Discussion Intel current and future Lakes & Rapids thread

[LightningZ71]

All of this is still reinforcing my belief that the 12400 is going to be a problem for the 5600x, the 12600kf is going to be a problem for the 5800x, and the 12700kf and 12900kf will put pricing pressure on the 5900 and 5950x.

And, lest we forget, this round, Intel has a platform advantage with PCIe 5.0 on the first slot and x8 PCIe 4.0 linking the chipset. I think that the DDR4 boards are going to land in x570 price range. Unless AMD hits hard with Zen3d/xt, I don't see the attraction.

 

[andermans]

And, lest we forget, this round, Intel has a platform advantage with PCIe 5.0 on the first slot and x8 PCIe 4.0 linking the chipset. I think that the DDR4 boards are going to land in x570 price range. Unless AMD hits hard with Zen3d/xt, I don't see the attraction.

I was curious about the motherboard thing since it seems easy to have outrage about the expensive boards but those really seem like enthusiast boards. Looking at some of the lower end boards (200-350 USD was the cheaper range of what I could find) in a couple of local webshops, and it seems like the DDR4 boards are about 50 USD more expensive than the same named X570 boards, and the DDR5 boards are about 100 USD more expensive than the same named X570 board. (Though in the DDR5 case one would also need to consider the memory pricing)

So it seems they're still more expensive but it doesn't sound like it is earth shattering.

 

[Hougy]

Thanks everyone.

Looks like Golden Cove is smaller than expected. That's good. ~30% increase in core area for 19% perf/clock increase and an additional pipeline is a very good result.

Gracemont is 2x the size of Tremont. Larger than expected, but little better performing than expected. Slightly on the disappointing side. Perhaps it's larger because they wanted to reach 4GHz?

Could you please give the source for that 30% number?

 

[IntelUser2000]

I know this is very rough maths and the i9 is probably power starved, the difference between the i5, i7 and i9 shows adding each extra E core gives roughly 40% of a P core's performance when the core's fully loaded.
I know they were expected to be around 60% of the performance but I wonder if when fully loaded they drop further behind, a limitation of the shared l2 and ring stop perhaps?

Doesn't the 11900K get 5700 points at 250W?

The improvement is 50% once you add in 19% gain for the Cove cores, and the clocks haven't changed. The E cores are still in the 50% range.

Golden Cove's 30% area gain for 19% improvement despite an extra pipeline stage is a very good step in the right direction, considering adding pipeline stages cost performance, power use and area and we got same ballpark 19% range with Sunny Cove but a 38% increase in area.

@mikk

AVX2 shouldn't take a lot, even if Gracemont has both dual 256-bit units and FMA support. The diagrams seem to point towards a 128-bit FMA configuration similar to earlier AMD chips. I think the clock focus to get it from 3 to 4GHz is costing them.

I calculated 0.25mm2 for AVX256 units. Really, the addition shouldn't have been more than 0.1mm2. Even then we're talking 1.6mm2, still in the 90 percentile increase in core area.

I am hoping Intel eventually gets into the Tablet business, but naturally, organically this time by making an efficient chip. Gracemont's size seems to push away from that idea.

Going from a 0.88mm2 Tremont, to 1.7mm2 Gracemont, even with 2-4x vector unit performance and 30% uarch improvement doesn't sit right with me. Hopefully it's really good in the perf/watt department at least. And I mean "really good" by making it worth it in the mobile space despite the scheduler headaches. If say the 15W 2+8 is 25% faster than 28W Tigerlake in MT, and 4+8 28W is twice as fast, then probably.

 

[JoeRambo]

AVX2 shouldn't take a lot, even if Gracemont has both dual 256-bit units and FMA support. The diagrams seem to point towards a 128-bit FMA configuration similar to earlier AMD chips.

They seem to have dual 256 bit pipes, both with separate FMUL/FADD units. What seems to have happened in this generation overall is Intel realizing that very few if any applications are using FMA and deprioritizing it.
That's why they have added FADDs on two ports in P core - to be able to do fast FP vec/scalar ADD. And that is why E core has separate FADD/FMUL instead of beastly FMA.
The (mostly imaginary) cost for E cores is that FMA instructions will have several cycles slower latency, but benefits are used each day - faster FMUL and faster FADD.

It is nasty tradeoff, even if it makes little sense ( FMA unit is obviuosly capable of both ADD and MUL ops ). The most obviuos of it happening in the past was Broadwell having low latency ADD/MUL (3 cycles) and Skylake dropping it in favour of executing them on FMA units and that took 4 cycles ( full FMA op was also 4 cycles latency ).

So E cores not having "real" FMA units is great news for everyone and so is Intel adding low latency FADD to P cores. ( and the guys who still care about FMA on CPU need to get thorough beating with first gen Tesla card ).

 

[Timorous]

All of this is still reinforcing my belief that the 12400 is going to be a problem for the 5600x, the 12600kf is going to be a problem for the 5800x, and the 12700kf and 12900kf will put pricing pressure on the 5900 and 5950x.

And, lest we forget, this round, Intel has a platform advantage with PCIe 5.0 on the first slot and x8 PCIe 4.0 linking the chipset. I think that the DDR4 boards are going to land in x570 price range. Unless AMD hits hard with Zen3d/xt, I don't see the attraction.

I am not so sure. The main market for the 5600X is gaming and I don't see the 12600K having that large an advantage over it if and I expect we will only get DDR5 results on launch day. Platform vs platform for a new build is going to favour the 5600X for gaming from what I can tell.

The 5800X might need a small drop but again the main audience is gaming and platform vs platform the 5800X may remain competitive.

The 5900X and 5950X might need more movement if ADL sweep the productivity benches but it will depend on workload.

 

[TESKATLIPOKA]

Twitter

35% higher score after increasing TDP to 241W.

This graph shows one problem with Alder Lake and that is clockspeed at lower TDP.
If we assume that at TDP:241W the clocks are P: 4.8GHz and E: 3.8GHz, then It would mean at 125W It is only P 3.6GHz and E: 2.8GHz based on the difference in scores. That is pretty low for desktop.
Then what about 65W? Clocks must be even lower.
Mobile Alder will be even more limited, although they have less P cores.

 

[mikk]

This graph shows one problem with Alder Lake and that is clockspeed at lower TDP.
If we assume that at TDP:241W the clocks are P: 4.8GHz and E: 3.8GHz, then It would mean at 125W It is only P 3.6GHz and E: 2.8GHz based on the difference in scores. That is pretty low for desktop.
Then what about 65W? Clocks must be even lower.
Mobile Alder will be even more limited, although they have less P cores.

i5-12400 6C 4.0 Ghz used 78W in AIDA64 FPU, it's not the worst stress test but it uses more power than Cinebench on Intel CPUs. i5-12400 at 4.0 Ghz is still faster than a Ryzen 5600x in Cinebench R20 which usually clocks at ~4.5 Ghz. This is a huge efficiency increase over Tigerlake-H, I think it looks really promising for mobile.

 

[AtenRa]

i5-12400 6C 4.0 Ghz used 78W in AIDA64 FPU, it's not the worst stress test but it uses more power than Cinebench on Intel CPUs. i5-12400 at 4.0 Ghz is still faster than a Ryzen 5600x in Cinebench R20 which usually clocks at ~4.5 Ghz. This is a huge efficiency increase over Tigerlake-H, I think it looks really promising for mobile.

Where did you see 12400 CB20 scores and power consumption ??

 

[TESKATLIPOKA]

In what you quoted is a link. Try to click on It.

Edit: You are welcome.

 

[AtenRa]

In what you quoted is a link. Try to click on It.

axaxaxaxa completely missed that
thx

 

[AtenRa]

i5-12400 6C 4.0 Ghz used 78W in AIDA64 FPU, it's not the worst stress test but it uses more power than Cinebench on Intel CPUs. i5-12400 at 4.0 Ghz is still faster than a Ryzen 5600x in Cinebench R20 which usually clocks at ~4.5 Ghz. This is a huge efficiency increase over Tigerlake-H, I think it looks really promising for mobile.

If those scores are indeed with that power then its extremely positive and a good competition both in desktop and mobile.
Only a few more days for the reviews, CPU landscape seems to heat up and thats good.

 

[Abwx]

i5-12400 6C 4.0 Ghz used 78W in AIDA64 FPU, it's not the worst stress test but it uses more power than Cinebench on Intel CPUs. i5-12400 at 4.0 Ghz is still faster than a Ryzen 5600x in Cinebench R20 which usually clocks at ~4.5 Ghz. This is a huge efficiency increase over Tigerlake-H, I think it looks really promising for mobile.

But if AIDA use 10% more power than CB, as you suggest, it means that the CPU can clock 5% higher with the latter at same power, so dont do such conclusions so fast.

 

[Asterox]

If those scores are indeed with that power then its extremely positive and a good competition both in desktop and mobile.
Only a few more days for the reviews, CPU landscape seems to heat up and thats good.

Yes they are, in short i5 12400/performance per watt there is no objection.

i5 12400 65W vs R5 5600 65W, short and simple.But i5 12400 will lounch in 2022, most likely February.

Page 544 - Discussion - Intel current and future Lakes & Rapids thread

Page 544 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[insertcarehere]

But if AIDA use 10% more power than CB, as you suggest, it means that the CPU can clock 5% higher with the latter at same power, so dont do such conclusions so fast.

Somebody here with a 5600x was wondering the exact same thing, turns out the 12400 might be more efficient... :

 

[DrMrLordX]

The 12900k limited to 125W doesn't seem too impressive though. At least not in R20. The 5900X (142W) wipes the floor with it in MT.

 

[TESKATLIPOKA]

i5-12400 6C 4.0 Ghz used 78W in AIDA64 FPU, it's not the worst stress test but it uses more power than Cinebench on Intel CPUs. i5-12400 at 4.0 Ghz is still faster than a Ryzen 5600x in Cinebench R20 which usually clocks at ~4.5 Ghz. This is a huge efficiency increase over Tigerlake-H, I think it looks really promising for mobile.

So 12400 has the same efficiency or a bit better than 5600x at least If It's at 4GHz, right? How did you come to the conclusion that ADL is hugely more efficient than Tiger Lake? Zen3 is not that much more efficient at higher power draw than Tiger Lake.
Techspot

At 75W Zen3 is about 10-13%. I can't say that is a huge increase in efficiency.

 

[maddie]

Somebody here with a 5600x was wondering the exact same thing, turns out the 12400 might be more efficient... :
View attachment 52118

Am I missing something?

13.3% higher score for 18.1% more power should be considered lower efficiency.

 

[Zucker2k]

So 12400 has the same efficiency or a bit better than 5600x at least If It's at 4GHz, right? How did you come to the conclusion that ADL is hugely more efficient than Tiger Lake? Zen3 is not that much more efficient at higher power draw than Tiger Lake.
Techspot

At 75W Zen3 is about 10-13%. I can't say that is a huge increase in efficiency.

Tiger Lake is on 10SF, Alder Lake is on 10ESF. Not the same.
Edit: Or, Intel 7. ;-)

 

[TESKATLIPOKA]

The 12900k limited to 125W doesn't seem too impressive though. At least not in R20. The 5900X (142W) wipes the floor with it in MT.

I think the same. The clocks should be pretty low at 125W TDP, as I calculated in my previous post, that's why I wonder about the laptop version, at least the Alder Lake H.
Ryzen 9 5950X is capable of 9801 points without OC or PBO.

 

[TESKATLIPOKA]

Tiger Lake is on 10SF, Alder Lake is on 10ESF. Not the same.

I don't understand what you wanted to say by this.

 

[Zucker2k]

Am I missing something?

13.3% higher score for 18.1% more power should be considered lower efficiency.

Assuming linear scaling? I think it's close. The example @mikk used had to do with FPU code so it's possible in that scenario, Zen 3 is not as efficient.

 

[Zucker2k]

I don't understand what you wanted to say by this.

ADL is on an improved node.

 

[TESKATLIPOKA]

ADL is on an improved node.

I know, and your point is?
You quoted my reply to mikk, where I asked him why he thinks Alder Lake will be hugely more efficient than Tiger Lake, even though I think It won't be by that much(I think ~15% max), at least If we compare only the models with Golden Cove cores and no Gracemont cores.

 

[insertcarehere]

Am I missing something?

13.3% higher score for 18.1% more power should be considered lower efficiency.

Except we all know that it takes more than 18.1% power to get 13.3% higher clocks/freq, as shown in the bottom spreadsheet where 5600x takes 83w (5w more than 12400) in AIDA 64 FPU at the freq needed to match 12400 in CPU-Z benchmark.