Discussion Intel current and future Lakes & Rapids thread

[wahdangun]

If we lock the 8700K what would it's TDP be? Would it be lower?
How much power does an 8700K use if you leave it stock?

Back in the day we had a thing with CPUs called a "stepping".

Often, a later stepping of a chip performed better.

Today, we don't seem to have different steppings of the same chip.

skylake, kabylake and coffelake is essentially a steeping.

 

[dooon]

added 2 new 8C SKUs: CFL-S 8C 95W & 80W

 

[LTC8K6]

added 2 new 8C SKUs: CFL-S 8C 95W & 80W

The 9700K and the equivalent Xeon chip?

 

[Dave2150]

added 2 new 8C SKUs: CFL-S 8C 95W & 80W

9700k, 8 core Skylake pretty much officially confirmed at this point.

Pretty sad that I've had my 6700k @ 4.8Ghz since August 2015, and the 9700k will be the exact same architecture and IPC

Sure, I'd get more cores and cache, though those will have a miniscule difference in the vast majority of games I play (gaming only system).

Is it impossible for Intel to release the new Icelake architecture on 14nm++++? Or is it only compatible with 10nm?

I hope the 10700k, or whatever it's named, will finally be a new architecture with at least +10% IPC over Skylake/Kabylake/Cannonlake/Coffeelake, that's looking to be my only upgrade option at this point.

 

[ondma]

You somewhat answered your own post. Is the 8700K truly living up to the 14nm++ potential? I don't think so. If 14 nm++ was really supposed to be 52% more power efficient than 14nm, then why does adding 2 cores require a base clock speed drop? And why can the 2176G do it with so much less TDP? I suspect that Intel was being highly conservative with the base speed to ensure high yields when they were having trouble producing enough 8700K chips to meet expected demand for this rushed chip.

They seem to have come close to this with the six core mobile chips. Granted base clock is low, but turbo is very high. The tests i have seen show around 50% gain in multi-threaded benchmarks over the previous gen quads at the same tdp.

 

[jpiniero]

Is it impossible for Intel to release the new Icelake architecture on 14nm++++? Or is it only compatible with 10nm?

Impossible, no, but the decision to backport Icelake to 14 nm would have to had been made some time ago and that didn't happen.

 

[ondma]

skylake, kabylake and coffelake is essentially a steeping.

This is simply not true.

 

[LTC8K6]

9700k, 8 core Skylake pretty much officially confirmed at this point.

Pretty sad that I've had my 6700k @ 4.8Ghz since August 2015, and the 9700k will be the exact same architecture and IPC

Sure, I'd get more cores and cache, though those will have a miniscule difference in the vast majority of games I play (gaming only system).

Is it impossible for Intel to release the new Icelake architecture on 14nm++++? Or is it only compatible with 10nm?

I hope the 10700k, or whatever it's named, will finally be a new architecture with at least +10% IPC over Skylake/Kabylake/Cannonlake/Coffeelake, that's looking to be my only upgrade option at this point.

6700K/7700K/8700K/9700K all have the same amount of cache per core, so you don't get more cache with the 8700K or 9700K, really.
You do get a heck of a lot faster multi thread performance over the 4 core chips, though.

 

[Dave2150]

6700K/7700K/8700K/9700K all have the same amount of cache per core, so you don't get more cache with the 8700K or 9700K, really.
You do get a heck of a lot faster multi thread performance over the 4 core chips, though.

I believe the cache on sky/kaby/coffee/cannon lake to be a shared L3 cache of 2MB per core - meaning the total L3 cache could be utilized even if all processing cores weren't being fully utilized?

 

[piesquared]

added 2 new 8C SKUs: CFL-S 8C 95W & 80W

Looks about right. What intel needs more than anything is an answer to AMD's low wattage 8 core chips, but probably can't get a 8 core 65W chip to match the 2700/1700. Ryzen on GloFo's 14nm is a more efficient chip at lower TDPs, and intel likely can't get the 8 core clocks high enough to compete with the 2700X. Looks like they are stuck in the middle.

 

[Dayman1225]

added 2 new 8C SKUs: CFL-S 8C 95W & 80W

Where is this from?

 

[mikk]

I believe the cache on sky/kaby/coffee/cannon lake to be a shared L3 cache of 2MB per core - meaning the total L3 cache could be utilized even if all processing cores weren't being fully utilized?

Yes it is shared, the app could use 16MB even if it loads only 1 core.

Looks about right. What intel needs more than anything is an answer to AMD's low wattage 8 core chips, but probably can't get a 8 core 65W chip to match the 2700/1700. Ryzen on GloFo's 14nm is a more efficient chip at lower TDPs, and intel likely can't get the 8 core clocks high enough to compete with the 2700X. Looks like they are stuck in the middle.

80W is possibly a no iGPU Mehlow version.

 

[LTC8K6]

The E-2176G is identical to the 8700K, save for being locked, includes the IGP, and is rated at 80W TDP.

So I would think that the 80W version of the "9700K" probably also includes the IGP.

 

[LTC8K6]

I believe the cache on sky/kaby/coffee/cannon lake to be a shared L3 cache of 2MB per core - meaning the total L3 cache could be utilized even if all processing cores weren't being fully utilized?

That's true, but under full loads, the cache is basically the same amount.
The effect of 16mb of cache over 12mb on say, a 4 core load, is probably non-existent.

 

[JoeRambo]

That's true, but under full loads, the cache is basically the same amount.
The effect of 16mb of cache over 12mb on say, a 4 core load, is probably non-existent.

Not really, unlike on AMD that is limited to 8MB for 4 cores, each core can access full 16 MB of L3. I think physical address hash decides which slice is getting what data. So in the end it all depends on access patterns.
Even in so called "toxic" case, where 1 or more threads are "destroying" L3 data by replacing lines and in turn invalidating lines in L2 for other cores ( L3 must include all L2 from all cores ), you still have more cache to play with before running into trouble.

The only thing to watch out for is increased L3 latency @ same uncore clock. Coffee lake degraded by maybe 1ns, it is not unreasonable to expect 1ns slower for 16MB of L3. Still I'd take 16MB over 12MB any day.

 

[Dayman1225]

CFL 8c w/3.1GHz base on KBL Platform



Details for Result ID Genuine Intel(R) CPU 0000 @ 3.10GHz (8C 16T 3.1GHz, 8x 256kB L2, 16MB L3)

 

[formulav8]

CFL 8c w/3.1GHz base on KBL Platform

On the Kaby Lake setup?

 

[LTC8K6]

Not really, unlike on AMD that is limited to 8MB for 4 cores, each core can access full 16 MB of L3. I think physical address hash decides which slice is getting what data. So in the end it all depends on access patterns.
Even in so called "toxic" case, where 1 or more threads are "destroying" L3 data by replacing lines and in turn invalidating lines in L2 for other cores ( L3 must include all L2 from all cores ), you still have more cache to play with before running into trouble.

The only thing to watch out for is increased L3 latency @ same uncore clock. Coffee lake degraded by maybe 1ns, it is not unreasonable to expect 1ns slower for 16MB of L3. Still I'd take 16MB over 12MB any day.

I think if we take an 8700K and a "9700K" and disable 2 cores of the 8 core chip, and clock them the same, we will be hard pressed to find any performance differences.

Hopefully, a poster will be able to try that soon.

 

[jpiniero]

3.1 Ghz base for the 95W K sounds about right.

 

[LTC8K6]

3.1 Ghz base for the 95W K sounds about right.

I bet $1 that it's higher than that.

(Based on the 8700's 65W tdp at 3.2)

 

[piesquared]

80W is possibly a no iGPU Mehlow version.

Well, I've never even entertained the idea that there would be an iGPU on this chip. I don't imagine anyone is expecting that. I'm not sure why intel would consider side saddling their anemic gpu on this, it seems utterly pointless.
The only way putting a GPU next to this thing makes any sense is if they use AMD's GPU again, at least it would be useful.

But, i don't think there is any chance of this containing a GPU on chip or on package. I'm not entirely convinced this mythical CPU even exists, with or without GPU.

-They would essentially be building a single CPU with really only one usable SKU. How would they position the harvested chips with failed cores and how can they can compete on price with AMD's low cost and high yielding modular 8 core chips.

-How long ago would have they had to add this to the roadmap in order to design and release it this year(?). Probably at least 3 years, at the time when they would have been designing 10nm chips. I suppose there is a glimmer of a possibility that they ported everything back to 14nm when they realized 10nm is a disaster. What would that do to their business model though? Moving it back to 14nm would balloon the die size, TDP and/or wreak havok on clocks, and likely hurt yields, compared to their 10nm plans.

-Another glimmer of a possibility is if this has an intel 'GPU' on board and was ported back to 14nm from the originally intended 10nm when they realized 10nm was a disaster, and first caught wind of what Ryzen was going to be. But that only makes sense if they replace the entire CL stack with harvested whateverLake dies. This would surely drop their margins dramatically due to the Ryzen factor.

 

[jpiniero]

-They would essentially be building a single CPU with really only one usable SKU. How would they position the harvested chips with failed cores and how can they can compete on price with AMD's low cost and high yielding modular 8 core chips.

I imagine Intel's 14 nm is pretty close to 100% yield at this point. And the die is still decently smaller than the Ryzen die even with the IGP. Margins are going to be hit yes.

 

[ehume]

Well, I've never even entertained the idea that there would be an iGPU on this chip. I don't imagine anyone is expecting that. I'm not sure why intel would consider side saddling their anemic gpu on this, it seems utterly pointless. . . .

For those of us who do not game the "anemic" Intel iGPU is plenty good enough for looking at pictures and videos.

 

[jpiniero]

By the way I do expect Intel to split the CPU and GPU into tiles during the EMIB era but obv we are not there yet.

 

[mikk]

3.1 Ghz base for the 95W K sounds about right.

Probably higher because all these samples are based on early 0000 samples.