Discussion Intel current and future Lakes & Rapids thread

[SiliconFly]

Exactly, which is part of what I said. Intel will likely not use any N3 node for Arrow Lake.



Parallel development would be the only way they could effectively switch the 8+16 die to 20a, yes.

Parallel development isn't a big issue for Intel anymore. Remember Jim Keller's involvement... starting from LNC all Intel core designs are purely node-agnostic. Makes it a lot easier to port from one node to another.

And considering some of the ARL designs are already on 20A, I don't think it's that difficult to do 8+16 on 20A. Main issues are yield and volume. Even if the yield is okay, they still might not make volume due to capacity. It's difficult.

 

[TESKATLIPOKA]

TPU tested i9-14900K at different power limits down to 35W. Check It out.
Performance suffers a lot, but efficiency is great.

Spoiler: Example

 

[Mopetar]

Basically the same story as with the 12 and 13000 series. You can get ~90% of the performance for nearly half the power in most cases.

 

[Geddagod]

Something interesting is that GNR appears to have taped out, in Q3 2023, while tape in was Q1 2022. So roughly 1.5 years between tape in and tape out, and an year between tape out and launch.
In total, you are seeing ~2.5 years between tape in and launch. So maybe a couple months shaved off the dev schedule when compared to ICL/SPR. Not exactly inspiring.

 

[H433x0n]

Something interesting is that GNR appears to have taped out, in Q3 2023, while tape in was Q1 2022. So roughly 1.5 years between tape in and tape out, and an year between tape out and launch.
In total, you are seeing ~2.5 years between tape in and launch. So maybe a couple months shaved off the dev schedule when compared to ICL/SPR. Not exactly inspiring.

I’d argue that ICL is not an apples/apples comparison. SPR taped in December 2019 to release in ~January 2023, which put it at just over a full 3 years. We’ll also get an idea of improvement by looking at final production stepping. If GNR release stepping is something like C0 or C1, that’s certainly an improvement over SPR.

If they consistently hit 2.5 year dev cycle time from generation to generation - I think it’d be okay (while not ideal). It really just needs to be a consistent.

 

[SiliconFly]

After seeing the following HY video, I'd like to reiterate:

Intel 14th gen desktop CPUs are an outdated monolithic design based on an outdated RPL-R architecture built on an outdated Intel 7 node and none of these should even exist in this day and age. It's a shame they released such a product in 2023. HY spelt it out clearly in no uncertain terms: youtube link

MTL on the other hand is something we can looking forward to. An amazing tile design. But cpu tile is still based on RWC which sadly is a rehash of RPL and is built on the slightly dated Intel 4 node. But the increase in efficiency & powerful iGPU is enough to make it standout.

But the real excitement starts next year with ARL. An all new architecture (LNC) built on a new advanced node 20A. I think ARL is gonna take efficiency to the next level. Unquestioned leadership? Maybe... maybe not... too early to speculate (but might even match Apple silicon). Just sayin.

But the 14th gen desktop parts in 2023 is a bad dream I wish to forget. May the almighty stop Intel from flooding the market with outdated parts. Looking forward to MTL & ARL.

 

[Markfw]

After seeing the following HY video, I'd like to reiterate:

Intel 14th gen desktop CPUs are an outdated monolithic design based on an outdated RPL-R architecture built on an outdated Intel 7 node and none of these should even exist in this day and age. It's a shame they released such a product in 2023. HY spelt it out clearly in no uncertain terms: youtube link

MTL on the other hand is something we can looking forward to. An amazing tile design. But cpu tile is still based on RWC which sadly is a rehash of RPL and is built on the slightly dated Intel 4 node. But the increase in efficiency & powerful iGPU is enough to make it standout.

But the real excitement starts next year with ARL. An all new architecture (LNC) built on a new advanced node 20A. I think ARL is gonna take efficiency to the next level. Unquestioned leadership? Maybe... maybe not... too early to speculate (but might even match Apple silicon). Just sayin.

But the 14th gen desktop parts in 2023 is a bad dream I wish to forget. May the almighty stop Intel from flooding the market with outdated parts. Looking forward to MTL & ARL.

I can't believe I am agreeing with you, but facts are facts !

 

[H433x0n]

Intel 14th gen desktop CPUs are an outdated monolithic design

I’ve got to push back on 1 point: a monolithic design is not inherently outdated. For a client CPU, it’s the best overall option. All of these desktop SKUs are relatively small dies where a monolithic design doesn’t really have any downsides other than a slight cost difference (but even that is debatable depending on the type of packaging used).

 

[poke01]

After seeing the following HY video, I'd like to reiterate:

Intel 14th gen desktop CPUs are an outdated monolithic design based on an outdated RPL-R architecture built on an outdated Intel 7 node and none of these should even exist in this day and age. It's a shame they released such a product in 2023. HY spelt it out clearly in no uncertain terms: youtube link

MTL on the other hand is something we can looking forward to. An amazing tile design. But cpu tile is still based on RWC which sadly is a rehash of RPL and is built on the slightly dated Intel 4 node. But the increase in efficiency & powerful iGPU is enough to make it standout.

But the real excitement starts next year with ARL. An all new architecture (LNC) built on a new advanced node 20A. I think ARL is gonna take efficiency to the next level. Unquestioned leadership? Maybe... maybe not... too early to speculate (but might even match Apple silicon). Just sayin.

But the 14th gen desktop parts in 2023 is a bad dream I wish to forget. May the almighty stop Intel from flooding the market with outdated parts. Looking forward to MTL & ARL.

Intel is "dead" till Arrow Lake and lunar lake

 

[igor_kavinski]

Intel is "dead" till Arrow Lake and lunar lake

Nah. Nova Lake is supposed to be the real deal. Those two won't move the performance needle much.

 

[lightisgood]

Intel is "dead" till Arrow Lake and lunar lake

Your view is simply lacking Gaudi3, Granite Rapids, Sierra Forest and so on...

 

[igor_kavinski]

Your view is simply lacking Gaudi3, Granite Rapids, Sierra Forest and so on...

Expensive boring stuff. Meh.

 

[poke01]

Your view is simply lacking Gaudi3, Granite Rapids, Sierra Forest and so on...

I don't care for server/AI products that are not for the client/consumer.

The point is Intel does not have a efficient product in client until the end of 2024

 

[lightisgood]

I don't care for server/AI products that are not for the client/consumer.

The point is Intel does not have a efficient product in client until the end of 2024

I don't think so.
Meteor Lake would be good product.

 

[Abwx]

Basically the same story as with the 12 and 13000 series. You can get ~90% of the performance for nearly half the power in most cases.

Those numbers are made up and do not make sense at all, no way that 11% better perfs require 100% more power when starting from 125W.

At 125W you get 83% of the stock perf of a 13900K@253W, at 142W there s 87% remaining perf, so still quite below TPU s fairy tales tests that put RPL at same efficiency than a 7950X...

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

 

[H433x0n]

Those numbers are made up and do not make sense at all, no way that 11% better perfs require 100% more power when starting from 125W.

At 125W you get 83% of the stock perf of a 13900K@253W, at 142W there s 87% remaining perf, so still quite below TPU s fairy tales tests that put RPL at same efficiency than a 7950X...

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

They're not made up.. they're testing different things. The TPU test isn't strictly a multithreaded test since it includes things like JetStream and Microsoft Apps.

The ComputerBase test is 10 total benchmarks with 7 of them focused on rendering. It's essentially a render benchmark (3 of the 10 are cinebench).

 

[TESKATLIPOKA]

Those numbers are made up and do not make sense at all, no way that 11% better perfs require 100% more power when starting from 125W.

At 125W you get 83% of the stock perf of a 13900K@253W, at 142W there s 87% remaining perf, so still quite below TPU s fairy tales tests that put RPL at same efficiency than a 7950X...

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

There is no problem with that test, you just didn't bother looking at what apps were used.

Here you have CB R23 and the scores looks fine.

Spoiler: Cinebench R23

 

[Abwx]

There is no problem with that test, you just didn't bother looking at what apps were used.

Here you have CB R23 and the scores looks fine.

Spoiler: Cinebench R23

I was commenting the claim by a member that at 125W there was still 90% of the stock perf available, in this CB run from TPU that s 81% of the stock perf at 125W, i pointed Computerbase who measure 83% using about 10-15 benches and a 13900K.

 

[Abwx]

They're not made up.. they're testing different things. The TPU test isn't strictly a multithreaded test since it includes things like JetStream and Microsoft Apps.

The ComputerBase test is 10 total benchmarks with 7 of them focused on rendering. It's essentially a render benchmark (3 of the 10 are cinebench).

5 are rendering, CB R20 and R23, Povray, Blender, Corona, 4 others are 7 Zip, Agisoft, Digicortex and Handbrake.

Beside measuring efficency in lowly threaded apps is non sense, at this rate you can say that for a single thread app the CPU will have the same efficency at 125W than at 253W, hence 100% better efficency but without mentioning that the CPU use only 50W in ST...

 

[Goop_reformed]

Your view is simply lacking Gaudi3, Granite Rapids, Sierra Forest and so on...

Granite rapids is at least 2 quarters after Turin, also slower. Same with Sierra Forrest vs Bergamo

 

[H433x0n]

5 are rendering, CB R20 and R23, Povray, Blender, Corona, 4 others are 7 Zip, Agisoft, Digicortex and Handbrake.

Beside measuring efficency in lowly threaded apps is non sense, at this rate you can say that for a single thread app the CPU will have the same efficency at 125W than at 253W, hence 100% better efficency but without mentioning that the CPU use only 50W in ST...

It's not simply lightly threaded applications - it's a wide range of scenarios. I would argue it's a more representative sample of CPU performance than a bunch of rendering applications that are trivially parallel and better offloaded to a GPU.

None of these results look particularly suspect to me and they all scale with power / architecture.

Spoiler: Benchmarks

 

[TESKATLIPOKA]

I was commenting the claim by a member that at 125W there was still 90% of the stock perf available, in this CB run from TPU that s 81% of the stock perf at 125W, i pointed Computerbase who measure 83% using about 10-15 benches and a 13900K.

It doesn't look like you were just correcting what @Mopetar said wrongly, but also criticized TPU's test in the same post as shown below.

Those numbers are made up and do not make sense at all, no way that 11% better perfs require 100% more power when starting from 125W.

At 125W you get 83% of the stock perf of a 13900K@253W, at 142W there s 87% remaining perf, so still quite below TPU s fairy tales tests that put RPL at same efficiency than a 7950X...

Intel Core i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks

i9-14900K, i7-14700K & i5-14600K im Test: Anwendungs-Benchmarks / Testmethodik / Multi-Core-Leistung / Single-Core-Leistung

www.computerbase.de

So I stand by what I wrote, there is no problem with that test.

 

[SiliconFly]

I don't think so.
Meteor Lake would be good product.

Is MTL a good product? I think so too. But definitely a lot better than previous gen.

ARL is gonna be the real deal I believe (assuming they can hit yield & volume on 20A).

Intel is "dead" till Arrow Lake and lunar lake

I'm pretty sure MTL will save their day until ARL comes out. 12 months isn't that long a time.

(As a side note: LNL is a niche product. More of a test vehicle.)

 

[SiliconFly]

Nah. Nova Lake is supposed to be the real deal. Those two won't move the performance needle much.

Too early to say. Looks like ARL is also all about efficiency like MTL (which isn't a bad thing I'd say).

But can't totally rule out ARL performance gains this early. LNC might move the needle a bit.

 

[Geddagod]

Parallel development isn't a big issue for Intel anymore. Remember Jim Keller's involvement...

Don't think Keller had anything to do with that. IIRC Keller claimed most of what he had to do with is the organizational changes of Intel.

starting from LNC all Intel core designs are purely node-agnostic. Makes it a lot easier to port from one node to another.

SNC actually. Intel claimed that even before RKL launched IIRC.
If you are talking about Intel + TSMC, then crestmont might be the first. Though I'm also not sure there aren't any irrelevant core ip (or tbh even gpu ip) that also wasn't straight up ported between the two fabs.

I think ARL is gonna take efficiency to the next level. Unquestioned leadership? Maybe... maybe not... too early to speculate

If ARL doesn't win the perf/watt crown, it's going to be extremely embarrassing for Intel, esp at lower power levels. I'm still holding out hope LNC isn't completely a bust, and that even if they threw away a bunch of perf, they might have kept the good perf/watt scaling aspect of the arch, but who knows.