Discussion Intel current and future Lakes & Rapids thread

[nicalandia]

Where are you seeing that they're both running at 15W?

The Gracemont CPU listed there is a direct replacement of the Jasper CPU. and that CPU is rated at 15 watt

Intel® Celeron® Processor N5095 (4M Cache, up to 2.90 GHz) - Product Specifications | Intel

Intel® Celeron® Processor N5095 (4M Cache, up to 2.90 GHz) quick reference with specifications, features, and technologies.

www.intel.com

The leaked Gracemont CPU will be the Intel Celeron N6095, but Geekbench is reading the CPU as Intel N95 instead.

 

[dullard]

We know from ADL silicon that Gracemont clocks way higher than Tremont, in addition to the IPC boost. It's completely nonsensical for the two to perform similarly in MT at iso-power, and some previous results support this conclusion. So either the result is just garbage, or they're not being run at the same power limits.

Where are you seeing that they're both running at 15W?

Geekbench doesn't tell you a lot about the setup. The N95 says it is in the balanced power plan, but the N5095 doesn't have the power plan listed at all. Plus, the power used on both would be highly dependent on other factors that are not controlled here (I assume these were in mini PC form factors where heat buildup is quite a concern. Did both have fans? Are both passively cooled? Etc.) The N5095 that was selected has scores that look to be roughly on par with a balanced power plan, but it would have been nice to know.

Does anyone know if there are effects from different memory amounts (16 GB vs 7.75 GB) or different Geekbench versions (5.4.5 vs 5.4.1)? Also, memory speeds aren't listed which I assume could impact multithreaded benchmarks.

 

[Exist50]

The Gracemont CPU listed there is a direct replacement of the Jasper CPU. and that CPU is rated at 15 watt

View attachment 72899

Intel® Celeron® Processor N5095 (4M Cache, up to 2.90 GHz) - Product Specifications | Intel

Intel® Celeron® Processor N5095 (4M Cache, up to 2.90 GHz) quick reference with specifications, features, and technologies.

www.intel.com

The leaked Gracemont CPU will be the Intel Celeron N6095, but Geekbench is reading the CPU as Intel N95 instead.

Just because it's likely the successor to a 15W SKU doesn't mean this particular SKU was run at 15W in this test. Or, as I said, the result could be off for some other reason.

Seriously, why are you debating this? We've had Gracemont results from real silicon for ages now. It's way better than Tremont across the board.

 

[BorisTheBlade82]

Isn't it the most likely that they just run in the TDP limit and that they basically have the same efficiency - just on different base clocks, as already mentioned by @nicalandia ?
Jasper Lake was also on an admittedly earlier version of 10nm version. So it would indeed be a bit disappointing but not entirely out of this world.

 

[nicalandia]

Isn't it the most likely that they just run in the TDP limit and that they basically have the same efficiency - just on different base clocks, as already mentioned

I am not sure why it's so hard to grasp. Same TDP limit and about 25% IPC boost. Sounds about right.

 

[Exist50]

Isn't it the most likely that they just run in the TDP limit and that they basically have the same efficiency - just on different base clocks, as already mentioned by @nicalandia ?
Jasper Lake was also on an admittedly earlier version of 10nm version. So it would indeed be a bit disappointing but not entirely out of this world.

We know that 10ESF (or whatever they're calling it now) clocks way better than the 10nm Ice Lake process that Tremont uses. We saw those gains from Ice Lake to Tiger Lake, much less Alder Lake. So why do we see a regression here? Additional power intrinsic to a bigger core would explain some iso-process regression, but the process improvement should more than compensate.

Let's do some math, shall we? At 15W, the i7-1065G7 has a base frequency of 1.3GHz. At the same TDP, the i7-1160G7 has a base frequency of 2.1GHz. These are practically the same architecture, just on a different process. So at the same power per core and same architecture, we should expect ~60% higher speeds from process alone. Yet we see 15% lower. Or instead of 2.00GHz * 1.6 = 3.2GHz, we see 1 - 1.70/3.20 = ~50% less than expected.

If we approximate cubic power scaling with frequency, then half the frequency => 1/8th the power. Gracemont as an architecture obviously doesn't take 8x the power of Tremont, so even giving plenty of error to the approximation, it really makes no sense.

Or better yet, we have actual numbers!

Alder Lake’s Power Efficiency – A Complicated Picture

Reviews across the internet show Alder Lake getting very competitive performance with very high power consumption.

chipsandcheese.com

At ~14W of power to the cores, Gracemont runs around 3.1GHz (note, almost exactly in line with the extrapolation above). So again, how are earth is it down to 1.7GHz? Uncore power doesn't explain such a gap.

God I wasted too much time on this...

I am not sure why it's so hard to grasp. Same TDP limit and about 25% IPC boost. Sounds about right.

At the frequencies listed, yes, but why are they so low? That's my question.

 

[dullard]

At the frequencies listed, yes, but why are they so low? That's my question.

You are looking at the first public benchmark of the lowest of the low chips (formerly Celeron). I wouldn't expect much from that. You certainly shouldn't be comparing them to the frequencies from i7 chips. There is supposedly an N100 and N200 coming too.

 

[nicalandia]

At the frequencies listed, yes, but why are they so low? That's my question.

I understand your point now. But I am just interested in IPC Gains from Tremont to Gracemont.

 

[nicalandia]

4C/4T Gracemont Cores(2.8 Ghz ST and 1.70 Ghz base) vs 2C/2T Skylake Cores(2.8 Ghz ST and 2.9 Ghz base)

Default string Default string vs System manufacturer System Product Name - Geekbench

Gracemont showing Skylake level IPC in ST and MT at Same speed

 

[fgxvcdsm]

I need to get caught up: I keep hearing about an Intel 8-core for mainstream/client platform. Someone's musing, rumor, or evidence of such hardware?

I heard that intel will introduce Direct
Memory access in the near future due to which pc would become much faster.

 

[Markfw]

I heard that intel will introduce Direct
Memory access in the near future due to which pc would become much faster.

Where did you hear that ? source ?

 

[nicalandia]

Yep, but Sapphire Rapid is at least 3 years away. Stuff could change.

That was back on 2017.... 13+ Stepping letter it will finally be released on 2023

 

[DrMrLordX]

That was back on 2017.... 13+ Stepping letter it will finally be released on 2023

Epic necro post.

 

[moinmoin]

Epic necro post.

The real funny part is what jpiniero's post was in response to:

Isn't Intel working on a complete revamp of X86 for post tiger lake chips? I read some new stories about it dated from last year. Supposedly leaner, faster and all that kind of stuff.

Sapphire Rapid is everything but leaner to me.

But I do recall that Sapphire Rapid was regularly brought up as the gen where Intel would finally overhaul the core. Still didn't happen either (unless that always has been about SDSi/Intel On Demand).

The longevity of this thread is legendary though.

 

[DrMrLordX]

The real funny part is what jpiniero's post was in response to:

Sapphire Rapid is everything but leaner to me.

But I do recall that Sapphire Rapid was regularly brought up as the gen where Intel would finally overhaul the core. Still didn't happen either (unless that always has been about SDSi/Intel On Demand).

The longevity of this thread is legendary though.

Yeah I have to admit, Sapphire Rapids is shaping up to be a disappointment in a lot of ways. At this point, radical redesigns of Core don't seem to be on the table at all.

 

[nicalandia]

First Apples to Apples Comparison of High Core Sapphire Rapids and High Core Milan than.

Intel Xeon Platium 8481C(Cloud version of the 56C/112T CPU)

AMD EPY 7B13 (64C/128T Cloud Milan)

Google Google Compute Engine vs Google Google Compute Engine - Geekbench

 

[Markfw]

First Apples to Apples Comparison of High Core Sapphire Rapids and High Core Milan than.

Intel Xeon Platium 8481C(Cloud version of the 56C/112T CPU)

AMD EPY 7B13 (64C/128T Cloud Milan)

View attachment 73129

Google Google Compute Engine vs Google Google Compute Engine - Geekbench

What is the power utilization of the 2 different setups ? That means more than a small difference in these scores.

 

[nicalandia]

What is the power utilization of the 2 different setups ? That means more than a small difference in these scores.

Unknown since it's a cloud setup hidden from the user. Intel Set Up has 22 Cores configured and The Milan has 24 cores(could not find anything lower than 24 cores and higher than 16 for the Milan part)

 

[nicalandia]

Some people in Twitter...



No, Alder Lake draws More power when using AVX512

 

[Markfw]

As far as Alder lake taking less power than AVX2, not sure, but my 12700F with e-cores disabled did run avx-512 , but the power went over 300 watts !

Yes, I agree with the post above.

 

[nicalandia]

As far as Alder lake taking less power than AVX2, not sure, but my 12700F with e-cores disabled did run avx-512 , but the power went over 300 watts !

Yes, I agree with the post above.

You would think that if it was true(that it draws less power) it would have made headlines news all of the world.. "AVX-512 actually makes the 12900K more efficient, why is Intel does not enable it by default?"

But, since it actually draws more power nobody cared or made a fuss about not being enabled by default.

 

[Exist50]

As far as Alder lake taking less power than AVX2, not sure, but my 12700F with e-cores disabled did run avx-512 , but the power went over 300 watts !

The 12700f has a 180W PL2. Yours was not taking that much power, and it's baffling that you keep insisting otherwise.

You would think that if it was true(that it draws less power) it would have made headlines news all of the world.. "AVX-512 actually makes the 12900K more efficient, why is Intel does not enable it by default?"

But, since it actually draws more power nobody cared or made a fuss about not being enabled by default.

There literally were articles like that: https://www.igorslab.de/en/efficien...ke-the-returned-command-set-in-practice-test/

But it turns out, not many people care about HPC benchmarks on client chips. Though it is funny seeing how many people immediately pivoted from AVX512 being useless to essential.

 

[nicalandia]

...

There literally were articles like that: https://www.igorslab.de/en/efficien...ke-the-returned-command-set-in-practice-test/

There has been numerous test to show how power hungry AVX-512 not only on Server but Client(all found on Phoronix) and you are countering that with a Igor's Lab benchmark?

 

[nicalandia]

Wrong thread

 

[Exist50]

...

There has been numerous test to show how power hungry AVX-512 not only on Server but Client(all found on Phoronix) and you are countering that with a Igor's Lab benchmark?

That article was in response to your claim that no one was talking about AVX512 efficiency on Alder Lake etc.

More to the point, where AVX512 is power hungry, it also provides a lot more performance, and thus, significantly better efficiency. That power consumption isn't coming from nothing. If you want to talk efficiency, you can't just ignore the performance half the equation.