Discussion Intel current and future Lakes & Rapids thread

[itsmydamnation]

'3' x N = PRQ
SRF is not PRQ yet.

X x N = 144
Repeatedly, Yuuk_AnS says that is A0 stepping (i.e. X = 1 , therefore, N = 144).
3 x 144 = 432

Consequently, SRF will be up to 432 cores.

I have same magic beans that double clock rates of any existing silicon. Are you interested? I take western union money transfer only.

Let me Know asap I have other interested buyers!!!!

 

[lightisgood]

350W is not enough for 432 cores, for 144 cores that s about 2W/core.

1. Birch Stream-AP will support up to 500W.
2. Gracemont@3.9GHz roughly consumes 4W/core.
3. Hopping Gracemont to SRF is two node jump on perf/watt gain. Potentially speaking, two node jump will deliver 0.36x power consumption at same clock.

4 x 0.36 x 432 = 622 watt

Therefore, I guess that Intel will launch SRF-AP as 432cores/3~3.5GHz/500W, the phenomenon products.

 

[itsmydamnation]

1. Birch Stream-AP will support up to 500W.
2. Gracemont@3.9GHz roughly consumes 4W/core.
3. Hopping Gracemont to SRF is two node jump on perf/watt gain. Potentially speaking, two node jump will deliver 0.36x power cosumption at same clock.

4 x 0.36 x 432 = 622 watt

Therefore, I guess that Intel will launch SRF-AP as 432cores/3~3.5GHz/500W, the phenomenon products

lol ,
how many memory channels mate?
how much power for the interconnect
given the "tiny"* amount of L3 / L2 memory pressure is going to be extremely high.

*Zen4D is 2mb of L3 and mb of L2 per core

edit: cant divide 16/8 properly

 

[lightisgood]

*Zen4D is 4mb of L3 and mb of L2 per core

So what?

Now, here... I have no interest in Zen topics, because I'm sure that we can output the inference about future Lakes & Rapids of Intel from it's own leak , rumor, and so on.

 

[TESKATLIPOKA]

So what?

Now, here... I have no interest in Zen topics, because I'm sure that we can output the inference about future Lakes & Rapids of Intel from it's own leak , rumor, and so on.

What he meant was that what you calculated is just the cores, but not only cores consume power in a CPU and by having little L2/L3, the memory subsystem and interconnects will have to be loaded more often, consuming even more power.

So what you wrote here:

4 x 0.36 x 432 = 622 watt

Therefore, I guess that Intel will launch SRF-AP as 432cores/3~3.5GHz/500W, the phenomenon products.

Could be a bit too optimistic.

I don't even know how you ended up with 4W for Gracemont at 3.9GHz?
From what you linked, the first graph I see 45.7W when 8 Gracemont cores are loaded, so 5.7W/core and I don't even know If they clocked at 3.9GHz.
I am also skeptical about lowering core power consumption down by 64% at the same clock.

 

[JoeRambo]

What he meant was that what you calculated is just the cores, but not only cores consume power in a CPU and by having little L2/L3, the memory subsystem and interconnects will have to be loaded more often, consuming even more power.

( I am not making any claims about core counts, just viability part )
These CPUs compete in so called "compute dense" part of SoC spectrum. ARM currently has up to 192-core CPUs that compete in same exact space and i think they have only recently made them sensible with stuff like 2MB core private L2 etc.
I don't see why Intel can't fit say 2x144 or 3x144 cores in such SoC if they have advanced enough node. The only question would be clocks and power and SoC smarts like turbo etc. Even if such chip would run dozens of VMs, that does not mean they all are loaded 100% all of the time, leaving space for turbo.
Cloud guys love density and they are the intended customers for such SoC, they will just invent another word after "hyper" and roll with these uber-converged systems.

 

[Geddagod]

At this point, it's getting real hard to say whether ARL has either RWC+ or LNC. Too many leaks say it can be either.

No, literally only witeken thinks it is RWC+. He isn't even a leaker.

This is A0 stepping.
So, probably SRF will be up to 432 cores.

That attachment you provided is in reference to GNR, no SRF.
SRF is most likely a one tile 144 core product.

 

[lightisgood]

I don't even know how you ended up with 4W for Gracemont at 3.9GHz?
From what you linked, the first graph I see 45.7W when 8 Gracemont cores are loaded, so 5.7W/core and I don't even know If they clocked at 3.9GHz.

The data of 1st image was measured by 12900K (excuse me, I had to mention this).

12900K has power budget as PL1_125W and boasts 3.9GHz E-cores turbo clock.
Up to 46W power consumption assures running at max turbo clock.

17.651W @1core loaded, 45.744W @8cores loaded
45.744 - 17.651 = 28.093 watt

This is pure 7 cores power consumption (i.e. Gracemont@3.9GHz/4W/core ).

What he meant was that what you calculated is just the cores, but not only cores consume power in a CPU and by having little L2/L3, the memory subsystem and interconnects will have to be loaded more often, consuming even more power.

Even same process node, 3.9GHz to 3.0GHz clock regression provides 0.2~0.3x power consumption.
Repeatedly, I guess that Intel will launch SRF-AP as 432cores/ "3~3.5GHz" /500W ...

Could be a bit too optimistic.

So, I think my guess is not optimistic at all, but a modest.

 

[TESKATLIPOKA]

17.651W @1core loaded, 45.744W @8cores loaded
45.744 - 17.651 = 28.093 watt

This is pure 7 cores power consumption (i.e. Gracemont@3.9GHz/4W/core ).

This doesn't make any sense.

I can also do:
17.651W @1core loaded, 32.7W @4cores loaded
32.7 - 17.651 = 15.049 watt

According to your logic this 15W would be pure 3 cores power consumption, in other words 5W per core.

Seriously, If 8 cores loaded consume 46W, then It's ~5.75W per core nothing else. If the cores consumed only 4W, then It would 32W instead of 46W.

 

[Henry swagger]

So what?

Now, here... I have no interest in Zen topics, because I'm sure that we can output the inference about future Lakes & Rapids of Intel from it's own leak , rumor, and so on.

How many cores is sf ap 144 or 288 mild said 288 ?

 

[lightisgood]

This doesn't make any sense.

I can also do:
17.651W @1core loaded, 32.7W @4cores loaded
32.7 - 17.651 = 15.049 watt

According to your logic this 15W would be pure 3 cores power consumption, in other words 5W per core.

Seriously, If 8 cores loaded consume 46W, then It's ~5.75W per core nothing else. If the cores consumed only 4W, then It would 32W instead of 46W.

Okay. I'd like to adoption your saying 6W/core scenario as worst case.
Anyway, clock regression and process node jump resolve the power consumption issue.

6 x 432 x 0.3 x 0.36 = 280 watt
// Gracemont * cores * clock_regression * node_jump = total_power

 

[JoeRambo]

The data of 1st image was measured by 12900K (excuse me, I had to mention this).

12900K has power budget as PL1_125W and boasts 3.9GHz E-cores turbo clock.

Why are we even talking about 12900K? There are E-Core only SKUS like i5-N300 etc that have 8C and 7W TDP that includes iGPU as well.
12900K testing cannot disable all P-Cores, has massive uncore and E-Cores are pushed WAYYYYY beyond their efficiency point and fed voltage that is shared by P-Core and Uncore.

 

[TESKATLIPOKA]

Okay. I'd like to adoption your saying 6W/core scenario as worst case.
Anyway, clock regression and process node jump resolve the power consumption issue.

6 x 432 x 0.3 x 0.36 = 280 watt
// Gracemont * cores * clock_regression * node_jump = total_power

By downclocking the core from 3.9GHz down to 3 GHz you won't reduce power consumption as drastically as you expect.
70% reduction in power consumption would mean 8 cores would consume 8*5.75*0.3=13.8W or 1.725W/core.

A pure Gracemont CPU -> Core i3-N305 with 8 cores has 15W TDP. Its boost is 3.8GHz but basic frequency only 1.8GHz according to Wiki.
Why do you think It's only that much? Because 15W is not enough for 8 cores loaded at higher clocks than that including uncore.

 

[DrMrLordX]

Intel's 144 Core Sierra Forest A0 Stepping Spotted: 108MB L3 Cache and 350W TDP | Hardware Times

The A0 stepping of Intel’s upcoming Xeon Sierra Forest processors has leaked out, courtesy of @yuuki_ans. It includes info regarding two chips: A 96-core and a 144-core CPU. Both have a TDP of 350W, the same as Sapphire Rapids and AMD’s Bergamo cloud offerings. On the memory side, the 96-core...

www.hardwaretimes.com

This is A0 stepping.
So, probably SRF will be up to 432 cores.

Allegedly the 512c and 344c Sierra Forest were cancelled. There never was a 432c Sierra Forest.

 

[JoeRambo]

Why do you think It's only that much? Because 15W is not enough for 8 cores at higher clocks than that including uncore.

I think N300 and N305 are ~1300-1400 ST GB6, and 4400'ish vs 5000'ish MT. So doubling TDP gives minor gains in MT and that most likely means TDP was not limiting factor that much to start with. Remember we are talking about 4C sharing a cluster of L2 and just 6MB of L3.

 

[TESKATLIPOKA]

I think N300 and N305 are ~1300-1400 ST GB6, and 4400'ish vs 5000'ish MT. So doubling TDP gives minor gains in MT and that most likely means TDP was not limiting factor that much to start with. Remember we are talking about 4C sharing a cluster of L2 and just 6MB of L3.

I also found this.
i3 N300: 4015 points in CB R23
i3 N305: 4505 points in CB R23
Weird, what is then the clockspeed? Only ~11% lower by halving TDP?

I found this in one of the reviews for N305. During load only 2.5GHz.Link

Ii will have to find other scores to make any conclusions.

 

[JoeRambo]

I found this in one of the reviews for N305. During load only 2.5GHz.Link

TDP "limitation" 100% depends on workload. They were designed to run CB23 real well and have competent vector FP units. So i'd expect them to drop to base with something like y-Cruncher.
GeekBenches are much more forgiving aside from several subtests and reflect the real world much better.

I think the main limiting factor is tiny L2 capacity when all cores are active - still great for CB23, but 512KB per core is from year 2000 and probably performing even worse due to conflicts. 6MB of L3 from two slices is not exactly strong point either.
I'd expect them to keep good clocks but mostly wait for memory. And also gain 30+% GB6 mt score just from Raptor Lake 4MB L2 setup.

 

[lightisgood]

A pure Gracemont CPU -> Core i3-N305 with 8 cores has 15W TDP. Its boost is 3.8GHz but basic frequency only 1.8GHz according to Wiki.

> Processor Base Power
> The time-averaged power dissipation that the processor is validated to not exceed during manufacturing while executing an Intel-specified high complexity workload at Base Frequency and at the junction temperature as specified in the Datasheet for the SKU segment and configuration.

Source: ark.intel.com


This "Intel-specified high complexity workload" is the so-called torture code, not general workload at all.
As You and I mentioned before, Gracemonte achieves higher clock than base clock in the real processor-intensive workload (libx264, CB23).

 

[TESKATLIPOKA]

> Processor Base Power
> The time-averaged power dissipation that the processor is validated to not exceed during manufacturing while executing an Intel-specified high complexity workload at Base Frequency and at the junction temperature as specified in the Datasheet for the SKU segment and configuration.

Source: ark.intel.com


This "Intel-specified high complexity workload" is the so-called torture code, not general workload at all.
As You and I mentioned before, Gracemonte achieves higher clock than base clock in the real processor-intensive workload (libx264, CB23).

That 1.8GHz is basically a guarantied clock. You can't get any lower than that with any kind of load in my understanding.
CB R23 is not as heavy as that, so It has higher clock in It.
I just want to know what that clocks is for both N300 and N305 to be so close to each other in It.
Then the question is for your calculation what should we use, a load similar to CB R23 or this Intel specific high complexity code.

 

[gdansk]

Isn't it merely what it says on the tin?

 

[Dayman1225]

Yes SRF maxes out at 144c, clear water forest goes higher

 

[SiliconFly]

Yes SRF maxes out at 144c, clear water forest goes higher

512 ?

 

[Abwx]

Isn't it merely what it says on the tin?

Less than18 hours OS booting.?.

What does it mean exactly.?.

And they state excellent silicon health, there s some irony here...

 

[gdansk]

Less than18 hours OS booting.?.

What does it mean exactly.?.

And they state excellent silicon health, there s some irony here...

Edit: misunderstood

 

[SiliconFly]

1. Birch Stream-AP will support up to 500W.
2. Gracemont@3.9GHz roughly consumes 4W/core.
3. Hopping Gracemont to SRF is two node jump on perf/watt gain. Potentially speaking, two node jump will deliver 0.36x power consumption at same clock.

4 x 0.36 x 432 = 622 watt

Therefore, I guess that Intel will launch SRF-AP as 432cores/3~3.5GHz/500W, the phenomenon products.

PPW gain for SF going from Intel 7 to Intel 3 would be 42%