Question Alder Lake - Official Thread

[Hulk]

With the release of Alder Lake less than a week away and the "Lakes" thread having turned into a nightmare to navigate I thought it might be a good time to start a discussion thread solely for Alder Lake.

 

[BorisTheBlade82]

If you are interested in power and performance efficiency numbers of i9 12900K with different power limits, just take a look at this: https://forums.anandtech.com/thread...pu-architectures.2597905/page-3#post-40624976

This looks really solid. Vermeer does not stand a chance due to the chiplet architecture with the only exception being the 5950x that is still on top thanks to its 16 middle cores.

 

[Grooveriding]

Never say never. But I grok. I play old games on my AMD APUs, to let my RTX and GTX cards rest. Not sure when value GPUs will return, so have to preserve the ones I have.

This is how crazy the GPU market is. I do exactly the same with my 3080. Try not to stress it or run it at 100% load too often. If it dies, at least with Gigabyte's abysmal RMA service, you don't even know if you'll get a replacement, and if you do it will take at least a month. Crazy times. Not like you can just go out and buy another unless you want to swallow paying 2xMSRP to get one immediately from a scalper.

What is up with this chip running like a furnace? Linus' video they used a DH-15 and the thing was hitting 100C under stress tests... Still it is nice to have some competition again from Intel, crazy as that is to be a thing. Don't think these are worth upgrading to from Ryzen 5 series for such small gains that come with high power consumption and ridiculous thermals. The gains over 3rd gen Ryzen are small enough that it's likely we see AMD come back to parity or ahead again with their new 3d v-cache chips in a few months, without the inefficiency of these Intel chips.

Whatever Intel refreshes these with is probably going to be a good chip, faster, not a furnace and by that time DDR5 will have been around for a bit. New DDR spec is always pretty poor out the gate and worth waiting to move to for a bit.

 

[coercitiv]

Given we now KNOW Gracemont is Skylake level IPC w/o HT, how is 8 Skylake cores w/o HT at 48 watts not good? This is not speculating, this is not yesterday. Now these are facts. 8 Skylake Cores @3.7GHz without HT using 48W on the desktop is great.

What facts have you used to come up to this conclusion?

The i7 9900 is an 8c/16t SKU with 65W TDP and a base clock of 3.1Ghz. If we remove the extra power used by HT, which we know scales almost linearly with added performance boost, the result would be a Skylake 8c/8t CPU with 48W TDP and 3.1Ghz. And that's on 14nm. Since then Intel has introduced not one, but two mature revisions of their 10nm technology (10SF and now Intel 7). A "modern" Skylake on Intel 7 would not only match E-core power efficiency @3.7Ghz, it would most likely clock at 4Ghz+ within the same TDP.

If you're still unconvinced, here's a 10510U doing 2.9Ghz in Prime95 while package power tracking is @24W. Remove 2W from uncore/SOC then remove at least 20-25% power from HT and you are left with 17W for 4c/4t SKylake @ 2.9Ghz under Prime95. Switch to a more realistic workload such as CB, increase power allocation by 40% (17W -> 24W) and you'll easily reach 3.4Ghz+. Now switch from 14nm to Intel 7 and you get at least 20% more performance, meaning 4Ghz+.

Spoiler: 10510U 2.9Ghz /w HT @24W

You have been told repeatedly, by Intel themselves nonetheless, that E-cores were built with two main goals:
1. power efficiency at low loads / very low clocks
2. area efficiency at high clocks

Intel made it abundantly clear in their engineering oriented presentations (such as the one at Hot Chips) that efficiency cores will be less power efficient than performance cores when churning through meaningful workloads. The point of E cores is to pack a lot of them in the same area.



Take a look at the graph above, I highlighted the area where P cores offers higher power efficiency over E core. See how it stretches over most of the frequency spectrum of the E core? While Skylake may be a grandpa' core, it still has P core pedigree and was built with high dynamic range in mind.

 

[TESKATLIPOKA]

I found one interesting thing in the anandtech review. Check out the difference in ST vs MT.

A single E-core manages 410 points, but 8x E-cores manage only 2572.
2572/8 = 321.5 points, that's 21.6% less than 410.
410 points should be at 3.9GHz, so 321.5 points per core would mean the clockspeed was only 3060Mhz.

 

[uzzi38]

Given we now KNOW Gracemont is Skylake level IPC w/o HT, how is 8 Skylake cores w/o HT at 48 watts not good? This is not speculating, this is not yesterday. Now these are facts. 8 Skylake Cores @3.7GHz without HT using 48W on the desktop is great. Do the math, 24 of them would use ~5950X power and be as or more performant in highly threaded apps that scale well with cores. The 5950X is still the performance/efficiency champ and e's alone could beat it in specific cases. Or I'm missing something? I'm always willing to learn.

Are you sure about that?

Looks to me like at 3.7GHz they're exactly the same perf/W as Zen 3, except that's without taking into account any extra performance you can extract with SMT. ~5W per core for Gracemonts at 3.7GHz vs 6.1W per core average for Zen 3 at 3.775GHz.

For a small core that's surprisingly low perf/W. I want to see what power/perf scaling is like on Gracemont, but it seems pretty sub-par at 3.7GHz.

 

[TheELF]

What is up with this chip running like a furnace? Linus' video they used a DH-15 and the thing was hitting 100C under stress tests...

Jay2c was amazed at how low the temps are, there is obviously a bug there somewhere and we will have to wait to see which one turns out to be closer to the truth, it's nothing new, how many years did AMD users think there was something wrong with their CPUs because software would show them the distance from max instead of the temps in degrees?

 

[JoeRambo]

Indeed. I think the 12900K is of very narrow appeal, bursty type loads like someone running Photoshop for work, weekly Premiere renders, etc. For heavy sustained compute, 5950 or TR builds just make more sense. For (very high end) gaming, 12600/700 make more sense, for budget gaming sticking with old Zen2/Coffee Lake is advisable IMHO.

While i agree with sustained compute part - i do expect every single guy who needs sustained compute in the form of compiles, renders ( quite a few of those guys are already on GPUs ) and some other cpu friendly HPC already noticed how good TR and EPYCs are. They are out of discussion here, irrelevant to casual desktop.

Now for what is the casual desktop - as in personal workstation, where peak speed is of top importance, 12900K is the fastest chip for day to day work in IDEs, in browser, in Adobe suite, CADs etc. There is just no escaping this fact from reviews other than JEDEC burdened Anandtech this is the fastest chip now.
On top of this mostly strong ST driven performance it can also provide amazing burst of MT performance. So when you need to recompile, or run some workload on WSL or virtualization overall - the capability is right there.

The power at high MT load is irrelevant, these CPUs won't be running anything sustained other than stress tests.

What Intel really did, is break AMDs MT capacity/capability reserve monopoly in this casual desktop segment and i fully expect them to proceed to thoroughly destroying AMD in this ( cap/cap) niche with those 8 +32 atom monsters in the future. AMD cannot afford to throw so many proper cores at casual desktop while keeping them fed with power, mem/interconnect bw in the same socket as proper casual desktop CPUs and APUs.
Competition is shifting to peak ST performance, SOC level engineering with larger and faster caches, speedier interconnects.

HEDT chips are completely different matter and different segment and everyone who needs them, already has them.

 

[NTMBK]

If you are on Zen 3, it's probably a wait and see what comes next scenario.

But if you are on any Intel based system or earlier Zen platform and want to throw a bunch of money at it (probably maybe you got a big GPU already) then it is interesting.

My gaming HTPC is still using a Haswell i5... Seeing it lose to even the E cores makes me think it's time for an upgrade!

 

[coercitiv]

The 12900k is not power efficient when you turn off the e-cores.

So, 8+0 consumes more power than 8+8 (6w) while incurring a 38% deficit in CB R23.

Two observations:

• that's a 28% deficit, not 38% (I understand what you meant though, the P+E score is 38% higher)
• to evaluate the benefits of E-cores we would need to compare with a hypothetical 10 core Golden Cove. A 25% increase in core count would pump the score by ~20%, meaning the performance deficit would shrink to ~15%. Power would still be higher.

Based on the CB score/power data I would say the E-cores in this 8+8 config bring 20-25% more MT perf when normalizing for area and power. Say hi to the new HT!

 

[biostud]

So...intel is back in the game, but a few observations:

Intel might have the top performance crown, but still uses a lot of power.
Intel has the edge with PCIe5 and DDR5 support, but once DDR5 modules are more reasonable priced and PCIe5 devices are available on the market, AMD should have launched AM5 socket with support for both.
Zen3D is just around the corner.
For gaming the GPU is going to be the limiting factor in most cases, whether you have a zen3 or ADL based computer.
GPU prices are still insane, so building a new rig ATM is going to be a costly affair no matter what CPU you end up with, and that is why I'm not going to upgrade my computer even if I wanted to.

So for me it seems pretty close to a tie between to two platforms, and depending on your usage you might want to go with one over the other. Maybe AMD need to lower the price on 5600X, to stay competitive in budget/midrange gaming.

 

[uzzi38]

Are you sure about that?

Looks to me like at 3.7GHz they're exactly the same perf/W as Zen 3, except that's without taking into account any extra performance you can extract with SMT. ~5W per core for Gracemonts at 3.7GHz vs 6.1W per core average for Zen 3 at 3.775GHz.

For a small core that's surprisingly low perf/W. I want to see what power/perf scaling is like on Gracemont, but it seems pretty sub-par at 3.7GHz.

I see now my phone decided not to send this image as context to what I'm talking about:

 

[coercitiv]

Zen3D is just around the corner.

That's their biggest problem, they were probably hoping for a healthy 3 quarters of leadership in gaming, and that won't work due to the 3D cache stopgap solution AMD has thanks to TSMC.

So for me it seems pretty close to a tie between to two platforms, and depending on your usage you might want to go with one over the other.

That's exactly the case, personally will be literally choosing the most comfortable solution between 12700K and 5900X, and the decision will come down to motherboard availability. I'm already committed to DDR4 and partially committed to mITX. Cooler compatibility is another issue, but Noctua has me covered on this front.

For Alder Lake the MB prices are insane, but I see some cash-back offers (equivalent to bundles) that would even out prices, will need to analyze the boards though. One of the reasons I would still like ADL is the way I'm dealing with upgrades: current machine becomes the new Plex server and does benefit from iGPU decode/encode capabilities as well as lower idle power consumption of an APU.

I'll just wait and watch through this Black Friday season to see what comes up for both platforms, for once it feels good to be able to pick and chose.

 

[Zucker2k]

What Intel really did, is break AMDs MT capacity/capability reserve monopoly in this casual desktop segment and i fully expect them to proceed to thoroughly destroying AMD in this ( cap/cap) niche with those 8 +32 atom monsters in the future. AMD cannot afford to throw so many proper cores at casual desktop while keeping them fed with power, mem/interconnect bw in the same socket as proper casual desktop CPUs and APUs.
Competition is shifting to peak ST performance, SOC level engineering with larger and faster caches, speedier interconnects.

Spot on. I've said this before and I'll repeat it here. What AMD did by bringing the 5950x down to mainstream desktop, charging $800 for it while starving it of memory bandwidth (it's clear from ADL DDR4 tests these many cores crave bandwidth in certain scenarios), was a move purely motivated by petty claim to total desktop dominance; both at HEDT, and mainstream. The consequence of it, however, is that AMD is going to be caught up in a core war it doesn't really want or need. What a lot of people are not concentrating on is that the 12600k is performing at 5800x levels at 5600x pricing. That performance (33%?) is a legit generational leap. At the halo level the impact is a bit muted with high power consumption by the 12900k but the same can't be said of the budget level where the performance gap is glaring. Intel has raised the performance/price bar and AMD has to respond, at a time when chip supply is at its highest demand in decades.

 

[JoeRambo]

Looks to me like at 3.7GHz they're exactly the same perf/W as Zen 3, except that's without taking into account any extra performance you can extract with SMT. ~5W per core for Gracemonts at 3.7GHz vs 6.1W per core average for Zen 3 at 3.775GHz.

Is there any source on the web that feeds proper voltage for 3.7Ghz and not those "1P core ran at 5.3Ghz @ 1.35V and 8 small cores consumed 666W due to being fed same, while only requiring 1V instead" sources ?

E core investigation is rather hard at the moment, the "best" is this sentence from Anandtech:

By contrast, in green, the E-cores only jump from 5 W to 15 W when a single core is active, and that is the same number as we see on SPEC power testing. Using all the E-cores, at 3.9 GHz, brings the package power up to 48 W total.

That is in PoV-RAY with P cores idle and package ramping up. HARD to speculate, but E cores are 3W affair @3.9Ghz and ZEN3 is 7W affair @ 3.9Ghz when both are fed proper voltages.

 

[uzzi38]

Is there any source on the web that feeds proper voltage for 3.7Ghz and not those "1P core ran at 5.3Ghz @ 1.35V and 8 small cores consumed 666W due to being fed same, while only requiring 1V instead" sources ?

The workload was locked to the 8 Gracemont's specifically, the GLC cores were idle.

E core investigation is rather hard at the moment, the "best" is this sentence from Anandtech:

That is in PoV-RAY with P cores idle and package ramping up. HARD to speculate, but E cores are 3W affair @3.9Ghz and ZEN3 is 7W affair @ 3.9Ghz when both are fed proper voltages.

How in the hell did you manage to extrapolate package power for the entire CPU going from 5->15W in a single core workload indicating that we're looking at 3W per core at 3.9GHz?

I'm genuinely interested in what kind of maths where involved here.

 

[JoeRambo]

How in the hell did you manage to extrapolate package power for the entire CPU going from 5->15W in a single core workload indicating that we're looking at 3W per core at 3.9GHz?

I'm genuinely interested in what kind of maths where involved here.

I even wrote it is speculation, but Anandtech provides following info for L3/MC/RING power burning:

When one core is loaded, we go from 7 W to 78 W, which is a big 71 W jump. Because this is package power (the output for core power had some issues), this does include firing up the ring, the L3 cache, and the DRAM controller, but even if that makes 20% of the difference, we’re still looking at ~55-60 W enabled for a single core. By comparison, for our single thread SPEC power testing on Linux, we see a more modest 25-30W per core, which we put down to POV-Ray’s instruction density.

So ~45W of power for package delta when 1 big core is firing up. So i generously reduced it in half to account for lower uncore clocks and whatever power savings Intel has for E-Cores and arrived to 24W of non core power in that 48W number.
Speculation, but i think it is even overestimating power for E cores @ 3.9Ghz.

 

[insertcarehere]

Another set of caveats of course, but w/ AIDA FPU E-cores look to be around 2-4w at their max freq (given that PL2 limits haven't been breached in all three scenarios).

 

[TheELF]

While i agree with sustained compute part - i do expect every single guy who needs sustained compute in the form of compiles, renders ( quite a few of those guys are already on GPUs ) and some other cpu friendly HPC already noticed how good TR and EPYCs are. They are out of discussion here, irrelevant to casual desktop.

Now for what is the casual desktop - as in personal workstation, where peak speed is of top importance, 12900K is the fastest chip for day to day work in IDEs, in browser, in Adobe suite, CADs etc. There is just no escaping this fact from reviews other than JEDEC burdened Anandtech this is the fastest chip now.
On top of this mostly strong ST driven performance it can also provide amazing burst of MT performance. So when you need to recompile, or run some workload on WSL or virtualization overall - the capability is right there.

The power at high MT load is irrelevant, these CPUs won't be running anything sustained other than stress tests.

What Intel really did, is break AMDs MT capacity/capability reserve monopoly in this casual desktop segment and i fully expect them to proceed to thoroughly destroying AMD in this ( cap/cap) niche with those 8 +32 atom monsters in the future. AMD cannot afford to throw so many proper cores at casual desktop while keeping them fed with power, mem/interconnect bw in the same socket as proper casual desktop CPUs and APUs.
Competition is shifting to peak ST performance, SOC level engineering with larger and faster caches, speedier interconnects.

HEDT chips are completely different matter and different segment and everyone who needs them, already has them.

Is this the year 3535? Did I take a wrong turn in albuquerque?
Casual desktop are the guys that buy i3 and below, simple things and some easy gaming, what you describe is already HEDT and what you consider HEDT is server/workstation.

That's their biggest problem, they were probably hoping for a healthy 3 quarters of leadership in gaming, and that won't work due to the 3D cache stopgap solution AMD has thanks to TSMC.

Even if z3d obliterates them in sterile game benches, the 12900k will run games on the p cores and thanks to the new scheduler the e cores will act as a completely different CPU, gaming without any loss due to OBS or whatever else you run alongside, not even the 5950x does that because it all looks like one single pool of cores so background tasks still hurt performance by a lot.
And that is something that vcache will not be able to change.

 

[diediealldie]

An observation from the compterbase review.

When limited to 88 W, a 5800x is 2% faster than a 12900k in multi-threaded performance.

View attachment 52360

If we then add in 8 E cores to the 12900k and compare that to adding in 4 or 8 Zen 3 cores on top of the 5800x

View attachment 52361

The advantage for Zen 3 grows to 8% for adding 4 Zen 3 cores or 16% for adding 8 Zen 3 cores. In other words, in this power limited scenario, you get a greater performance increase with an additional 4 Zen3 cores than adding an additional 8 E cores for ADL. It would be interesting to see if this comparison changes at all at even lower power levels.

Quite interesting results in 12900K 8+0(125W) vs 8+8(65W). This kinda indicates that Monts are running far away from their sweet spot in default PL2. Makes me to think that running Alder lake on 125W will give me a cool and decent desktop just like a laptop.
Quite curious how Alder Lake-P will work next year especially a 2+8 variant.

 

[JoeRambo]

Is this the year 3535? Did I take a wrong turn in albuquerque?
Casual desktop are the guys that buy i3 and below, simple things and some easy gaming, what you describe is already HEDT and what you consider HEDT is server/workstation.

Jesus, lets not be so hard on non-English native guys. That is not HEDT, cause that is mixed use, does not require ECC, does not require 64PCIE lanes, does not want to deal with 100W idle and so on. Call it prosumer, call it whatever.

 

[coercitiv]

What Intel really did, is break AMDs MT capacity/capability reserve monopoly in this casual desktop segment and i fully expect them to proceed to thoroughly destroying AMD in this ( cap/cap) niche with those 8 +32 atom monsters in the future.

It's funny you see it this way, as I read Intel's move as reactionary still. In terms of being able to pack more cores, think of the following relative size ratios between the current cores involved: Golden Cove 4 units , Zen3 2 units, Gracemont 1 unit. In terms of silicon area 8+32 Coves+Monts would be roughly equivalent to 32 Zen cores. The more Intel gets engaged in this core count war, the more it scales up to Zen versus Gracemont. It won't be pretty, Zen is a nimble, flexible core.

Intel's first real occasion to genuinely take the lead is Meteor Lake, since they're finally breaking the monolith apart. Until then they'll have to rely on aggressive market moves based on their superior production capacity. (12600K is such an example)

 

[Hans Gruber]

It's all fun and games until AMD responds with their revised Alder Lake pricing. The early adopters can drive the DDR5 pricing down. I will jump on that boat in about two years myself. It looks like AMD will have to bring 3D memory stacking to all the SKU's as well as some "XT" refresh like they did with Zen 2. The 5950 will remain a beast. The very late Intel 10nm will mature and it will be a good time for consumers who need new CPU's.

Let us not forget the wisdom of Intel. AMD was 1st to 1ghz. Intel didn't believe 64 bit CPU's were necessary and Intel believed 4 core processors were all consumers would ever need.

 

[JoeRambo]

Disclaimer: I don't understand a word in this video, but some info is incredibly useful in decision making for ADL, shows what to expect from DDR4 latency in ADL and so on.

 

[jpiniero]

Gotta say regardless of reviews I was expecting Alder Lake to be completely sold out. DIY demand must be really lousy given GPU prices.

 

[diediealldie]

You're quite right there, @Saylick I didn't have high hopes for Alderlake myself. And even with those leaks in the weeks leading up to today's release I was left feeling underwhelmed as more info about power and heat came out. To reach parity with a 1 year old release all the while using more power, or beating a 1 year old release while using significant levels of power that make air cooling fiddly on the higher end of the spectrum feels almost like a DOA platform at this point. Even if future BIOS releases improve performance, it's still at parity or have a slight edge that make it not worth the investment, at least not at current hardware prices and relatively immature DDR5.

All eyes on AMD and what Zen3 3D will offer up. I think I'll stick with the cheap 10th gen I bought for now. I can't see myself buying a more newer Intel platform until Intel manages to deliver a wallop of a performance stride compared to AMD.

This feels more and more like historic Intel. Coming out with a half baked idea and pushing the power on it hoping to beat AMD who cruised along with a better product. Up until AMD fumbled, of course.

Actually, it's possible to pack more than 40Gracemont cores in 12900K die space. Noting that 5950X uses almost 160mm2 core complex(80 x 2), Gracemont MT grinder is actually quite a good machine if similar die space is allocated(thus more core to reduce clock speed).
If we make a hypothetical die that allocates all GC area + IGPU to Gracemonts then that one will be running 40~50 Gracemonts and will be a true MT beast in the desktop segment.