps misleading if I worded it differently. The statement is entirely accurate as is.

AllBodies said:

Perhaps misleading if I worded it differently. The statement is entirely accurate as is.

Click to expand...

Sorry, I forgot to say thanks for sharing, that's a great article.

joeyjojo said:

It's a bit less than that probably. Clock for clock an i5-6600 is about 43% faster than an i5-760 in cinbench, x264, and 7zip.

There's only 5 years between these chips (Q3'15 and Q3'10), so a 7.4% improvement year on year (compounded).

Edit: actually you said bloomfield/westmere, so we could compare an i7-920 and an i7-6700 (ark). Surprisingly the mean efficiency difference is only +38%, over 6.75 years that's 8.8% per year. (The median is probably more reliable, +49%, because those 3d particle movement benchmarks drag the average right down.)

Click to expand...

Yeah I was looking at i7-920 v i7-6700K with obviously the i7-920 being hugely clock adjusted to match 4 GHz. You have to be careful what benchmarks you look at because certain tasks benefit hugely from new instruction sets and others don't. An average improvement of 50-60% seems about right, but Bloomfield came out in 2008 so there's been 7 years between them.