Discussion Intel current and future Lakes & Rapids thread

[jpiniero]

No, it's just later than 8+8.

What repoman wrote first is correct, it should be around the beginning of next year for a launch. Could be before, could be after, idk.

Whenever the non-K SKUs launch, it should be then.

That was the rumor. Intel did the same thing with 11th Gen if you think about it. We'll have to see if that ends up being true. It obviously would remove a ton of 10 nm wafers that would be needed since i3 is still very popular with OEMs.

 

[scineram]

According to the hot chips presentation: 50%:

View attachment 49195

So Golden Cove is more efficient than Gracemont all the way except lowest end of the curve. All the power efficiency crap is gaslighting. It is all about area.

 

[dullard]

So Golden Cove is more efficient than Gracemont all the way except lowest end of the curve.

That is why they are running Gracemont at the lowest end of the curve. See the right area in red on their slide #32 below. They are operating Gracemont at voltages and low power states that Golden Cove cannot handle.

It is all about area.

Yes, Intel said so themselves. See the left area in red on their slide #32. Gracemont is all about area to get in as many cores as possible over time.

 

[uzzi38]

That is why they are running Gracemont at the lowest end of the curve. See the right area in red on their slide #32 below. They are operating Gracemont at voltages and low power states that Golden Cove cannot handle.

Yes, Intel said so themselves. See the left area in red on their slide #32. Gracemont is all about area to get in as many cores as possible over time.
View attachment 49258

Not on desktop they're not. Gracemonts don't really go above 4GHz.

 

[dullard]

Not on desktop they're not. Gracemonts don't really go above 4GHz.

? I said they are running at low power (which means low frequency). Then you reply with no, they are running at low frequency. Something is lost in translation there.

 

[uzzi38]

? I said they are running at low power (which means low frequency). Then you reply with no, they are running at low frequency. Something is lost in translation there.

They don't clock over 4GHz. At all. They're being run right at their limit, not low down on the V/f curve.

My point is they're not being run at a "low frequency" at all. They're being run right at their max.

 

[coercitiv]

All the power efficiency crap is gaslighting. It is all about area.

Bingo! Based on Intel's own estimate they're doubling the throughput in the same area. 4 E-cores can do the work of 2 P-cores.

 

[dullard]

They don't clock over 4GHz. At all. They're being run right at their limit, not low down of the V/f curve.

Base is rumored to be 1.8 GHz. Max turbo hasn't been announced as far as I know, but might go up to 4.0 GHz (I've seen rumors of 3.4 GHz, 3.6 GHz, and 4.0 GHz). 1.8 GHz is quite a ways down the power curve.

 

[eek2121]

They don't clock over 4GHz. At all. They're being run right at their limit, not low down on the V/f curve.

My point is they're not being run at a "low frequency" at all. They're being run right at their max.

Low frequency means "Not 5 GHz".

Base is rumored to be 1.8 GHz. Max turbo hasn't been announced as far as I know, but might go up to 4.0 GHz (I've seen rumors of 3.4 GHz, 3.6 GHz, and 4.0 GHz). 1.8 GHz is quite a ways down the power curve.

The leaks claim 3.9 ghz for Gracemont for the top i9 SKU and 3.7 GHz for the i7. The base is 1.8 GHz supposedly.

 

[uzzi38]

Low frequency means "Not 5 GHz".



The leaks claim 3.9 ghz for Gracemont for the top i9 SKU and 3.7 GHz for the i7. The base is 1.8 GHz supposedly.

No, it doesn't. @dullard claimed Intel were running the Gracemonts close to the voltage floor which Golden Cove can't drop down to. That would be FAR less than 3.9GHz. Actually probably close to or below the 1.8GHz base clock you say here.

3.9GHz is close to the absolute maximum they can clock to full stop. It is very much high frequency for the Gracemont cores, even if not high frequency in relation to Golden Cove cores.

 

[Ajay]

Bingo! Based on Intel's own estimate they're doubling the throughput in the same area. 4 E-cores can do the work of 2 P-cores.

Seriously? I missed that. I would have expected a ratio of less than 2:1. Intel really needs a new architecture. Seems like the endless extensions of 'Core' are losing steam. Or, is Intel simply shackeled by the x86-64 instruction set limitations? Seems like AMD is not, or, at least, not to the same extent.

 

[Hulk]

What do you think the top frequencies are for the two parts indicated here?
3.9GHz for Gracemont?
4.5GHz for Skylake?

 

[dullard]

What do you think the top frequencies are for the two parts indicated here?
3.9GHz for Gracemont?
4.5GHz for Skylake?

1) Your red grid is a bit off from Intel's claims. For example, at your 140 performance line (where they stopped graphing Skylake), it is listed as ~97% power for Skylake. But, Gracemont at your 140 performance line is ~43. 43/97 is 44% of the power. But Intel claims that it is even better than that (<40% of the power is needed). If you redo it with a different x-axis scaling, then you can at least match what Intel claims. Most likely the problem is that you put 0% power well to the left of the "Performance" label.

2) Skylake topped out at 4.2 GHz single core in the 6700K unless you overclocked it. I would doubt that Intel is putting overclocked values on the graph, but it is possible to make a graph with overclocked settings. You could argue that Skylake-X got to 5.0 GHz, but then it would make more sense for Intel to label it as Skylake-X. Heck, we don't even know if that graph goes up to 4.2 GHz, they could have truncated it.

 

[Abwx]

What do you think the top frequencies are for the two parts indicated here?
3.9GHz for Gracemont?
4.5GHz for Skylake?

View attachment 49264

The E core curve is irrealistic and defy the laws of physics...

In the most favourable part of a performance/power curve perf should increase as a square root of power, anything that display a steeper slope is pure marketing construction...

 

[Hulk]

Based on the Intel provided "Latency Performance" graph for Skylake vs. Gracemont I plotted the reduction in power usage for Gracemont vs. Skylake assuming in the Intel chart Gracemont tops out at 3.9GHz.
y-axis is reduction in power, ie assume Skylake is using 100%, then if this charts shows 40% at 1800MHz, this means at equal performance Gracemont is using 60% of the power of Skylake. Except for the slight deviation around 3.1GHz it's as expected.

Some other interested "finds" from this chart. Assume Gracemont tops out at 3.9GHz. If Skylake tops out at 5.3GHz then Gracemont is producing 48.5% better IPC. If Skylake tops out at 5.0GHz then Gracemont is 40.1% better. If Skylake tops out at 4.5GHz then Gracemont is 12.1% better. Clock-for-Clock I'm seeing 9.3% better for Gracemont.

Again, all with lots of grains of salt. Just messing around.

 

[dullard]

The E core curve is irrealistic and defy the laws of physics...

In the most favourable part of a performance/power curve perf should increase as a square root of power, anything that display a steeper slope is pure marketing construction...

CPU power is proportional to frequency cubed. Assuming a typical workload where frequency is proportional to performance, then CPU power is proportional to performance cubed. Or, flipped around, performance is related to the CPU power to the 1/3 power. The graphs should look something like this:

 

[dullard]

Based on the Intel provided "Latency Performance" graph for Skylake vs. Gracemont I plotted the reduction in power usage for Gracemont vs. Skylake assuming in the Intel chart Gracemont tops out at 3.9GHz.
y-axis is reduction in power, ie assume Skylake is using 100%, then if this charts shows 40% at 1800MHz, this means at equal performance Gracemont is using 60% of the power of Skylake. Except for the slight deviation around 3.1GHz it's as expected.

Some other interested "finds" from this chart. Assume Gracemont tops out at 3.9GHz. If Skylake tops out at 5.3GHz then Gracemont is producing 48.5% better IPC. If Skylake tops out at 5.0GHz then Gracemont is 40.1% better. If Skylake tops out at 4.5GHz then Gracemont is 12.1% better. Clock-for-Clock I'm seeing 9.3% better for Gracemont.

Again, all with lots of grains of salt. Just messing around.

View attachment 49265

Intel didn't claim a 40% reduction of power. They claimed even better: to need less than 40% of the power. That would be at least a 60% reduction.

 

[Abwx]

CPU power is proportional to frequency cubed. Assuming a typical workload where frequency is proportional to performance, then CPU power is proportional to performance cubed. Or, flipped around, performance is related to the CPU power to the 1/3 power. The graphs should look something like this:
View attachment 49266

I said in the most favourable case, wich is theorical but still almost realized at mid frequency, at wich point power is a polynomial of degre close to 2 of frequency.

As frequency increase the power/perf curve converge to a cubic law before getting quadratic at the higher end of the frequency range, so perf in function of power is reciprocally increasing first as a square root then a cubic root to end as a quadratic root.

 

[Hulk]

Intel didn't claim a 40% reduction of power. They claimed even better: to need less than 40% of the power. That would be at least a 60% reduction.

Not trying to back-up/deny their claim. Only looking at the data. They can get to 60% as Skylake starts to go nuclear, and beyond of course. But not knowing frequencies makes this kind of ridiculous. 60% less power than Skylake when Skylake is at 4GHz, ... Very impressive. 60% of Skylake when Skylake is at 5.3GHz,... Big deal.

But coertitiv wrote this a page or two back. He's right. The frequency where these power reductions occur are important.

 

[dullard]

I said in the most favourable case, wich is theorical but still almost realized at mid frequency, at wich point power is a polynomial of degre close to 2 of frequency.

As frequency increase the power/perf curve converge to a cubic law before getting quadratic at the higher end of the frequency range, so perf in function of power is reciprocally increasing first as a square root then a cubic root to end as a quadratic root.

I put a theoretical Performance = (power)^(1/3) graph on top of Intel's claims. The shape is pretty accurate.

 

[dullard]

But coertitiv wrote this a page or two back. He's right. The frequency where these power reductions occur are important.

That is true. I would assume that the graph is a variety of frequencies. That is, low frequencies to the left and high frequencies to the right. Thus, it isn't a graph at a single frequency.

 

[IntelUser2000]

ADL 2+8+2 LP: 18.4 mm x 11.2 mm = 206 mm².
(For reference, CML 10+2 HP is 22.4 mm x 9.2 mm = 206 mm².)

ADL 8+8+1 HP: 25.2 mm x 12.6 mm = 318 mm².
(For reference, RKL 8+1 HP is 24.0 mm x 11.5 mm = 276 mm².)

And wait for it... ADL 6+8+2 LP: 29.8 mm x 14.6 mm = 435 mm².
(For reference, SPR 15C tile is 20.8 mm x 20.5 mm = 426 mm².)

There is no way Intel can maintain anything like their traditional margins with 10nm client dies that size.

I'm highly skeptical the 6+8+2 is that large. Your calculations are basically saying the size difference between 32EU and 96EU Xe GPU is over 100mm2, and approaching 150mm2.

Actually the entire 96EU GPU which includes the FF units, are only about 40mm2. 32EU only cuts that down by half, meaning it's a 20mm2 difference.

Update: I just did measurements myself and those package renderings seem seriously off. It would indicate the P part is just larger everywhere for no apparent reason.

 

[Ajay]

Old, somewhat complex power calculation from former member idontcare. The dominate term is Vcc (4th power). Of course, increases in voltage increase frequency. So knowing frequency = f(Vcc) would be a must.

 

[IntelUser2000]

For example, if we pick up Cinebench R20 single-core then we can find its single-core performance is 294 which is slightly higher than Cascade Lake 8253(single-core 286, 3Ghz Turbo Skylake) without any AVX support. Cinebench multi also says the same thing, Tremont(2Ghz x 4, 10W) is comparable to i3 8109U(3Ghz x 2 + HT, 20~28W)

Tremont is about Ivy Bridge, not Haswell+.

It does really well in certain benchmarks like Geekbench and on there it approaches Haswell performance.

@uzzi38 6700K is 4.2GHz but they don't have to compare to the 6700K. If it's against the mobile parts it's pretty much a perf/clk comparison. They did say it's "higher IPC" than Skylake.

If you assume Golden Cove is 33% faster, than at 5GHz, that's 30% higher frequency and easily claim 50% better ST performance over Gracemont.

 

[uzzi38]

Tremont is about Ivy Bridge, not Haswell+.

It does really well in certain benchmarks like Geekbench and on there it approaches Haswell performance.

@uzzi38 6700K is 4.2GHz but they don't have to compare to the 6700K. If it's against the mobile parts it's pretty much a perf/clk comparison. They did say it's "higher IPC" than Skylake.

If you assume Golden Cove is 33% faster, than at 5GHz, that's 30% higher frequency and easily claim 50% better ST performance over Gracemont.

Oh wait, whoops, you're right, it's 4GHz base instead, not boost. That's my bad. Still, agreed with all of that, and I expect something similar things from Gracemont too. In some workloads it'll probably exceed SKL IPC, in others it may be more like HSW. Will be interesting to see how they're received overall