Discussion Intel current and future Lakes & Rapids thread

[mikk]

The UHD 750 seems to overclock great. There's the review with 1.7GHz, which is a 30% overclock.

Performance gains are surprisingly low over the UHD 630: https://hexus.net/tech/reviews/cpu/147440-intel-core-i9-11900k/?page=11

The 50% is really an up to figure. I wonder if negatives of the backport also apply here? Or if Gen 11 and Xe were just focused on power efficiency and die size which is relevant for mobile?

There is not enough data apart from 3dmark, we have only a few games. F1 2020 from hardwareluxx gained 53% and Strange Brigade 1080p 50%.

Review Intel Core i9 11900K y Core i5 11600K [LGA-1200]

Hoy se lanzan de forma definitiva los nuevos procesadores de Intel, pertenecientes a la 11va Generación, conocidos como Rocket Lake-S.

www.madboxpc.com

Benchmark: Integrierte Grafikeinheit - Seite 17 - Hardwareluxx

Die Rakete mit Fehlstart: Core i5-11600K, Core i7-11700K und Core i9-11900K im Test.

www.hardwareluxx.de

edit: the timespy scores look strange on hexus. If it's an overall score the UHD 750 score seems to be too low and if it's a graphics score the 10900k score too high compared to guru3d. I found an iGPU review with 4 games (mostly in the 50-60% range): https://www.hwupgrade.it/articoli/c...0k-e-i5-11600k-rocket-lake-in-test_index.html

From the limited data I wouldn't say 50% is an up figure, it looks more like the norm.

 

[DrMrLordX]

@dullard

CineBench R20 and R23 already support AVX512 via Embree, or at least they should. That should give you some idea of what AVX512 will bring to the table in a real-world application like a 3D rendering suite.

 

[Hulk]

The 11700k is the 5800x competitor. It'll do almost everything the 11900k will do. Overclocking is a wash. The UHD 750 isn't free either. The i9 is priced like that for no reason other than the fact that it's a segment filler and competes with the R9 12 core.

I don't think the 11900K is a legitimate competitor with the 5900X because Intel prices it that way. It offers a miniscule performance advantage over the 11700K and both are beaten by the 5800X.

As much as I was hoping Intel would come up with something great when moving Sunny Cove to the desktop with its faster memory subsystem it didn't happen. The Zen 3 core is demonstrably superior to Rocket Lake, and no amount of relative pricing vs AMD by Intel can convince me otherwise. Eventually there is only truth shown by the numbers.

 

[eek2121]

Didn't you read Gelsinger IDM2.0 announcement?

So whether 7nm comes or not Intel has leadership products coming from TSMC - for all market segments.

https://newsroom.intel.com/wp-content/uploads/sites/11/2021/03/CEO-Gelsinger-quotes-57285720.pdf

...and given how long it takes to design or port an existing design to a new node, we will likely have 7nm before you ever see an Intel CPU on TSMC unless Intel has been secretly working on it for a while. It seems to me like the TSMC talk is a “hedge our bets” type of deal. I guess we will see. GPUs on the other hand...

 

[naukkis]

...and given how long it takes to design or port an existing design to a new node, we will likely have 7nm before you ever see an Intel CPU on TSMC unless Intel has been secretly working on it for a while. It seems to me like the TSMC talk is a “hedge our bets” type of deal. I guess we will see. GPUs on the other hand...

You didn't read it? There it is, both Intel 7nm and additional leadership products from TSMC are coming 2023. If like to bet, my money on TSMC parts come before Intel own 7nm. So yes, Intel have been working cpu's on TSMC for a while. But they still except majority of sold chips to come from their own factories 2023 - for next years things obviously depend on how well Intel's own 7nm matures.

 

[Pneumothorax]

.....
That's called reputation, and one Intel is losing quickly. If they keep this up AMD will become #1 in revenue and marketshare.

Right now they are riding on the reputation they have built over the years. As it can be gained, it can also be lost.

Of course the transition doesn't happen instantaneously.

We wish....
As long as AMD is severely production constrained with producing dies for consoles and GPU's on the same process and competing with the likes of Apple and Qualcomm for fab capacity, it will never happen.
Tragically this is just like the Athlon 64 days.... before Core hit and will likely replay the exact same way again.

 

[dullard]

@dullard

CineBench R20 and R23 already support AVX512 via Embree, or at least they should. That should give you some idea of what AVX512 will bring to the table in a real-world application like a 3D rendering suite.

CineBench R23 does not use AVX-512:

https://twitter.com/x/status/1327358373373898752

CineBench R20 is 3 years old. I haven't checked it's Embree version, but AVX-512 performance problems have been fixed recently. Does CineBench R20 automatically get these updated Embree versions?

 

[jpiniero]

Rocket Lake doesn't appear to have any kind of automatic downclock. How much power it uses is another question. Pretty sure Tiger Lake is the same deal. We'll have to see about Ice Lake Server.

 

[Hulk]

We wish....
As long as AMD is severely production constrained with producing dies for consoles and GPU's on the same process and competing with the likes of Apple and Qualcomm for fab capacity, it will never happen.
Tragically this is just like the Athlon 64 days.... before Core hit and will likely replay the exact same way again.

This is true but I think AMD's current architecture and process lead over Intel is going to be longer lasting. I could be wrong but I just don't see Alder Lake being the breakthrough that Core was. Intel has made their first attempt to achieve architectural superiority, or even parity with Rocket Lake. Strike one. At equal clock speed Zen 3 has a comfortable lead over Cypress Cove. When you also consider AMD has parts with double the number of cores and double the efficiency you have to realize that it is quite the hole Intel has dug for itself.

The reason Intel is "fine" now is because AMD simply can't supply the processors that Intel can, and Intel's processors are "good enough."

Alder Lake has a lot riding on it's shoulders. If it can beat Zen 3 but is behind Zen 4 (and Zen 4 is released within a quarter or so of ADL) then Intel will remain in the "good enough" category and have to hope that AMD can't source enough parts to put a dent in their bottom line. But I would consider this outcome strike two.

We know that TMSC is putting a huge amount of money into fabs and one would assume that part of that projection is AMD being able to maintain an architectural lead on Intel. If this is the case and they can supply the parts and Intel can't regain architectural superiority then this would be strike 3 for Intel and they may go into a death spiral.

Yes I know, lots of "ifs" there.

So if you are in AMD/Intel's corner between rounds what do you tell them?

Intel's corner:
"You've gotta transition everything to 10ESF or at least 10SF as soon as possible. You're gettin' killed out there! You're burning up! It's taking you too much energy to get going, you're just too darn big. Also we've gotta get smarter, Alder Lake has got to KO Zen 3 and be able to stand up to Zen 4."

AMD's corner:
"Okay we've got him on the run. Don't take any big risks, just follow the fight plan. Keep Zen 4 on track and work with TMSC to secure the number of parts that we know we'll need. Oh yeah, and if we want the big numbers we've gotta get the APU production going."

 

[Hulk]

I'm also very interested to know about the following that Ian wrote about way back when Intel released RKL info.

Another feature on the GPU side, along with enhanced performance moving to Xe graphics, is AV1 decode support, as well as always-on Quick Sync Video. With Rocket Lake, Intel is enabling both the integrated graphics and the discrete graphics to supply compute simultaneously, allowing users to enable their discrete graphics for gaming and the integrated graphics for stream encoding. This will be managed through additional software, such as that found on dual graphics laptops, that enable pointing specific software to different devices.

 

[Panino Manino]

What about Global Foundries?
Wouldn't they make killer money if they hadn't decided to give up 7nm? Can't they revert this decision now that the marked is so constrained?

 

[Hitman928]

What about Global Foundries?
Wouldn't they make killer money if they hadn't decided to give up 7nm? Can't they revert this decision now that the marked is so constrained?

Most likely they could have made a decent profit had they not given up on 7nm (assuming they could have it at least functional with decent yields by now). But this crazy demand situation wasn't foreseen by anyone so at the time they decided to end development, it made sense because they knew they would just keep falling further and further behind which wouldn't allow them to even break even on the R&D costs of new nodes. This also assumes that they wouldn't hit a bottleneck in other supplies as well.

For instance, if there is a crucial lack of supplies needed for assembling the fabricated die, then GF wouldn't be in a good situation anyway because the bottleneck would be post fabrication and common to all foundries, so people would just pick the best process to go with for whatever amount of chips they could get assembled. Anyway, it's too late now. By the time they actually got 7 nm up and running, they'd be at least 2 (maybe more) nodes behind and interest in using their node would no longer exist, not in large numbers anyway.

 

[scannall]

What about Global Foundries?
Wouldn't they make killer money if they hadn't decided to give up 7nm? Can't they revert this decision now that the marked is so constrained?

You have to be able to buy the equipment to do it. And ASML has sold everything they can produce for the next couple of years. By the time they would actually get equipment they would be behind again.

 

[Hulk]

You know me. I can't resist working the numbers a bit. Everything held to 3 significant digits, which is what the data used had.

Quick and dirty "throughput" comparison of some recent cores using SPEC2006 1T from Anandtech Bench results. Equalized for clock speed and average value from 3 or so tests from each core. I used SPEC2006 because it seems to correlate decently to real-world performance. While we may argue the percentages, I don't think anyone would dispute the order of "throughput" for the following cores as being:

Zen 3>Cypress Cove>Skylake>Zen 2>Haswell>Zen+

You've gotta give SPEC2006 credit for at least getting the order right in terms of what we see in terms of real-world performance.

Rocket Lake (Cypress Cove) halves the distance of Skylake to Zen 3. Zen 3 30% faster than Skylake and 15% faster than Rocket Lake.
Alder Lake has some ground to make up just to catch Zen 3 let alone Zen 4. If Golden Cove is 20% better than Cypress Cove then Intel could relax frequency by 5% and equal Zen 3 performance-wise. But then there's Zen 4 coming...

The interesting thing to me here is that while the 5800X isn't that much faster than the 11900K in real world performance but is significantly better in "throughput," much of that is because the 5800X is giving up quite a bit in terms of clock speed to the 11900K. The more time goes by the more I am impressed by the Zen 3 architecture. The 5800X is giving up 500MHz to the 11900K and still performing better in SPEC2006.

Zen 3 is so good I'm wondering if AMD is going to significantly change the internals for Zen 4 or just make some tweaks that would improve IPC a bit and allow for higher clocks? The flowcharts on these designs tell much of the story, Zen 3 is just so wide and it must have really good algorithms to keep instructions moving along.

Okay, feel free to tear into me!

	MHz/SPEC2006
Zen 3	63.6	14.9%​	Zen 3 Faster than Rocket Lake
Rocket Lake	73.1	13.4%​	Rocket Lake Faster than Skylake
Skylake	82.9	3.1%​	Skylake Faster than Zen 2
Zen 2	85.5	13.5%​	Zen 2 Faster than Haswell
Haswell	97.0	10.1%​	Haswell Faster than Zen+
Zen+	106.8
	MHz/SPEC2006
Zen 3	63.6
Rocket Lake	73.1	14.9%​	Zen 3 faster than Rocket Lake
Skylake	82.9	30.3%​	Zen 3 faster than Skylake
Zen 2	85.5	34.4%​	Zen 3 faster than Zen 2
Haswell	97.0	52.5%​	Zen 3 faster than Haswell
Zen+	106.8	67.9%​	Zen 3 faster than Zen+

 

[ondma]

That is a nice table, but the Zen 2 to Zen 3 gain is certainly an outlier compared to other benchmarks. Here is an article by Techspot comparing Zen 3 to Zen 2 and Skylake at 4ghz. Productivity gains range from 23% to 4.7% (Zen 3 vs Zen 2). Gaming "IPC" gain averaged 22% vs Zen 2 and 7% vs Skylake. link

If the IPC gain from CL to RL had translated correspondingly to gaming, Intel would have been equal or slightly ahead. Unfortunately for Intel it didnt.

 

[repoman27]

I noticed that Intel made the Rocket Lake datasheets publicly available. The Processor DDR Memory Speed Support table has no less than 10 footnotes. They essentially boil down to the following caveats as far as what is officially supported by Intel:

• 400 series chipset: DDR4-2933 Gear 1 maximum
• 4-layer motherboard: DDR4-2933 Gear 1 maximum
• more than 2 ranks per channel: DDR4-2933 Gear 1 maximum
• i9-11900K and i9-11900KF only: DDR4-3200 Gear 1 maximum
• all other Rocket Lake SKUs: DDR4-3200 Gear 2 maximum

So there's binning based on memory controller for the i9-11900K[F], and the relaxed platform segmentation based on memory overclocking features is probably a response to the otherwise weak memory controller.

 

[samboy]

From a technical perspective; I think the Intel design team has done a great job with what they had to work with (accepting that 10nm or 7nm wasn't going to happen until later). They have taken a 2015 design node (14nm) and come out with a processor that is right at the top with single threaded performance and with decent multi-threaded performance for 8 cores.

They obviously had to make compromises; 8 core tops; higher power usage. In my mind, this was always going to be the case with 14nm fabrication. I personally think that for typical desktop usage the reviewers have been overly critical; the average user isn't bench marking 24/7 and the processor idles back in most common usage scenarios (i.e. the power consumption at full load is overplayed). When you do need the performance, then its there..... just not as efficient as AMD. My only concern would be reliability as I recall the instability with the Pentium III when Intel pushed the clocks too high to compete with AMD at the time (around the 1GHZ level)......

I'm really not sure what everyone was expecting; but if Intel can provide volume then this is a reasonable choice for many folks. All things being equal, AMD is likely a better choice but I don't think they are anywhere close to being able to provide the volume of CPU's that Intel can to major OEM's etc. Intel has done pretty well with their 14nm node in my opinion.

Full disclosure: My current system is an AMD 5950x

 

[Hulk]

That is a nice table, but the Zen 2 to Zen 3 gain is certainly an outlier compared to other benchmarks. Here is an article by Techspot comparing Zen 3 to Zen 2 and Skylake at 4ghz. Productivity gains range from 23% to 4.7% (Zen 3 vs Zen 2). Gaming "IPC" gain averaged 22% vs Zen 2 and 7% vs Skylake. link

If the IPC gain from CL to RL had translated correspondingly to gaming, Intel would have been equal or slightly ahead. Unfortunately for Intel it didnt.

Ahh! Don't do that to me! I can't stop working numbers!
Seriously though, that's great data. Thanks.

Skylake down 17% vs Zen 3 for apps, 6% down for games, and with 11 apps/8 games weighting, down 12% to Zen 3. You are right, much more realistic than the 30% number using SPEC2006.

Using relative SPEC2006 numbers comparing Rocket to Skylake then this 17% deficit for Skylake to Zen 3 would be more like half of that for Rocket Lake or about 8.5%, most of which is made up in Rocket vs. Zen 3 clocks, which makes sense as we're seeing 5800X just a bit better than 11900K overall with applications.

However you slice and dice it Zen 3 has about 3% (games) to 8.5% (apps) "throughput" advantage over Cypress Cove (Rocket Lake). The clock speed advantage of Rocket Lake can't make up for that deficit in applications (generally) but it can probably in some games.

This also makes sense in that with the slightly lower IPC and clocks of Zen 2 vs Zen 3, Skylake was just able to stay ahead of Zen 2.

Techspot - Ryzen 5000 IPC	5800X	3800X	10700K	2700X	1800X
Cinebench R20 MT	5333	4789	4192	4051.0	3944
Cinebench R20 ST	517	465	417	411.0	404
7-Zip File Manager (comp)	83095	70514	66865	54819.0	52732
7-Zip File Manager decomp)	99695	89535	81282	79539.0	73694
Blender Open Data	935	979	1081	1103.0	1134
	0.00107	0.001021	0.000925	0.000907	0.000882
V-Ray Benchmark	15592	13445	12395	11854.0	11498
Corona 1.3	94	108	110	120.0	126
	0.010638	0.009259	0.009091	0.008333	0.007937
DaVinci Resolve Studio 16	1371	1283	1225	1211.0	1175
Adobe Premiere	839	789	738	735.0	727
Adobe Photoshop 2020	1051	867	894	781.0	766
Adobe After Effects 2020	1177	1013	959	912.0	881
Geomean Total	334	297	278	263	254
Compared to Zen 3		-11.1%	-16.7%	-21.3%	31.3%
Compared to Zen 2			-6.4%	-11.5%	-14.4%
Compared to Skylake				-5.5%	-8.5%
Compared to Zen+					-3.2%
Techspot - Ryzen 5000 IPC
Games 1% Min fps	5800X	10700K	3800X	2700X	1800X
Far Cry New Dawn (DX11)	112	114	97	88.0	86
Tom Clancy's Rainbow Six Seige	406	371	336	285.0	278
Watch Dogs: Legion	68	64	56	50.0	49
F1 2020 (DX12)	222	205	180	152.0	150
Horizon Zero Dawn	106	100	94	83.0	81
Death Stranding	162	146	131	121.0	119
Shadow of Tomb Raider	125	120	106	96.0	92
Hitman 2	120	112	98	91.0	89
Geomean Total	143	135	120	107	104
Compared to Zen 3		-5.9%	-16.4%	-25.5%	-27.3%
Compared to Skylake			-11.2%	-20.9%	-22.7%
Compared to Zen 2				-10.9%	-12.9%
Compared to Zen+					-2.3%
Techspot - Ryzen 5000 IPC	5800X	10700K	3800X	2700X	1800X
App+Game Geomean Total	234	205	203	180	175
Compared to Zen 3		-12.3%	-13.4%	-23.1%	91.2%
Compared to Zen 2			-1.2%	-12.3%	-14.8%
Compared to Skylake				-11.3%	-13.8%
Compared to Zen+					-2.8%

 

[DrMrLordX]

CineBench R23 does not use AVX-512

Hah! Seriously? They're that worried about clockspeed reduction due to AVX512 mode? Bizarre. Well if R23 doesn't support it then I'm guessing R20 doesn't either.

CineBench R20 is 3 years old. I haven't checked it's Embree version, but AVX-512 performance problems have been fixed recently. Does CineBench R20 automatically get these updated Embree versions?

Not sure if Maxon updates their old Cinebench versions but I'm guessing not.

 

[Panino Manino]

Zen 3 is so good I'm wondering if AMD is going to significantly change the internals for Zen 4 or just make some tweaks that would improve IPC a bit and allow for higher clocks? The flowcharts on these designs tell much of the story, Zen 3 is just so wide and it must have really good algorithms to keep instructions moving along.

A Zen 3 refresh, die shrink, if it can give it more clocks would be enough to stay competitive with Intel now, but about Zen 4 I believe it'll bring more, because Zen 3 don't have that much of changes from Zen 2, isn't that right? Is much of the same but "made again, this time better". Most of the gains come from better and faster cache and memory, right?
I know that Zen 4 should bring much more than a new CPU arch and this alone would justify a new generation and purchase, but the promise and to continue with the aggressive performance increases and this can only be achieve change the core, I suppose.

But... does this matter? The market is changing, isn't there plenty of other things more important now? Like, if AMD will add AVX-512 or something equivalent, if it'll at SMT4 for the server market, how will it integrate Xilinx and so on?

 

[Thala]

Hah! Seriously? They're that worried about clockspeed reduction due to AVX512 mode? Bizarre. Well if R23 doesn't support it then I'm guessing R20 doesn't either.
Not sure if Maxon updates their old Cinebench versions but I'm guessing not.

AVX512 is not that useful in Embree even if enabled as it is directly bound to vuint<16> and vfloat<16> class. However this class has limited use within Embree in general. I guess the most used datatype in Embree is vint<4> and vfloat<4>. However these datatypes are directly coupled with SSE (and optionally NEON) registers and instructions.

 

[Hulk]

A Zen 3 refresh, die shrink, if it can give it more clocks would be enough to stay competitive with Intel now, but about Zen 4 I believe it'll bring more, because Zen 3 don't have that much of changes from Zen 2, isn't that right? Is much of the same but "made again, this time better". Most of the gains come from better and faster cache and memory, right?
I know that Zen 4 should bring much more than a new CPU arch and this alone would justify a new generation and purchase, but the promise and to continue with the aggressive performance increases and this can only be achieve change the core, I suppose.

But... does this matter? The market is changing, isn't there plenty of other things more important now? Like, if AMD will add AVX-512 or something equivalent, if it'll at SMT4 for the server market, how will it integrate Xilinx and so on?

I think Zen 2 to 3 architectural changes were pretty significant. The OP-cache has significant changes as well as the much wider execution engine. Overall it seems a very "balanced" design as supported by evidence of greatly increased performance over Zen 2.

Yes, I think the architecture does still matter quite a bit but that's just my opinion of course! We're still "looking" for applications where AVX-512 really matters. I've found a lot of these ISA changes promise huge performance increases that never materialize in actual day-to-day use. As for SMT4, yeah I don't get too excited about servers as long as my web pages load quickly I'm good

 

[IntelUser2000]

Using relative SPEC2006 numbers comparing Rocket to Skylake then this 17% deficit for Skylake to Zen 3 would be more like half of that for Rocket Lake or about 8.5%, most of which is made up in Rocket vs. Zen 3 clocks, which makes sense as we're seeing 5800X just a bit better than 11900K overall with applications.

Hulk, you need to differentiate between SpecInt and SpecFP. The latter is useless for client as it represents HPC workloads and they are extremely memory bandwidth bound.

With the microcode updated 11700K, it's getting 6.98 in SpecCPU2017_Int versus 6.04 for 10700K. That's 5GHz ST Turbo for 11700K vs 5.1GHz Turbo for 10700K. Assuming 85% scaling and not erroneously assuming 100% scaling like Ian does in his reviews,* it'll put 11700K at 7.09.

That means Rocketlake is 17.5% faster per clock compared to Cometlake. Then normalizing for 5.1GHz clocks again, the 5950X gets 7.715, or 8.7% faster than Rocketlake.

*Again if you are not being like AT is being in their reviews, the 3% "deficit" that Tigerlake has compared to Icelake disappears entirely. Higher clocked parts perform less per clock in real-world scenarios.

 

[Hulk]

Hulk, you need to differentiate between SpecInt and SpecFP. The latter is useless for client as it represents HPC workloads and they are extremely memory bandwidth bound.

With the microcode updated 11700K, it's getting 6.98 in SpecCPU2017_Int versus 6.04 for 10700K. That's 5GHz ST Turbo for 11700K vs 5.1GHz Turbo for 10700K. Assuming 85% scaling and not erroneously assuming 100% scaling like Ian does in his reviews,* it'll put 11700K at 7.09.

That means Rocketlake is 17.5% faster per clock compared to Cometlake. Then normalizing for 5.1GHz clocks again, the 5950X gets 7.715, or 8.7% faster than Rocketlake.

*Again if you are not being like AT is being in their reviews, the 3% "deficit" that Tigerlake has compared to Icelake disappears entirely. Higher clocked parts perform less per clock in real-world scenarios.

The 8.5% I came up with is pretty close to the 8.7% you calculated for Zen 3 over Rocket Lake right?

The Anandtech benchmark data for Rocket Lake doesn't differentiate FP vs INT performance as far as I could tell like it does normally for some reason?

I avoided any mobile benchmark data for two reasons. First, mobile memory subsystems as I've been told here are not valid comparison points against desktop. Second, with thermal throttling going on with mobile designs it's really impossible to even estimate clock speeds.

Of course you are correct... when extrapolating clocks higher, results are higher than reality due to less than 100% scaling. Never thought about that with the Ice Lake vs. Tiger Lake comparison. Tiger is 1GHz higher so is negatively affected by extrapolation when comparing IPC. ie Tiger should have a "built in" decrease IPC (as do all designs) as frequency increases and this shouldn't be part of an IPC core vs. core comparison.

Very good point. The specifics are very important in these discussions.

 

[IntelUser2000]

The 8.5% I came up with is pretty close to the 8.7% you calculated for Zen 3 over Rocket Lake right?

Yea. That's a margin of error difference.

The Anandtech benchmark data for Rocket Lake doesn't differentiate FP vs INT performance as far as I could tell like it does normally for some reason?

We're arguing about 1-2% differences. The specifics are very important.

Taking into account, SpecInt is actually quite representative of differences in client.