The Intel Execution Thread [2021]

Status
Not open for further replies.

Rootax

Veteran
I don't know where to post this, but, I've a dumb question about Intel...

I think we can all agree that the delay of 10nm is the main reason for their trouble (even if they did a hell of a job pushing 14nm this far). I always wondered why they didn't release a cpu without igpu. I don't mean just disabled, but not present at all, to gain some space/transistors ?

I believe that the igpu presence is a main factor for Dell, hp, etc for desktop products, but for people buying cpu separatly / gamer ?

I guess it's kind of moot now that alderlake is coming on 10nm at last.
 
I guess it's kind of moot now that alderlake is coming on 10nm at last.
It's not moot. Without iGPU ADL would probably be able to pack 16 "big" cores in the same area as it will have with 8 "big" and 8 "little" and iGPU.
Intel's insistence on putting iGPUs in their high end desktop processors has been very weird for some time now.
 
I don't know where to post this, but, I've a dumb question about Intel...

I think we can all agree that the delay of 10nm is the main reason for their trouble (even if they did a hell of a job pushing 14nm this far). I always wondered why they didn't release a cpu without igpu. I don't mean just disabled, but not present at all, to gain some space/transistors ?

I believe that the igpu presence is a main factor for Dell, hp, etc for desktop products, but for people buying cpu separatly / gamer ?

I guess it's kind of moot now that alderlake is coming on 10nm at last.
People buying separately / gamers are drop in the ocean of OEM sales and re-designing whole chips isn't cheap (they can't just cut the GPU off and call it a day)
 
I always wondered why they didn't release a cpu without igpu. I don't mean just disabled, but not present at all, to gain some space/transistors ?
Cuz commercial.
iGP will be mandatory on AMD platforms starting with AM5, too.
but for people buying cpu separatly / gamer ?
That's a very, very small niche both vendors do mostly for goodwill and kicks.
Without iGPU ADL would probably be able to pack 16 "big" cores in the same area as it will have with 8 "big" and 8 "little" and iGPU.
Yeah good luck cooling 16 GLCs in pkg that small.
lol
 
Yeah good luck cooling 16 GLCs in pkg that small.
lol
16C 10SF GLC design wouldn't be exactly "small".
And as for cooling 4C TGL is able to run at 4.3GHz on all cores in 28W power.
As naive as it is 28Wx4=112W for a similar 16C part.
Don't see much problems with cooling a 112W CPU.
"lol"
 
16C 10SF GLC design wouldn't be exactly "small".
And as for cooling 4C TGL is able to run at 4.3GHz on all cores in 28W power.
As naive as it is 28Wx4=112W for a similar 16C part.
Don't see much problems with cooling a 112W CPU.
"lol"
Man you surely read the brains of people doing the ADL-S bringup and figured out the GLC target power I c.
Also 4.3 4c TGL is 50W, see 11375H.
The actual TGL-U42 runs low 3s loaded at 28W.
 
When there is a regression of performance on the same node compared to an architecture from 2015 then it's not an issue of the node.
The regressions seem to come from the botched L3 cache and IMC (higher latency than Comet Lake and has something like Zen 2/3’s FCLK where running memory above 3733 MT/s incurs another 10ns+ latency penalty). Andrew Frumusanu of AT speculates that the L3 cache is a consequence of backporting the core.
Rocket Lake just seems like it has no place in the market. It can’t touch the value of Comet Lake with current discounts, nor the raw performance of Zen 3. Seems like a waste of effort if Alder Lake-S is really coming in Q4 2021.
 
Andrew Frumusanu of AT speculates that the L3 cache is a consequence of backporting the core.
It's an issue of Intel's engineers doing god know what recently. Again, if an architecture from 2015 is better in these areas then it's not an issue of a process or backporting or anything - they've just fucked up with the design, plain and simple.
 
I see quite a few tests where it is substantially faster. Interesting results though.
 
Last edited:
The power usage also spikes through the roof too, if I parsed the AT Preview correctly.
Up to 295 or so watts. Just terrible. If you want an Intel CPU and reasonable temps you're forced to use water cooling. Even their 5lb air cooler couldn't keep temps under 100c.
 
It's an issue of Intel's engineers doing god know what recently. Again, if an architecture from 2015 is better in these areas then it's not an issue of a process or backporting or anything - they've just fucked up with the design, plain and simple.

But it' not the original design. Backporting is not free, they must have changed things last minute to make it work imo.
I wonder why they even released this since Alder Lake seems on the starting block...
 
But it' not the original design. Backporting is not free, they must have changed things last minute to make it work imo.
I wonder why they even released this since Alder Lake seems on the starting block...

Is it bold to assume that their goal with the next product line will be an overall improvement rather than reaching parity with their previous products ?
 
Is it bold to assume that their goal with the next product line will be an overall improvement rather than reaching parity with their previous products ?
Rocket Lake is really just a sad filler product since Skylake reached its limits in increasing the core count and frequency. Back in Tick-Tock days, a 10nm part was planned right after the original 2015 Skylake.

For RL Intel took a 2 years old Ice Lake core and ported that to a 7 years old 14nm process. AL is manufactured at brand new 10nm SFF+, features Golden Cove cores with a rumored ~20% IPC increase over current Tiger/Ice Lake, and is rumored to be presented in September. To me RL and AL are quite different.
 
Up to 295 or so watts. Just terrible. If you want an Intel CPU and reasonable temps you're forced to use water cooling. Even their 5lb air cooler couldn't keep temps under 100c.
It's firing on all cylinders, with not only record power usage but also record performance in that single hand-tuned AVX512-benchmark of Ian's.
It's 32k points for 290 watts, while most of the other 8-core-processors sit at 5-6k points. Which in itself seems like there's more to those optimizations than just increased vector width (2x) and half as much decoding load (0.5x): AVX512, if you ask the right persons, is much more than 2x AVX2.

I think, this enormous power draw figure is more a testament to the board freely extending boost times and clock allowances according to the capabilities of it's VRM array, as Ian duly notes. AMD is much more restrictive with this, unless you manually overclock. And - IMHO - that's how it should be handled.
 
Status
Not open for further replies.
Back
Top