Discussion Intel current and future Lakes & Rapids thread

[IntelUser2000]

Do we know that this holds true for ADL though?

It shouldn't matter a whole lot. It's more than a victim cache for Intel chips, that's the whole point of the ring bus.

Also the mobile AMD CPUs perform few % less than on desktops. That's clear from SPEC comparisons.

I think Intel will lead significantly on the H, as Zen 3+ is few single digit % faster and it has similar clocks. It's the U segment where Remembrandt improves a lot in regards to clocks.

Because of DDR5 memory subsystem - Rembrandt will have also have it.

No. And this goes all the way back to Nehalem. Back when they had L2 as LLC such as Core 2 it mattered more. Even then it was 5-7% for doubling the size.

 

[insertcarehere]

If we're gonna use SPEC for IPC then let's not pretend that Zen 3 desktop with double the L3 per CCD is the most valid comparison for RMB, when Anandtech has SPEC figures for Cezanne (5950x is speced at 4.9 ghz boost vs 5980HS w/ 4.8ghz boost):

 

[eek2121]

That shows in SPECint that ADL is 4.7% faster with ~3% faster clock. Very little difference in IPC between the two.

To quote AnandTech:

"Alder Lake and the Golden Cove cores are able to reclaim the single-threaded performance crown from AMD and Apple. The increases over Rocket Lake come in at +18-20%, and Intel’s advantage over AMD is now at 6.4% and 16.1% depending on the suite, maybe closer than what Intel would have liked given V-cache variants of Zen3 are just a few months away. "

The 6.4% is Integer, the 16.1% is Floating Point.

 

[IntelUser2000]

The larger caches for Intel chips are partially negated by the slightly higher latency due to extra ring stops.

I mean, the L3 cache might help situationally(like in modern fat games) but when it comes to CPU Integer performance I pretty much ignore it on the Intel side.

SpecInt is the decent one. SpecFP I ignore as it's a bandwidth test.

 

[Hitman928]

How do you figure?

SPECint 2017 - Zen3 vs. Golden Cove (hopefully no typos)

500.perlbench_r: GC is 22.69% faster
502.gcc_r: GC is 17.18% faster
505.mcf_r - memory latency bound, both are tied.
520.omnetpp_r: GC is 6.47% faster
5.23.xalancbmk_r: GC is 10.21% faster
5.25.x264_r: GC is 13.07% faster
5.31.deepsjeng_r: GC is 6.1% faster
5.41.leela_r: Zen 3 is 3.63% faster
5.48.exchange2_r: GC is 4.58% faster
5.56.xz_r: Zen 3 is 9.54% faster

Zen 3 only wins in 2 of those tests above.

There is a 'total' graph that takes all of the individual charts into account. The total graph is the actual SPECint or SPECTfp score.

 

[mikk]

Golden Cove has roughly 10% more IPC, some people really don't get. Also I have to ask how relevant will be Rembrandt in the 15W space against ADL-U? There are only two Rembrandt-U models and they only support DDR5/LPDDR5 which is premium. 15W ADL-U is likely targeted for cheaper notebooks with less capable cooling whereas the faster 28W models with more cores and clock speed goes into the premium laptop series. I really doubt Rembrandt with DDR5/LPDDR5 is a competitor to the cheaper ADL-U 2+8.

 

[naukkis]

It shouldn't matter a whole lot. It's more than a victim cache for Intel chips, that's the whole point of the ring bus.

Alderlake L3 is also victim cache, and Zen3 has also ringbus L3.

Also the mobile AMD CPUs perform few % less than on desktops. That's clear from SPEC comparisons.

I think Intel will lead significantly on the H, as Zen 3+ is few single digit % faster and it has similar clocks. It's the U segment where Remembrandt improves a lot in regards to clocks.

Yes IPC is a bit lower than desktop but Rembrandt has higher max boost clocks than any AMD desktop chip. And for halved L3 Rembrandt have ondie DDR5-memory controller instead of offdie DDR4 as in desktops.

And Intel's biggest problem is that Goldencove consumes way too much power for mobile - look those Intel published benchmarks - for their 14c Alderlake they won't promise more than ~45% more Mt performance against their previous 8C model - that's not a great achievement.

 

[eek2121]

Yes, and I listed the total difference above. 6.4% vs. 16.1%. On top of that, as @insertcarehere has stated, the mobile variants are slower due to having half the cache (as well as lower TDP, etc.) ADL-P is cut down as well, but not to the extent that the AMD chips are. Further, all this talk is excluding the "E" cores and only counts the "P" cores. The intel chips have "E" cores which will help it win out in multicore workloads.

 

[Hitman928]

To quote AnandTech:

"Alder Lake and the Golden Cove cores are able to reclaim the single-threaded performance crown from AMD and Apple. The increases over Rocket Lake come in at +18-20%, and Intel’s advantage over AMD is now at 6.4% and 16.1% depending on the suite, maybe closer than what Intel would have liked given V-cache variants of Zen3 are just a few months away. "

The 6.4% is Integer, the 16.1% is Floating Point.

It's 6.4% on integer when ADL is using DDR5 due to some tests being more sensitive to memory bandwidth. When using same DDR4 as Zen3, ADL is 4.7% faster. Both ADL and RMB support DDR5 so ADL won't have the same advantage so it is more apt to use the 4.7% figure.

 

[IntelUser2000]

The 6.4% is Integer, the 16.1% is Floating Point.

The SpecFP test for client is irrelevant as it's too memory bandwidth intensive. It's relevant for HPC but we do not care about that. It does not matter if you are using AVX or not.

Even in FP for client SpecInt is the way to go, because that shows uarch differences rather than focusing on specific parts of the CPU(such as the FPU).

 

[eek2121]

Yes IPC is a bit lower than desktop but Rembrandt has higher max boost clocks than any AMD desktop chip. And for halved L3 Rembrandt have ondie DDR5-memory controller instead of offdie DDR4 as in desktops.

My 5950X boosts to 5050 MHz at stock, so I wouldn't say that.

 

[IntelUser2000]

And Intel's biggest problem is that Goldencove consumes way too much power for mobile - look those Intel published benchmarks - for their 14c Alderlake they won't promise more than ~45% more Mt performance against their previous 8C model - that's not a great achievement.

And you ignore that 8 of those cores are E cores? That's a convenient excuse to downplay the changes. 45% gain in the same process is a lot, considering the predecessor wasn't exactly slow. It is a great achievement. The die area is also barely bigger than Tigerlake-H.

 

[naukkis]

And you ignore that 8 of those cores are E cores? That's a convenient excuse to downplay the changes. 45% gain in the same process is a lot, considering the predecessor wasn't exactly slow. It is a great achievement.

45% figure was for desktop, those mobile gains are more like 10-30%. It's not even in same process, process is improved. That's not a great achievement and because of that AMD MT gains with Rembrandt match those Intel's.

 

[IntelUser2000]

45% figure was for desktop, those mobile gains are 20-30%. It's not even in same process, process is improved. That's not a great achievement and because of that AMD MT gains with Rembrandt match those Intel's.

AMD's claims with Remembrandt is for the 6800U, because the clocks are much higher.

The H chips clock identical. Meaning maybe it's 5-10% faster.

Also when I quoted you, you were talking about the laptop models. They also claim 44% gain for the laptop SKU.

What you said.

for their 14c Alderlake they won't promise more than ~45% more Mt performance against their previous 8C model - that's not a great achievement.

 

[Hitman928]

Further, all this talk is excluding the "E" cores and only counts the "P" cores. The intel chips have "E" cores which will help it win out in multicore workloads.

I think the E cores are clocked too low with too little IPC to make up the difference at the U level. The P level will be the most interesting battleground, IMO.

 

[IntelUser2000]

I think the E cores are clocked too low with too little IPC to make up the difference at the U level. The P level will be the most interesting battleground, IMO.

I think the 2+8 15W might be similar to 8+8 S where the efficiency falls off. 2+8 is better suited for 9W. Should have seen at least 1 4+8 15W model.

 

[mikk]

12400 is limited to 4000 Mhz in MT and this is enough to match 5600x running at 4500-4600 Mhz. This is because Golden Cove has the better IPC.

 

[uzzi38]

The larger caches for Intel chips are partially negated by the slightly higher latency due to extra ring stops.

I mean, the L3 cache might help situationally(like in modern fat games) but when it comes to CPU Integer performance I pretty much ignore it on the Intel side.

SpecInt is the decent one. SpecFP I ignore as it's a bandwidth test.

Both FP and INT gain a lot from DDR5



~33% on ADL-S.

 

[Hitman928]

I think the 2+8 15W might be similar to 8+8 S where the efficiency falls off. 2+8 is better suited for 9W. Should have seen at least 1 4+8 15W model.

I really hope someone gets individual core power numbers on ADL mobile CPUs for both E and P cores. It will be really interesting to see.

 

[IntelUser2000]

AMD MT gains with Rembrandt match those Intel's.

The 30% gain they claim is 28W Rembrandt vs 15W Cezanne.

15W vs 15W they claim 17% CPU and 81% GPU. CPU is respectable but that's on the U.

Intel is totally screwed in the iGPU department until Meteor Lake and I bet you since that's nearly 2 years away AMD will have an answer for it too.

After all the fanfare regarding Iris Xe, they just coast. Maybe I'll call it the "Vega effect" as that's what AMD did with mobile Vega.

Both FP and INT gain a lot from DDR5

Rate benchmarks stress all cores and thus much more bound but we're also talking 15W processors here. LPDDR4x didn't anything on the CPU side for Tigerlake.

 

[dullard]

Intel is totally screwed in the iGPU department until Meteor Lake and I bet you since that's nearly 2 years away

I've never seen anything putting Meteor Lake into late 2023 / early 2024. I've seen multiple things putting Meteor Lake as soon as 1 year away (probably more like 5 quarters, I would bet). Intel 7 ramped up Q3 2021 and we are finally getting most chips Q1 2022. Intel 4 is supposed to be just 1 year later.

https://images.anandtech.com/doci/16823/AnandTechRoadmaps3.png

 

[IntelUser2000]

I've never seen anything putting Meteor Lake into late 2023 / early 2024.

Except there's Raptor Lake, and we just had Alderlake mobile announced.

 

[Exist50]

I've never seen anything putting Meteor Lake into late 2023 / early 2024. I've seen multiple things putting Meteor Lake as soon as 1 year away (probably more like 5 quarters, I would bet). Intel 7 ramped up Q3 2021 and we are finally getting most chips Q1 2022. Intel 4 is supposed to be just 1 year later.

https://images.anandtech.com/doci/16823/AnandTechRoadmaps3.png

Meteor Lake should start shipping to OEMs what? H1'23? Add another quarter before designs show up, and Q3 or so is quite believable.

 

[dullard]

Except there's Raptor Lake, and we just had Alderlake mobile announced.

Raptor Lake is this year. "01:16PM EST - Next Gen Raptor Lake already booting, coming later in 2022" https://www.anandtech.com/show/1717...bryant-at-ces-2022-live-blog-10am-pt-1800-utc

 

[IntelUser2000]

Raptor Lake is this year. "01:16PM EST - Next Gen Raptor Lake already booting, coming later in 2022" https://www.anandtech.com/show/1717...bryant-at-ces-2022-live-blog-10am-pt-1800-utc

Yes, late this year or early next year for mobile.

Since Meteor Lake is mobile focused I cannot see how it'll arrive that early.

100+ system wins for Alderlake-h, 200+ for Rembrandt.....

So Alderlake P is 350+ they said.