Discussion Intel current and future Lakes & Rapids thread

Page 437 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,682
136
So it looks like Rocket Lake's inter-core latency issues aren't just the result of the move to a bigger node. Tiger Lake H is doing better than the 14nm port, but it's still behind Comet Lake.

tgl-latencies.png
 
  • Like
Reactions: Tlh97 and Elfear

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
It is a "normal boost". Just because AMD boosts differently to Intel, doesn't mean it's not normal. Intel laptops just have a very different approach to boosting. While AMD slowly try and max out thermal saturation of a chassis over a 5 minute period before mating thermal dissipation of the device afterwards, Tiger Lake-H's base specification of 107W PL2 and 45W PL1 indicates that in that 56s PL2 they hit thermal saturation and then match thermal dissipation afterwards with PL1.
"Normal" may not be the word, but max boost for the 5980HS can go up to 90w!

1621291144061.png



At the start of the test we see the power peak at 62 W for a second or two, before coming down to a 300 second turbo of around 42 W. During this time the frequency comes down to around 3650 MHz, before reducing to 3400 MHz as the turbo budget runs out and the system drops to 35 W for a sustained mode. The second part of the test, which is multi-threaded, starts at around 500 seconds and showcases a sustained temperature of around 81ºC and an opportunistic frequency peaking at 4500 MHz, all while staying at 35 W. Towards the end of the test, as we start looking at single threaded workloads, the system peaks above 60 W on two occasions, but also shows 4500 MHz for ST activity. The CPU temperature rises more to 90ºC as the preferred core is used, but as it is only a single core the thermals are better managed and adjust accordingly.
 

Attachments

  • 1621291227471.png
    1621291227471.png
    85.7 KB · Views: 18
  • 1621291424658.png
    1621291424658.png
    124.8 KB · Views: 15
  • Haha
Reactions: maddie

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,572
146
"Normal" may not be the word, but max boost for the 5980HS can go up to 90w!

View attachment 44493
View attachment 44495
Um, that's the frequency line, not the power consumption line.

Ironically, the 11980HK - despite all the throttling - is the chip that actually boosts that high with a peak of 86W. Power is in blue and they actually noted peak power at 56W for the 5980HS on the chart.
c9d0f113539adf74f50fa5315d57b868.jpg
 

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
Um, that's the frequency line, not the power consumption line.

Ironically, the 11980HK - despite all the throttling - is the chip that actually boosts that high with a peak of 86W. Power is in blue and they actually noted peak power at 56W for the 5980HS on the chart.
c9d0f113539adf74f50fa5315d57b868.jpg
You're right, but look at those frequencies compared to Tiger Lake H. Also, shouldn't the comparison be Mobile vs Mobile for fairness?
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,572
146
You're right, but look at those frequencies compared to Tiger Lake H. Also, shouldn't the comparison be Mobile vs Mobile for fairness?
Yes, frequencies at the same power are higher. We've already established this.

A theoretical Tiger Lake-S would use the exact same die as Tiger Lake-H. I see no reason why the 5950X would not be a suitable comparison provided it is given JEDEC spec memory to equalise platforms as much as possible. As it just so happens, Anandtech test their desktop platforms using JEDEC spec memory.

If you want to try and compare Willow Cove to Zen 3 in IPC, I don't see why we shouldn't use the full Zen 3 core as a comparison point just as much as the castrated version.
 
  • Like
Reactions: Tlh97

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
Yes, frequencies at the same power are higher. We've already established this.

A theoretical Tiger Lake-S would use the exact same die as Tiger Lake-H. I see no reason why the 5950X would not be a suitable comparison provided it is given JEDEC spec memory to equalise platforms as much as possible. As it just so happens, Anandtech test their desktop platforms using JEDEC spec memory.

If you want to try and compare Willow Cove to Zen 3 in IPC, I don't see why we shouldn't use the full Zen 3 core as a comparison point just as much as the castrated version.
So you don't think the thermal/power limits and relatively short boost period all could have an impact on the results?
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,570
136
Regarding Zen 3 vs Tiger Lake power draw comparisons based on the new Tiger Lake H review.

Looks like at 65W given adequate cooling Tiger Lake H on 10SF could sustain 3800MHz at 65Watts.

Looking back at the Zen 3 5950X review we see 120 Watts at 3775MHz. So basically 120W vs. 130W for TMSC Zen 3 7nm vs TGL 10SF. Pretty darn close. I believe both of these measurements are package power. TGL also includes iGPU so we could call this even.

This is quite unfair to Zen 3 as the MCM I/O die is a relative absolute power hog. 65W Desktop Cezanne would be a better comparison, once it's released.

You're right, but look at those frequencies compared to Tiger Lake H. Also, shouldn't the comparison be Mobile vs Mobile for fairness?

BTW there is no contradiction in what Andrei says

1. He lists the higher clock speed (sustained near 4 ghz range) as a bonus for everyday tasks which it absolutely is.
2. He can still very well mention that this gives it no edge in their benchmarks as they are automated and run for hours with no time "off the gas pedal" .

The latter is just a fact, as SPEC alone demonstrates
 
  • Like
Reactions: Tlh97 and scineram

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,572
146
So you don't think the thermal/power limits and relatively short boost period all could have an impact on the results?
As @coercitiv mentionned above, not in SPEC as it's a 6 hour long test. a 5 minute +7W increase to power limits provides essentially nothing in terms of performance. And it also shouldn't have an effect on any single thread tasks either, as the 5980HS doesn't require >35W on a single core for maximum single-threaded performance. Not even close to that even.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Come on, Rocket Lake does not perform "horribly". If tuned properly, it comes very close to Zen 3 in gaming and productivity. If RL performs "horribly", they Zen 3 is only some percent above "horrible". I will grant that performance per watt is very poor, and Zen 3 is clearly a better processor, but RL is a very powerful chip.

I don't know why people ignore the multi-threaded benchmarks as if it's nonexistent. Even the 5900X is often 30% faster in such scenarios. That meets every definition of horrible.

If those tests don't matter to you then, you won't have in mind the 5900X or 11900K, you will get something like the 5600X.

And Zen 3 isn't any slower in those low threaded scenarios either. In some cases like gaming it's just as fast/faster than Rocketlake.

Not just AMD. How is it better than Cometlake?

Tigerlake-H at least keeps up with Zen 3 mobile, and laptops have wildly varying setups and configurations and optimizations that can skew to it's favor.

@uzzi38 I don't think it's just caches. The mobile chips perform less per clock than desktops. Plus most Zen mobile configurations are very conservative in it's boost.

I honestly don't understand how much more performance Intel is supposed to squeeze from an aged process on a back-ported design.

Are we really excusing Intel for screwing up transition to future processes? Because it doesn't matter they've achieved being able to backport - it shouldn't have happened.

So it looks like Rocket Lake's inter-core latency issues aren't just the result of the move to a bigger node. Tiger Lake H is doing better than the 14nm port, but it's still behind Comet Lake.

View attachment 44492

Did we completely rule out the possibility that it's not doing some weird power adjustments that affect latencies? How does it compare to Tigerlake-U? What about Cometlake-H?

The slower L3 caches will likely affect C2C latencies somewhat.
 
Last edited:
  • Like
Reactions: Tlh97

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
It is worth noting that AT tested a reference platform that Intel could not even be bothered to limit to spec. The AT review was also rushed. Both Intel AND AMD need to stop dropping products for review at the last minute.

I have both a 5950X and an 11800H system coming soon. Please feel free to reach out to me if there are any benchmarks you would like to run on either. I will try and post it here when both chips come.
 

Exist50

Platinum Member
Aug 18, 2016
2,445
3,043
136
That's an extension of the idea that Intel designs aren't somehow locked to a specific node anymore, starting with Sunny Cove. Which, after Rocket Lake, I'm beginning to think is more smoke and mirrors than truth.

A post-Redwood Cove reality only.
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,682
136
Did we completely rule out the possibility that it's not doing some weird power adjustments that affect latencies?
No, we didn't rule out anything just yet. However, this review is quite impressive in the amount of detail it is able to provide, and we get information about core latencies for TGL-U and Cezanne as well.
  • TGL-U has worse overall latencies over TGL-H. Since the ring-bus is smaller, the only explanation is better uncore speeds, either to save power or simply due to optimization and better uncore clocks on TGL-H.
  • Cezanne however has very close numbers to Vermeer. This shows that power adjustments shouldn't play that much of a role when talking about 35-45W TDP chips.
 
  • Like
Reactions: Tlh97

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Or, it could simply mean AMD isn't adjusting interconnect speeds while Intel is. We know ring bus speeds are adjusted by Intel nearly every generation.

What the hell is Intel planning with Raptor Lake if it's still on 10nm? Better hybrid changes for improved performance? What does that mean?

It's not like they have much space to grow so they can't add many more cores. Alderlake with GT2 graphics is easily going to get to 250mm2 die size or more. I can't see the cores changing much either.
 

uzzi38

Platinum Member
Oct 16, 2019
2,565
5,572
146
Or, it could simply mean AMD isn't adjusting interconnect speeds while Intel is. We know ring bus speeds are adjusted by Intel nearly every generation.

What the hell is Intel planning with Raptor Lake if it's still on 10nm? Better hybrid changes for improved performance? What does that mean?

It's not like they have much space to grow so they can't add many more cores. Alderlake with GT2 graphics is easily going to get to 250mm2 die size or more. I can't see the cores changing much either.

For RNR/CZN-H they aren't. FCLK runs at a static 1600MHz clock (assuming DDR4-3200 memory). I believe CZN-U acts the same as RNR-U and now TGL-H where FCLK/ring bus is dynamic depending on the workload.

I feel like Raptor Lake is an improvement to the Atom uArch alongside a Willow Cove style core upgrade (AKA just a cache improvement). I feel like they'd want to replace the Atom cores with a new uArch that supports AVX-512 to not have to worry about mismatched ISAs any more.

Agreed on the last bit. Raptor Lake is going to be physically huge from how large each Golden Cove tile looked in those SPR die photos.
 
  • Like
Reactions: Tlh97 and scineram

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I feel like Raptor Lake is an improvement to the Atom uArch alongside a Willow Cove style core upgrade (AKA just a cache improvement).

Not sure about this. They put "next" mont as being 2 years after Gracemont. Unless they are *just* adding 2-cycle AVX-512 as with client Core cores.
 

Hulk

Diamond Member
Oct 9, 1999
4,191
1,975
136
I don't think it's interesting at all, the die area allocated for 8+8 is equivalent to 10 GC cores. With a 20% IPC advantage and similar clocks the 10 GC chip would obviously match 5900X. The 8+8 not matching 5900X would be a big problem for the first gen desktop hybrid.

Maybe it's time people start comparing 8+8 versus 10+0 of the same kind of cores, and see where exactly you think there's gains to be had for hybrids, because so far all I see is power consumption, which is pure irony considering Intel's attitude in the last few years (squeeze performance no matter the power cost).

Well if we actually had silicon to compare these opinions/discussions would be moot. And it was interesting enough for you to take the time to respond;)
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,570
136
Interesting, so Overclocked DDR5 is coming at pretty decent latencies:


The GeIL DDR5 memory specification starts at 4800MHz with the sub-timing latencies of CL40-40-40 at 1.1 volts. The overclocking products are also under development, including 6000MHz CL32-36-36, 6400MHz CL32-36-36, 6800MHz CL36-44-44, and 7200MHz CL36-44-44, and will be available with non-RGB product versions at the same time. ]
The true latency for 6400 Mhz CL32 and 7200 Mhz CL36 kits is 10ns. This is exactly the same as DDR4 3600 CL18 or 3200 Mhz CL16 while providing double the bandwidth (actually more due to DDR5's better design).

That's much better than I expected ever since reading this Anandtech's take on DDR5.
Actually that's very similar to what was initially availably for DDR4 (here is a "blast from the past" Anandtech's first DDR4 scaling article from February 2015)

Another interesting dibit is:
It has been seven years since DDR4 launched into the market, and GeIL has put countless hours into developing the new DDR5 memory modules. And in doing so, GeIL has designed the Polaris RGB to provide RGB illuminated high-performance DDR5 gaming memory and has been working closely with motherboard makers to guarantee the best compatibility and reliability among the latest Intel and AMD motherboards.
So AMD does have motherboards with DDR5 support out there for testing with the memory makers. I seriously doubt these are for products released in 2022 H2 (e.g. Zen 4). This to me seems like a hint for Zen 3+.
 
Last edited:

2blzd

Senior member
May 16, 2016
318
41
91
I just orderd 64gb of DDR4-4000 because I was anticipating DDR5 ramp to be slow out the gate...now I'm not so sure..@_@
 

2blzd

Senior member
May 16, 2016
318
41
91
It’s going to be a whacky year in terms of supply. It’ll still be really valuable even if DDR5 is miraculously easily available in all speeds and capacities from the jump. I’d stay the course.

Thats what I was thinking too..but there seems to be a steady stream of news from different companies talking about DDR5..and we're 5-6 months away from the first DDR5 platform, so thats promising imho for DDR5..Im sure there will still be a steep price premium, especially with these ram manufactures seeing all the AIB partners making a killing by jacking up videocard prices, they're gonna get their opportunity w/ DDR5.
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,570
136
There is also the possibility that most initial modules will be JEDEC only and the XMP versions will come during later months