Intel Architecture Day 2020 - Improved 10nm, Tiger Lake, and Xe LP

[porina]

 

Tiger Lake and Process Improvements

Quote

Tiger Lake uses Intel’s 10nm ‘SuperFin’ manufacturing process technology. As part of this launch, Intel has replaced the 10+ nomenclature and instead renamed it to 10nm SuperFin, or 10SF. This is in part due to some of the updates Intel has made to its 10nm process in order to enable some of the features in Tiger Lake.

Intel have not stood still with their well known 10nm problems. While 14nm "pluses" were a joke, at 10nm they will refer to major node milestones instead of pluses.

 

Look at that block in bottom right. "PCIe GEN4". As a reminder, Tiger Lake is the next generation Mobile CPU due to be released in September, so don't expect to see this on desktop.

 

Quote

The story of Willow Cove is going to be a bit confusing for a lot of people. It certainly was to me when it was first explained. But I’m going to rip the band-aid off quickly for you, just to get it over and done with.

The microarchitecture of a Willow Cove core is almost identical to that of a Sunny Cove core.

Willow Cove's higher performance comes in 3 parts.

1. More clock from the improved 10nm process
2. Bigger cache in CPU
3. Faster ram support

It remains to be seen how this will fit in the picture with Zen 3's impending arrival also.

 

The 10nm Sunny Cove cores in Ice Lake were good, but somewhat held back from their potential by their limited clock. Removing that barrier should be interesting. At the same voltage, we get a higher clock out of Willow Cove than Sunny Cove, but we also get more headroom at higher voltages than possible with Sunny Cove.

 

The other notable change is the L2 cache is now going up to 1.25MB per core, up from 512kB in Ice Lake, or 256k in Coffee Lake. This reminds me a bit of the re-balancing they did in Skylake-X when it came out, with 1MB of non-inclusive L2 cache.

 

Quote

A big part of the Tiger Lake/Ice Lake comparison will be the performance difference in graphics. Where Ice Lake has 64 Execution Units of Gen11 graphics, Tiger Lake has 96 Execution Units but of the new Xe-LP architecture. On top of that, there’s the new SuperFin transistor stack that promises to drive frequencies (and power windows) a lot higher, making Tiger Lake more scalable than before.

 

This is similar to the CPU, but with more scaling potential. At the same voltage, far higher clocks can be attained, but it can also scale to lower and higher voltages. Want lower power, or higher performance? There is more room to adapt. 

 

 

My thoughts

"Reports of my death have been greatly exaggerated"

Intel have been working in the background and we now have some indication of where they are going. A combination of process and other design improvements should put them in a good position to keep in the game against AMD.

 

I know many will wish to see these new updates in a desktop CPU, we can still see what we have in the Tiger Lake mobile CPUs soon, and hope for similar on desktop in the not too distant future. While Intel's 10nm problems are well known, it is good to see they are eventually overcoming them and can offer more competitive 10nm offerings, especially given the previous reports of delays on their 7nm process.

 

On the Willow Cove core, while it may seem disappointing at first that there are no significant architectural updates from Sunny Cove, it will be interesting to see how much uplift the increased cache size and ram speed support will give. Sunny Cove already gave a good uplift from Skylake-Comet Lake family, so a further improvement over that would put Intel in as good a position as we can realistically expect to go against Zen 3's expected launch later this year.

 

The improved Xe LP graphics should also help keep Intel keep up. Where performance eventually falls we will have to wait and see, but it shouldn't be too long a wait.

 

There is a LOT more information than covered in this post. Do check out the links below for more information.

 

Sources

https://www.anandtech.com/show/15971/intels-11th-gen-core-tiger-lake-soc-detailed-superfin-willow-cove-and-xelp

 

Intel source

https://newsroom.intel.com/press-kits/architecture-day-2020/

Includes a 2h40m+ video.

[Soppro]

Woohoo, finally some details!! Let me tuck into this ..

[porina]

Just now, Soppro said:

Woohoo, finally some details!! Let me tuck into this ..

Check out the sources and enjoy. Updating the post could take a while!

[RejZoR]

Realistically, only reason I mock Intel is because of their lame insisting on 4c/8t configuration and they would drag this nonsense for another several years if AMD didn't kick their butts. And the fact they haven't really improved their architecture in any way other than just pump out higher clock. Generally nothing wrong with that as it yields performance, but they claim IPC improvements and frankly, there haven't been any since first Skylake. It's just hard to justify such high prices when the tech seems prehistoric and we could be so much further if it wasn't artificially dragged.

 

If they can bring something exciting and new, I'm all for it. Hell, even if it's still on 14nm, I don't care. If they can get it to work better, then great. But they haven't really delivered any of that. It's just rehashing of same old for years now. It's why that 5 core CPU with one big core and 4 Atom cores is exciting thing for me. No one has done it in the x86 space and I'm excited. I'm also excited about Xe. But all their mainstream CPU's are such big fat meh.

[Blade of Grass]

They're saying their new SuperFin tech can give them as much as a 20% boost in performance gen-over-gen, which is apparently the largest step they're ever made. If it works out, I'll be super impressed.

[LAwLz]

32 minutes ago, RejZoR said:

Realistically, only reason I mock Intel is because of their lame insisting on 4c/8t configuration and they would drag this nonsense for another several years if AMD didn't kick their butts. And the fact they haven't really improved their architecture in any way other than just pump out higher clock. Generally nothing wrong with that as it yields performance, but they claim IPC improvements and frankly, there haven't been any since first Skylake. It's just hard to justify such high prices when the tech seems prehistoric and we could be so much further if it wasn't artificially dragged.

It shouldn't really be a surprise that Intel haven't had any IPC gains in the generations since Skylake because Skylake (the architecture) is still the foundation for their cores released today. If you look up Comet Lake you will see that it is a bunch of Skylake CPU cores.

 

The i5 10600K is the most modern mid to high end Intel CPU right now, right? So it's called "Comet Lake" but let's look up what CPU architecture it uses... Oh wouldn't you know it it's Skylake cores!

 

So yeah, Intel has put a lot of effort into their chips in pretty much every aspect, except the CPU architecture which has stayed the same for about 5 years now. Each individual core in Intel's 10th gen processors are like 99% the same as the ones you will find in their 6th gen processors (6XXX series).

 

 

But! This is about to change. This article is about Intel's next gen CPU architecture called Willow Cove. Willow Cove is a refinement of Sunny Cove (only available on mobile CPUs today) and that already has a 15-20% IPC gain over Skylake. Willow Cove will (according to Intel) have an additional 10-20% performance increase over Sunny Cove. Then in 2021 we will get Golden Cove which will increase single threaded performance even more.

 

 

I think Intel is about to make a pretty good comeback after a couple of years of beating from AMD. The CPU landscape will probably get very exciting in the next 2 or so years.

[porina]

5 minutes ago, Blade of Grass said:

They're saying their new SuperFin tech can give them as much as a 20% boost in performance gen-over-gen, which is apparently the largest step they're ever made. If it works out, I'll be super impressed.

We have to be a little cautious in how they measure that improvement. Which 10nm are they comparing to? Note they say Broadwell is the baseline for 14nm performance, and we see it step through the generations. They don't list a 10nm design, not that there are many to choose from. Its first iteration with Canon Lake wasn't great, and I'm not sure how much Ice Lake was over that. On their chart, the improvement in 10nm would be more than that of Broadwell to Coffee Lake. The process chart is at "iso leakage" and I'm not enough into process to understand how that measure works in predicting performance.

 

On the other side, the chart showing Sunny Cove to Willow Cove clocks does suggest about 20% potential improvement at the top end.

[Blade of Grass]

3 minutes ago, porina said:

We have to be a little cautious in how they measure that improvement. Which 10nm are they comparing to? Note they say Broadwell is the baseline for 14nm performance, and we see it step through the generations. They don't list a 10nm design, not that there are many to choose from. Its first iteration with Canon Lake wasn't great, and I'm not sure how much Ice Lake was over that. On their chart, the improvement in 10nm would be more than that of Broadwell to Coffee Lake. The process chart is at "iso leakage" and I'm not enough into process to understand how that measure works in predicting performance.

 

On the other side, the chart showing Sunny Cove to Willow Cove clocks does suggest about 20% potential improvement at the top end.

They say it's coming on Tiger Lake in fall, I guess it would be 20% over whatever the tiger lake predecessor was? 

[LAwLz]

6 minutes ago, porina said:

We have to be a little cautious in how they measure that improvement. Which 10nm are they comparing to? Note they say Broadwell is the baseline for 14nm performance, and we see it step through the generations. They don't list a 10nm design, not that there are many to choose from.

I am fairly sure it's compared to Sunny Cove.

Sunny Cove had issues with clock speed and according to Anandtech Willow Cove will be closer to 5GHz, and also that "most of the performance uplift comes from the process node". So it's most likely 20% over Sunny Cove. It doesn't really make sense otherwise.

 

 

  

1 minute ago, Blade of Grass said:

They say it's coming on Tiger Lake in fall, I guess it would be 20% over whatever the tiger lake predecessor was? 

The predecessor to Tiger Lake was Ice Lake, which had Sunny Cove cores and was 10nm+.

[leadeater]

1 hour ago, porina said:

Willow Cove's higher performance comes in 3 parts.

1. More clock from the improved 10nm process

Well that's good, regression in frequency wasn't exactly optimal, not that I would have taken such a thing if performance increase or such things.

 

1 hour ago, porina said:

The other notable change is the L2 cache is now going up to 1.25MB per core, up from 512kB in Ice Lake, or 256k in Coffee Lake. This reminds me a bit of the re-balancing they did in Skylake-X when it came out, with 1MB of non-inclusive L2 cache.

I'll look it up later unless you post but any idea what the L3 cache size is? Also do you know the L3 cache sizes from the last few archiectures? Not HEDT/Xeon but consumer desktop, can't remember if Intel dropped L3 cache like they did with Skylake-SP.

[Ash_Kechummm]

I know nothing about how the tech industry works, so forgive me if my question sounds dumb (which it probably is), but how does this fit with the rumours that Xe would be cancelled? is that rumour dead now, or can Intel still surprise us in the wrong way?

 

[CTR640]

1 hour ago, RejZoR said:

Realistically, only reason I mock Intel is because of their lame insisting on 4c/8t configuration and they would drag this nonsense for another several years if AMD didn't kick their butts. And the fact they haven't really improved their architecture in any way other than just pump out higher clock. Generally nothing wrong with that as it yields performance, but they claim IPC improvements and frankly, there haven't been any since first Skylake. It's just hard to justify such high prices when the tech seems prehistoric and we could be so much further if it wasn't artificially dragged.

 

If they can bring something exciting and new, I'm all for it. Hell, even if it's still on 14nm, I don't care. If they can get it to work better, then great. But they haven't really delivered any of that. It's just rehashing of same old for years now. It's why that 5 core CPU with one big core and 4 Atom cores is exciting thing for me. No one has done it in the x86 space and I'm excited. I'm also excited about Xe. But all their mainstream CPU's are such big fat meh.

That's ultimate pride, laziness and greed. Intel deserves to be in deep shit for all the bs they've been pulling all those years.

If AMD didn't come back, Intel seriously would've been releasing i7 10900K with 4c/8t! I imagine them CEO's and other VIP's laughing all the way to the bank while calling us morons that keeps buying their shit over and over again when AMD was not back in the game.

[MyName13]

2 hours ago, RejZoR said:

And the fact they haven't really improved their architecture in any way other than just pump out higher clock.

Neither has AMD.Remember, they have only recently made something that is close to Intel's architecture.

[RejZoR]

2 minutes ago, MyName13 said:

Neither has AMD.Remember, they have only recently made something that is close to Intel's architecture.

You can't compare Ryzen to Bulldozer. They cut ties to that thing entirely. Zen 1 was big thing. Zen+ was just a refresh, basically a Skylake. They just improved the clocks sligthly, IPC remained the same iirc. Zen 2 had significant IPC boost compared to Zen 1 and Zen+. Zen 3 is planned to have yet another big IPC bump. Biggest yet actually. So, yeah, not quite the same thing. Not to mention how many moar cores they offer. Not quite the same is it? They made 2 major IPC boosts. Intel had none in several years going into the past from this point now.

[WereCatf]

I really, really hope the Xe HPG GPUs will support GVT-g. If they do, I'll be looking into buying one for my server and use Parsec inside my VMs instead of VNC or Spice -- that'd be damn sweet and Intel would be the only player in the market that supports virtual GPUs on consumer-grade, proper gaming-capable GPUs.

[MyName13]

28 minutes ago, RejZoR said:

. Intel had none in several years going into the past from this point now.

What's the point of AMD's architectural improvements if they were catching up to Intel's architecture?Improving the best thing on the market is harder than improving a crappy architecture that was around 6 years old when Zen 1 came out.

[porina]

1 hour ago, leadeater said:

I'll look it up later unless you post but any idea what the L3 cache size is? Also do you know the L3 cache sizes from the last few archiectures? Not HEDT/Xeon but consumer desktop, can't remember if Intel dropped L3 cache like they did with Skylake-SP.

12MB L3 for a 4 core, or 3MB/core (shown in the Tiger Lake SoC slide). This compares with most i7's through the years at 2MB/core, so that's an increase also. Zen 2 is still ahead on L3 alone, at 4MB/core (assuming max 4 core CCX configuration). Given Intel are now going non-inclusive, and Zen has been exclusive, the total L2+L3 on Willow Cove is now 4.25MB/core, and on Zen 2 is it 4.5MB/core. So in a simple quantity comparison, they are roughly comparable.

 

I assume/hope that L3 remains ring bus even with the L2 change. I'm not sure if that detail has been described anywhere official yet, but it would seem to suffice for 8 cores.

[RejZoR]

1 hour ago, MyName13 said:

What's the point of AMD's architectural improvements if they were catching up to Intel's architecture?Improving the best thing on the market is harder than improving a crappy architecture that was around 6 years old when Zen 1 came out.

Catching up? Zen 1 was equal in IPC to everything Intel had to this moment. They weren't reaching game performance because of raw clock, not because of lack in IPC. People still seem to have hard time understanding what IPC really is...

[MyName13]

24 minutes ago, RejZoR said:

Catching up? Zen 1 was equal in IPC to everything Intel had to this moment. They weren't reaching game performance because of raw clock, not because of lack in IPC. People still seem to have hard time understanding what IPC really is...

Are you sure?I remember r3 1200 (a 4c4t CPU) being barely better than Intel's 2c4t pentium.

[yian88]

Its 10nm+ time already? soon 10nm++ 2021.

[RejZoR]

1 hour ago, MyName13 said:

Are you sure?I remember r3 1200 (a 4c4t CPU) being barely better than Intel's 2c4t pentium.

Yup, you don't understand what IPC is. Also I HIGHLY doubt a proper quad core from Zen class was worse than shitty dual core Pentium... Maybe in games, purely because of clocks. Nothing else. No way.

[porina]

40 minutes ago, yian88 said:

Its 10nm+ time already? soon 10nm++ 2021.

Officially is is 10SF for upcoming tiger lake, and 10SFE for beyond that.

[Kisai]

5 hours ago, RejZoR said:

Realistically, only reason I mock Intel is because of their lame insisting on 4c/8t configuration and they would drag this nonsense for another several years if AMD didn't kick their butts. And the fact they haven't really improved their architecture in any way other than just pump out higher clock.

Pentium 4 says hi. The only time Intel has had to go back to the drawing board when the P4 and the marrying it to RAMBUS utterly sunk it. The Itanium (Itanic) was a failure from the beginning. So going back to the Pentium 3 was the only option, and all the C2D parts were basically Pentium 3 tech.

 

Skylake+++++ just feels like a joke. If they were having problems with the die shrink they should have just not made the commitment to it, as it is the 10nm process will be too far behind when other CPU's are on 5nm. So if you need a new Intel CPU for some reason, ok fine this might buy some time, but the lack of PCIe4/PCIe5, DDR5, and so forth is going to be the deathnell for the Skylake+++++ x infinite. Once DDR5 is available, Intel better have a 5nm cpu in the pipe or they're going to be left behind.

[DrMacintosh]

I live for any and all integrated graphics performance improvements. Sadly I already own the last generation of MacBook that will have Intel Processors. Year over year graphics bumps in low power packages is the only thing that interests me anymore. 

[-rascal-]

1 hour ago, MyName13 said:

Are you sure?I remember r3 1200 (a 4c4t CPU) being barely better than Intel's 2c4t pentium.

 

Closet comparison I can find is the:

i3-10100 (4c / 8t 3.6 GHz base, up to 4.3 GHz boost)

Ryzen 3300X (4c / 8t, 3.8 GHz base, up to 4.3 GHz boost)

 

3300X is ahead most of the time.

 

Take UserBenchmark as a grain of salt, but i3-10300 vs Ryzen 3300X.

https://cpu.userbenchmark.com/Compare/Intel-Core-i3-10300-vs-AMD-Ryzen-3-3300X/4074vs4076