Discussion Intel current and future Lakes & Rapids thread

[uzzi38]

I don't think they would bother if it was that bad. Remember I am talking about 200 W.

And I'm talking about that power draw for OG Skylake clocks. Best case scenario.

 

[jpiniero]

And I'm talking about that power draw for OG Skylake clocks. Best case scenario.

If it was that bad, why bother? Nothing would surprise me with Intel these days though.

 

[uzzi38]

If it was that bad, why bother? Nothing would surprise me with Intel these days though.

Well yeah, that's why I think it's a dumb idea that has no basis in reality.

Rocket Lake may not be Skylake, but I can certainly tell you it's neither Sunny nor Willow Cove. Both would be extremely stupid moves.

I'm still sticking by the Skylake-X cores on a ring bus theory I had ages ago. It's the only thing that makes sense.

 

[JasonLD]

Well yeah, that's why I think it's a dumb idea that has no basis in reality.

Rocket Lake may not be Skylake, but I can certainly tell you it's neither Sunny nor Willow Cove. Both would be extremely stupid moves.

I'm still sticking by the Skylake-X cores on a ring bus theory I had ages ago. It's the only thing that makes sense.

Well, can it be heavily modified Sunny or willow Cove, maybe with smaller cache? Skylake-X cores on ring bus will provide no benefits on mainstream over existing 8 cores they have. Might as well be cometlake 8 cores

 

[uzzi38]

Well, can it be heavily modified Sunny or willow Cove, maybe with smaller cache? Skylake-X cores on ring bus will provide no benefits on mainstream over existing 8 cores they have. Might as well be cometlake 8 cores

AVX512 and significantly larger caches.

Sunny Cove doubled L2 and has 1.5x the L1d cache, trimming down caches from there wouldn't help much for power draw.

As for Willow Cove, well that's just Sunny Cove with tweaked caches as far as we currently know. So... not all that different to the above statement really.

 

[jpiniero]

The reason to use anything other than good ol regular Skylake would be to get better regular application/gaming performance, not AVX-512. It's definitely either Willow or Skylake.

 

[mikk]

Assuming the AVX512 and PCIe4 support is accurate, it pretty much rules out the possibility of Skylake. Plus there is Gen12 graphics. Yes they could theoretically use a different chip for the graphics part but this is unlikely for a GT1 with only 32 EUs. Overall this is clearly an indication of a Tigerlake version in 14nm.

 

[IntelUser2000]

The reason to use anything other than good ol regular Skylake would be to get better regular application/gaming performance, not AVX-512. It's definitely either Willow or Skylake.

If it's AVX-512 and based on Skylake, then its Skylake-X.

Otherwise, its Sunny Cove or Willow Cove.

The leaker(and he's reliable) said its Tigerlake, without the integrated Thunderbolt controller.

Sure, being on 14nm will change things a lot, but it may be low level circuit changes that we won't know.

Yes they could theoretically use a different chip for the graphics part but this is unlikely for a GT1 with only 32 EUs. Overall this is clearly an indication of a Tigerlake version in 14nm.

On 14nm, it might end up more than 200mm2 even with GT1. The graphics part gets a real good shrink using the 10nm, closer to their 2.7x figure. If they want to alleviate the pressure on 14nm fabs, that wouldn't be the way to go.

I could see a 14nm + 14nm part, aligning with earlier roadmaps. 14nm + 10nm seems too small unless they are planning on a refresh with the GPU getting a full 96EU version.

 

[jpiniero]

On the one hand, the iPhone 5G will be out by the time Rocket Lake-S enters production. On the other, Cooper Lake. So tough to say if the shortage will still be an issue at that point.But it

Now I am in total agreement that if this thing is using Willow Cove that it will be drawing a ton of juice. But it should (better be?) able to still do 5 Ghz on all cores.

 

[IntelUser2000]

Now I am in total agreement that if this thing is using Willow Cove that it will be drawing a ton of juice. But it should (better be?) able to still do 5 Ghz on all cores.

I have doubts it'll reach 5GHz as its a new core while all 5GHz Intel chips have been refined little by little for many years now.

Still, with 25% improvement per clock, even at 4.3GHz it'll still be an advancement over any Skylake based cores. Maybe it'll have some overclocking headroom too, unlike with current chips that sell factory overclocked chips as stock.

4GHz is also the point where power problems start running totally out of control so power use may be lot more manageable at 4.3GHz.

 

[DrMrLordX]

TGL-S is back on the table now

According to whom? Am I missing something?

 

[uzzi38]

According to whom? Am I missing something?

I mean it's a possibility. Not that it's definitely happening.

 

[uzzi38]

Assuming the AVX512 and PCIe4 support is accurate, it pretty much rules out the possibility of Skylake. Plus there is Gen12 graphics. Yes they could theoretically use a different chip for the graphics part but this is unlikely for a GT1 with only 32 EUs. Overall this is clearly an indication of a Tigerlake version in 14nm.

Yeah, if only it wasn't also on a roadmap that has been proven to be pretty accurate, I'd agree, the 10nm chiplet is a dumb sounding idea.

But then, backporting Willow Cove is an even dumber idea.

Anyway, if Intel want to deal with a furnace that barely outperforms their other furnaces at the cost of millions wasted in R&D whilst making themselves look like utter laughing stocks in the industry, then yeah, backporting Willow Cove sounds like a great idea. Good luck with that.

Also, to whoever said 5GHz might be possible with Rocket Lake, I'd like to say, how much do you enjoy running your CPU under LN2 all the time?

If you think I'm being harsh on Intel, well that's a bloody shame. But I like to think they have even a shred of competency left.

 

[DrMrLordX]

I mean it's a possibility. Not that it's definitely happening.

I don't see it. Intel has . . . shall we say, highly morphological roadmaps as of late, but one of the things that's consistently on all their roadmaps is: no mention of TigerLake parts with more than 4c, and definitely not anything for their desktop sockets.

 

[uzzi38]

I don't see it. Intel has . . . shall we say, highly morphological roadmaps as of late, but one of the things that's consistently on all their roadmaps is: no mention of TigerLake parts with more than 4c, and definitely not anything for their desktop sockets.

Well there's a reason I said non-zero. I don't think it's likely, just a possibility. Would help if we knew properly which roadmaps were torched, but alas.

Well, even if TGL-S does launch, it'll make Broadwell look well adopted and in large volume, so not like it's actually a relevant product anyway.

 

[jpiniero]

Also, to whoever said 5GHz might be possible with Rocket Lake, I'd like to say, how much do you enjoy running your CPU under LN2 all the time?

Intel is definitely going to be encouraging Water even with Comet. Remember that the base TDP is now 125 W.

 

[uzzi38]

Intel is definitely going to be encouraging Water even with Comet. Remember that the base TDP is now 125 W.

That's still Skylake.

WC on 14nm wouldn't even get close to 5GHz even on water.

 

[mikk]

If it's AVX-512 and based on Skylake, then its Skylake-X.

Yes but SKL-X would be a downgrade for the mainstream, gaming IPC is much lower and only 8C, any additional work on this is a waste.

On 14nm, it might end up more than 200mm2 even with GT1. The graphics part gets a real good shrink using the 10nm, closer to their 2.7x figure. If they want to alleviate the pressure on 14nm fabs, that wouldn't be the way to go.

I could see a 14nm + 14nm part, aligning with earlier roadmaps. 14nm + 10nm seems too small unless they are planning on a refresh with the GPU getting a full 96EU version.

In the tweakers roadmap there was only one RKL-U with 14+10nm, the other RKL-U and RKL-S was only 14nm. About the 14nm fabs pressure, because RKL is a 2021 product I'm not even sure if this a good call hoping for 14+10nm. Tigerlake and its successor in 2021 should get a much bigger chunk of the mobile ULV market which should help their 14nm fabs. And Intel also needs 10nm capacity for other products like bigger discrete graphics or server etc, so they need much more 10nm capacity.

 

[jpiniero]

In the tweakers roadmap there was only one RKL-U with 14+10nm, the other RKL-U and RKL-S was only 14nm.

Probably something like the higher end Rocket U gets the 10 nm GPU chiplet, while the rest gets the 14 nm one. S would only get the 14 nm.

 

[NTMBK]

Yes but SKL-X would be a downgrade for the mainstream, gaming IPC is much lower and only 8C, any additional work on this is a waste.

How much of that is down to the core, and how much is down to the higher latency from mesh fabric? Put the same core with AVX-512 and extra cache on a ringbus, and I expect to see improved gaming performance.

Its a good way to get AVX-512 support across their whole lineup (finally), and would help encourage developers.

 

[uzzi38]

Yes but SKL-X would be a downgrade for the mainstream, gaming IPC is much lower and only 8C, any additional work on this is a waste.

Not on a ring-bus it wouldn't be.

The only reason SKL-X etc are downgrades for gaming is memory latency caused by the mesh. What do you think will happen when that's gone?

 

[mikk]

How much of that is down to the core, and how much is down to the higher latency from mesh fabric? Put the same core with AVX-512 and extra cache on a ringbus, and I expect to see improved gaming performance.

Its a good way to get AVX-512 support across their whole lineup (finally), and would help encourage developers.

Intel told the highend Skylake is optimized for datacenter workloads. I don't think it's only a mesh issue, SKL-X only has 1 MB L3 whereas Skylake-S has 2 MB and with Willow Cove Intel goes up to 3MB per core. This is not a simple stepping change converting a SKL-X core onto a ringbus system, I would say this is a huge work if they really would do this, definitely not a trivial task. So much effort for a dead architecture which in the end won't be faster than CML, with 8 cores most likely slower. AVX 512 won't help them, they might gain a few percent in x265 benchmarks and that's it. To be honest this a a stupid idea to even think about.

 

[NTMBK]

Intel told the highend Skylake is optimized for datacenter workloads. I don't think it's only a mesh issue, SKL-X only has 1 MB L3 whereas Skylake-S has 2 MB and with Willow Cove Intel goes up to 3MB per core. This is not a simple stepping change converting a SKL-X core onto a ringbus system, I would say this is a huge work if they really would do this, definitely not a trivial task. So much effort for a dead architecture which in the end won't be faster than CML, with 8 cores most likely slower. AVX 512 won't help them, they might gain a few percent in x265 benchmarks and that's it. To be honest this a a stupid idea to even think about.

More stupid than spending millions of dollars porting Tigerlake back to 14nm, for a small market of enthusiasts?

 

[uzzi38]

More stupid than spending millions of dollars porting Tigerlake back to 14nm, for a small market of enthusiasts?

Not to mention the fact that it'd barely be able to outperform the 'SKL-X on ringbus' idea too. Maybe an extra 5%, if you're incredibly, incredibly lucky.

 

[Richie Rich]

Its a good way to get AVX-512 support across their whole lineup (finally), and would help encourage developers.

There is something like 20 different instruction subsets under AVX512. Every Intel CPU lineup supports different subset group. Even IceLake doesn't support all of them (14 out of 20). https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512
AVX512 is messy and that's why there's so slow adoption from SW developers and AMD (and partly from Intel too). Developers will be encouraged after CPUs will support all subsets IMHO.