Discussion Intel current and future Lakes & Rapids thread

[JoeRambo]

What is amazing is that reviewers finally realised how important those cooling vents are for and hopefully will put fire on manufacturers. There simply should be no such performance variance in same TDP/weight/size/thickness class devices as we see now

 

[mikk]

ASUS ZenBook Flip S UX371: https://hothardware.com/photo-galle...ttom-asus-zenbook-flip-s-ux371e.jpg&tag=popup
Full test: https://hothardware.com/reviews/asus-zenbook-flip-s-ux371e?page=1

UX371 has air intakes on the back!!! Even more interesting performance mode was working in this test:

The Fan Profiles menu includes a total of three profiles: Whisper, Standard, and Performance. ASUS says the Whisper profile just reduces the fan speed down a bit from Standard for quieter operation. On the other hand, Performance mode enables a higher TDP, higher fan speeds, and as a result -- more performance.

Then we fired up MyASUS and turned the fan profile to the Performance setting. As mentioned previously, this does more than kick up the fans; it also pumps up the TDP to unlock higher performance.

The performance is much better than other UX371 reviews have shown, Cinebench R20 MT compared:

Hothardware: 1975 (Performance) 1607 (Standard)
Notebookcheck: 1285
Digitaltrends: 1254

Performance even below the standard setting from Hothardware, there was clearly something wrong in other UX371 reviews, I wonder if this Bios related or something else. Gaming performance is quite good as well in performance mode: https://images.hothardware.com/cont...th-performance-asus-zenbook-flip-s-ux371e.png

 

[Spartak]

...typo

Meant 2021.

Goddamn it me.

I think people here are failing to take into account the distinction between mobile and desktop. Desktop usually is about half a year behind mobile.

My money is on Alder Lake release for mobile late '21 and early '22 for desktop, so I don't see how both (Rocket Lake & Alder Lake launch in early & late '21) are mutually exclusive.

 

[jpiniero]

I think people here are failing to take into account the distinction between mobile and desktop. Desktop usually is about half a year behind mobile.

My money is on Alder Lake release for mobile late '21 and early '22 for desktop, so I don't see how both (Rocket Lake & Alder Lake launch in early & late '21) are mutually exclusive.

I could see Alder Lake being released on desktop first if they felt like they had to do it for competitive reasons, or if Microsoft is falling behind on the big.LITTLE optimizations. But I agree that if Rocket Lake does get released that they would be unlikely to also release Alder Lake in 2021.

 

[DrMrLordX]

if Rocket Lake does get released

You think Intel will can it? They already released box art and everything! They paid good money for those Avengers characters.

 

[mikk]

I think people here are failing to take into account the distinction between mobile and desktop. Desktop usually is about half a year behind mobile.

I wouldn't say typically. In the old Intel days desktop typically was first and mobile second. Even with a concurrent launch desktop models were available before mobile. The mobile first strategy started with the process struggle. If they expect to be ready with Sapphire Rapids next year you would expect they can be ready for smaller desktop parts as well. Fab 42 is fully operational now, they can ramp their 10nm volume up in the coming months.

As for the Alder Lake generation desktop is cleary first and mobile second. All the leaks, , driver and software work on windows/github and Intels own technical library points to desktop first. They have fully enabled 8+8+1 ADL-S ES parts in the lab. A few days ago the have added Alder Lake-S for the Linux version of Media SDK and nothing about ADL-P, just one example of many.

MFX_PLATFORM_TIGERLAKE = 40,
MFX_PLATFORM_ROCKETLAKE = 42,
MFX_PLATFORM_ALDERLAKE_S = 43,

{ 0x4600, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4680, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4681, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4683, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4690, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4691, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4693, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4698, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4699, MFX_HW_ADL_S, MFX_GT1 },//ADL-S

 

[Spartak]

I wouldn't say typically. In the old Intel days desktop typically was first and mobile second. Even with a concurrent launch desktop models were available before mobile. The mobile first strategy started with the process struggle. If they expect to be ready with Sapphire Rapids next year you would expect they can be ready for smaller desktop parts as well. Fab 42 is fully operational now, they can ramp their 10nm volume up in the coming months.

As for the Alder Lake generation desktop is cleary first and mobile second. All the leaks, , driver and software work on windows/github and Intels own technical library points to desktop first. They have fully enabled 8+8+1 ADL-S ES parts in the lab. A few days ago the have added Alder Lake-S for the Linux version of Media SDK and nothing about ADL-P, just one example of many.

MFX_PLATFORM_TIGERLAKE = 40,
MFX_PLATFORM_ROCKETLAKE = 42,
MFX_PLATFORM_ALDERLAKE_S = 43,

{ 0x4600, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4680, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4681, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4683, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4690, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4691, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4693, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4698, MFX_HW_ADL_S, MFX_GT1 },//ADL-S
{ 0x4699, MFX_HW_ADL_S, MFX_GT1 },//ADL-S

Interesting, that would be a break from their usual modus operandi the last 5+ years. Like many I have a hard time seeing Rocket Lake only have a shelf life of half a year. It's probably delayed by half a year but my guess is they can't afford to let a full year go by.

Anyhow, I would be the last to complain as I'm looking to upgrade when Alderlake hits the shelves.

 

[IntelUser2000]

@mikk Technical reasons are only one factor in release cycles. Business(financial) and OEM interests matter too.

Also, that UX371 even on performance mode doesn't perform that well. The Swift is still better, and 25W versions should be even faster than the Swift at 17W. A good 25W implementation should outperform the reference design.

 

[mikk]

Also, that UX371 even on performance mode doesn't perform that well. The Swift is still better, and 25W versions should be even faster than the Swift at 17W.

Better than the Swift 3 SF313-53 and slower than Swift 3 14 SF314-59 on Ultrabookreview. Temperature limit is key, SF314-59 can go to 90-92 degrees and SF313-53 63-65 degrees Celsius. The limit could change with every new bios though.

 

[coercitiv]

Like many I have a hard time seeing Rocket Lake only have a shelf life of half a year.

We've seen worse with Broadwell, which only got into desktops to tick a roadmap checkbox. Unlike Broadwell, RKL-S does have a serious job to attend to during it's limited shelf life.

 

[Spartak]

We've seen worse with Broadwell, which only got into desktops to tick a roadmap checkbox. Unlike Broadwell, RKL-S does have a serious job to attend to during it's limited shelf life.

Yes i agree they need Rocket Lake NOW. Still stunned they let it come to this. I had a hard time believing jpiniero 2yrs ago when he claimed comet lake was yet another skylake iteration but they couldnt even get sunny/willow cove to the desktop one year later.

I read some time back some heads have rolled and rightly so. As soon as 10nm issues turned serious plan B (backporting) had to be a top priority but they continued to muddle through in denial for at least a year.

There have been made comparisons to the Athlon 64 / P4 era but I don't think the performance gap was as bad back then as it is now with Zen3. On top of that the performance/watt gap is reversed from the dark FX days.

Who would have thought this 5 yrs ago when Intel launched skylake on 14nm and Zen wasnt yet announced.

 

[mikk]

We've seen worse with Broadwell, which only got into desktops to tick a roadmap checkbox. Unlike Broadwell, RKL-S does have a serious job to attend to during it's limited shelf life.

I have to add that depending on the SKU variants and amount of LGA 1700 chips they can initially sell.... it does not neccessarily mean it's the end for RKL-S, they can sell RKL-S and ADL-S concurrent for some time like they do with CML-U and ICL-U/TGL-U. Or may be RKL-S will be a limited SKU release similar to Broadwell DT who knows.

 

[LightningZ71]

Well, for Intel on the desktop computing front, this is substantially worse for them. Not only does AMD have a tentative performance lead and a performance per watt lead (and a rough parity in mobile with Tiger lake vs. Renoir), for Intel, they are essentially selling what used to be workstation class processors in their desktop line at desktop prices. The 10900k would have been a several thousand dollar workstation Xeon just three or so years ago, yet, now, it's being sold for a third of that, and doesn't even difinitively lead it's segment, falling behind competing chips in its price range in MP tasks.

Fo them, they have to sell a bigger, more expensive chip at soimilar prices to the competition. This hurts margin in a big way in a volume segment.

 

[jpiniero]

Fo them, they have to sell a bigger, more expensive chip at soimilar prices to the competition. This hurts margin in a big way in a volume segment.

14 nm has to be dirt cheap at this point. Might even out that way.

 

[piokos]

14 nm has to be dirt cheap at this point. Might even out that way.

I'm not sure why people keep saying 14nm is "dirt cheap at this point". It's not. You still have to provide all the production line devices, wafers, chemicals, energy. People who operate all of this earn as much as they used to (maybe more).
It's the exact same cost... + inflation.
The only thing that may have improved are yields.

You may think it's cheap because R&D was somehow "paid".

If you look at TSMC financial statement for 2019 (6.2.2 Financial Performance):

https://www.tsmc.com/download/ir/annualReports/2019/english/pdf/e_all.pdf

(All in NT$)
Revenue: 1069B
Cost of revenue: 577B
Operating expenses: 119B (inc. 90B for R&D)

Therefore:
R&D costs: 90B, or 8.5% of revenue
Non-R&D costs: 606B, or 56.7% of revenue

8.5% of revenue is clearly a lot. But it's not like chips start to mate and reproduce when you stop developing them...

 

[jpiniero]

You still have to provide all the production line devices, wafers, chemicals, energy.

Right but all that's cheap. The cost is pretty much depreciating the tools and the fab space. The R&D for 14 nm has long been paid for.

IIRC Fab workers don't make much. Being a fab worker is pretty crappy.

 

[IntelUser2000]

Better than the Swift 3 SF313-53 and slower than Swift 3 14 SF314-59 on Ultrabookreview. Temperature limit is key, SF314-59 can go to 90-92 degrees and SF313-53 63-65 degrees Celsius. The limit could change with every new bios though.

You sure you are reading the same review? The Asus was weak on 3D. The 17W Swift still outperforms it on Night Raid and Middle Earth game.

You can also see from their previous reviews of Icelake-based UX325 and UX425 that the graphics performance is seriously below expectations. It's not a one-off thing.

I still want to see a proper 25W system such as the XPS and the Yoga.

 

[jj109]

Therefore:
R&D costs: 90B, or 8.5% of revenue
Non-R&D costs: 606B, or 56.7% of revenue

8.5% of revenue is clearly a lot. But it's not like chips start to mate and reproduce when you stop developing them...

Your numbers are off.

TSMC's Q2 2020 reported $2.8B COGS and $2.5B depreciation on $10.8B of revenue, so equipment depreciation is 23% of their revenue.

Intel's ratio is roughly the same at $9.2B COGS+D&A ($5.2B) on $19.7B of revenue. Their depreciation accounting is slightly higher because of all the 10nm lines that were piddling along until recently.

 

[Accord99]

The 10900k would have been a several thousand dollar workstation Xeon just three or so years ago, yet, now, it's being sold for a third of that, and doesn't even difinitively lead it's segment, falling behind competing chips in its price range in MP tasks.

Fo them, they have to sell a bigger, more expensive chip at soimilar prices to the competition. This hurts margin in a big way in a volume segment

Comet Lake is still only about 206 mm^2 in size, smaller than Sandy Bridge 4C and considerably smaller than Lynnfield.

The Zen 2 CCD may only be 74 mm^2 in size but it also needs a 125 mm^2 IOD. The higher multitasking performing Ryzens CPUs end being 2x74 mm^2 + 125 mm^2 worth of silicon, so it's questionable whether Intel is really at a disadvantage in that particular area.

 

[LightningZ71]

Even with excellent yields, a chip that is over twice the area will still yield significantly fewer per wafer, and those wafers don’t come cheap. Then, you are comparing top end parts, which the 10900k must be, against two processors (the 3900 and the 3950) that are either taking CCDs that didn’t make the cut for Threadripper, or are harvested dies with an out of spec core. That’s also a cost difference. The cost of the GloFo I/o die has to be near trivial in the grand scheme of things as they are using it in $99 3100s.
Going against AMD is their cost of packaging the MCM with the various dies. That’s not cheap, though, in volume, it’s not exorbitant.

Comparing comet lake’s die size to Sandy Bridge is largely irrelevant. The nodes were significantly different back then, market conditions were different, competition was different. Relative ASPs were quite different. I’m talking about the recent and current market. Without AMD being successful with Zen, the 10900k just wouldn’t exist. I doubt they would even have bothered with the whole generation, nor would they have even pushed a workstation Xeon like it.

 

[jj109]

Even with excellent yields, a chip that is over twice the area will still yield significantly fewer per wafer, and those wafers don’t come cheap.

It's not just yields. A 7nm wafer requires so many more process steps that, to keep a competitive production rate, the capital costs are enormous in addition to higher material costs. TSMC charges over twice per 7nm wafer over their 12nm wafer for example.

You even acknowledge it in your own argument. Why is the high speed DDR4 and PCIe4 IO die so cheap? It's because all the capital costs have been amortized over half a decade of service that a 12nm GloFo wafer is a bargain compared to TSMC 7nm.

But all of this is missing the forest for the trees. Your claim that Intel has higher costs compared to AMD's equivalent products doesn't hold water. Look at the quarterly reports. AMD has been accepting lower margins for their products to chase market share. People unaware of this fact may have assumed AMD has lower costs which is driving their disappointment with the Ryzen 5K price hikes, but in reality this is AMD normalizing against Nvidia and Intel.

 

[piokos]

Your numbers are off.

TSMC's Q2 2020 reported $2.8B COGS and $2.5B depreciation on $10.8B of revenue, so equipment depreciation is 23% of their revenue.

Intel's ratio is roughly the same at $9.2B COGS+D&A ($5.2B) on $19.7B of revenue. Their depreciation accounting is slightly higher because of all the 10nm lines that were piddling along until recently.

I copied those figures from the 2019 yearly statement. Link is in the post. I don't know what you mean by "being off". You can check yourself.

Yes, equipment depreciation is included. Building fabs is expensive and is also a cost of manufacturing something. I'm not sure what you mean here as well.

Right but all that's cheap. The cost is pretty much depreciating the tools and the fab space. The R&D for 14 nm has long been paid for.

Non-R&D costs consume 56.7% of revenue. How is that "cheap"?

Seriously, sometimes when I see comments like that, it makes me think that the general "understanding" of chip making on this forum is that you spend billions on development, but then it just needs some sand and watering. And chips just magically jump out of the cutting edge machinery.

IIRC Fab workers don't make much. Being a fab worker is pretty crappy.

I'm not saying they do.
This is the "direct labor" figure in the statement: NT$ 13.7B ($480M), just 1.3% of revenue.

But TSMC employs people outside of fabs as well.
R&D payroll* was NT$ 28.6B (so twice that of fabs).
Administrative and sales: NT$ 9.5B.

This position is actually called "payroll and related expense", so it probably includes most of training, travelling etc.

 

[jj109]

I have to add that depending on the SKU variants and amount of LGA 1700 chips they can initially sell.... it does not neccessarily mean it's the end for RKL-S, they can sell RKL-S and ADL-S concurrent for some time like they do with CML-U and ICL-U/TGL-U. Or may be RKL-S will be a limited SKU release similar to Broadwell DT who knows.

RKL was obviously a backup plan in the event of an even more delayed 10nm. Now that RKL-U/H seems to be canned in favor of TGL-U/H, it's going to be a limited release to fulfill obligations to desktops and workstations with CML filling in lower tiers. Had 10nm taken 6 months more, I reckon RKL then would've been a full stack release.

They've essentially taken the entire cost of producing a new CPU architecture and set it on fire for a couple quarters of revenue until it's relegated to the bargain bin. It's questionable whether it would've cost less to just sell SKL refresh^5 at a discount until ADL launched but they're too committed now.

 

[DrMrLordX]

Now that RKL-U/H seems to be canned in favor of TGL-U/H, it's going to be a limited release to fulfill obligations to desktops and workstations with CML filling in lower tiers.

So what do you think is actually going to happen on the consumer desktop side wrt Rocket Lake-S? Are we seeing a paper launch like the 10980XE?

 

[jpiniero]

Non-R&D costs consume 56.7% of revenue. How is that "cheap"?

That's the cost of the fab equipment and building the fab space. Not the cost of actually running the fab.