Discussion Intel current and future Lakes & Rapids thread

[IntelUser2000]

Are you sure that the 128, 256, and 512 reference the EU count and not an L4 or graphics working memory cache of some sort similar to the old Iris-Pro setup? That's an awful lot of EUs for a highly memory constrained environment...

Yes. Those are discrete graphics, not integrated, or even on package. The 15W Tigerlake chip has 96EUs, FYI.

With 10nm Intel will transition from eDRAM to HBM generation. PCWatch has stated it was always their plan to use HBM, but they didn't have parts that supported it(hence, eDRAM). We might see Iris versions come with HBM, and certain Kabylake-G successors use it too.

I believe though the PCI Express cards will use GDDR6 to lower costs.

 

[mikk]

Are you sure that the 128, 256, and 512 reference the EU count and not an L4 or graphics working memory cache of some sort similar to the old Iris-Pro setup? That's an awful lot of EUs for a highly memory constrained environment...

We aren't sure. But from past experience the last digits were referred to the EU count, quite often in Intels graphics inf we can see EU numbers there. Of course this is Intels dedicated graphics with fast memory.

 

[Bouowmx]

4096 core/512 EU doesn't seem ambitious enough for 2020 in raw throughput (not really a useful measurement), but judgement is still reserved.

 

[aos77s]

4096 core/512 EU doesn't seem ambitious enough for 2020 in raw throughput (not really a useful measurement), but judgement is still reserved.

with just the math on current HD630's performance with 24 EUs that puts the top card with 512 EU's at 9.40Tflops FP32, slightly under the RTX 2080's 10Tflops

if they have architecture improvements like going down to 10nm then i would see their top mid range card beating out the 2080.

 

[Dayman1225]

Are you sure that the 128, 256, and 512 reference the EU count and not an L4 or graphics working memory cache of some sort similar to the old Iris-Pro setup? That's an awful lot of EUs for a highly memory constrained environment...

Those are references to dGPUs not iGPUs

 

[mikk]

with just the math on current HD630's performance with 24 EUs that puts the top card with 512 EU's at 9.40Tflops FP32, slightly under the RTX 2080's 10Tflops

if they have architecture improvements like going down to 10nm then i would see their top mid range card beating out the 2080.

And higher clock speeds. It would be disappointing if they can't exceed the clock speeds of their low power optimized iGPUs. Nvidia can go up to 2.0 Ghz.

 

[aos77s]

And higher clock speeds. It would be disappointing if they can't exceed the clock speeds of their low power optimized iGPUs. Nvidia can go up to 2.0 Ghz.

You’re correct. My math is based on the average of min and max clock speeds so if these things get 1600-1900mhz core clocks I could see that number peak at or above 2080ti fp32. This sounds reasonable if the 512 count is to be seen as the high tier. And the other two mid and entry

 

[ikjadoon]

You mean like AMD is already doing with PBO?

Perhaps? I don't know, unfortunately.

Intel, FWIW, has been using TVB shenanigans since April 2018...

It's all so lesser reviewers copy-paste the manufacturer-determined "specifications" and have a ready written narrative because testing CPU clock speeds in devices consumers will buy is a bridge too far, unfortunately.

 

[jpiniero]

So I did a Q&D estimate of how big the 512 EU part would be, and sure enough I got about 260 mm2. Yeah, that's just about what Navi 10 is. Being competitive with the 5700 XT seems like a good goal to reach for, but can they actually do that, especially on 10 nm?

 

[Guru]

t's not that top clocks have really advanced, they are just getting better at taking what used to be overclocking headroom, and selling it as stock.

Yes. Even 4GHz is on the not sane front.

It's not that top clocks have really advanced, they are just getting better at taking what used to be overclocking headroom, and selling it as stock.

Skylake did worse than Sandy Bridge for overclocks. Since 2011, by refining the same core and process for nearly 5 years they got a 10% increase.

Sure Ryzen 3K can reach 4.6GHz, but with 1 core auto boost enabled. Manual overclock is same as 2K! The top overclock records usually have SMT and most cores disabled and only run on 1 core.

Both Ryzen and Coffeelake chips are blowing their power budgets reaching stock performance. The 2002 then Intel CTO Pat Gelsinger prediction that such processors will have insane heat density is true today. The HSF combo we have on modern day processors are monstrosities and we also have lot more people running water cooling setups.

You might think its trivial because all you have to do is watch, but these engineers are basically trying to add a floor on the world's tallest skyscraper, or trying to make the fastest ever plane go 100km/h faster.

It's the same line of thinking people like Ray Kurzweil has. Just plot a straight line and assume it'll continue!

Fact is on AMD's first 7nm implementation they beat their max clocks over Ryzen 2000 by 400MHz with the 3950x reaching 4.7GHz. Everyone way saying they will regress and offer lower clock speeds than Ryzen 2000, yet its significantly higher clock speeds. 6nm used the same tools, but provides more density and thus about 12% increase in performance or even lower power consumption. I fully expect 6nm to bring in additional 300MHz frequencies, hoping for up to 500MHz, but ultimately it doesn't matter, they can probably squeeze out 2-4% ipc improvement without clock speed boost.

 

[IntelUser2000]

And higher clock speeds. It would be disappointing if they can't exceed the clock speeds of their low power optimized iGPUs. Nvidia can go up to 2.0 Ghz.

Yea. So far all Gen 11/12 implementations were called "LP". The one in Arctic Sound and the one in DG is called "HP".

Clock speed is a reliable way of getting performance improvements.

So I did a Q&D estimate of how big the 512 EU part would be, and sure enough I got about 260 mm2.

It might be slightly larger because it'll need a 256 or even 384-bit GDDR6 PHY plus Tigerlake is supposed to boost performance per EU significantly.

 

[DrMrLordX]

Fact is on AMD's first 7nm implementation they beat their max clocks over Ryzen 2000 by 400MHz with the 3950x reaching 4.7GHz. Everyone way saying they will regress and offer lower clock speeds than Ryzen 2000, yet its significantly higher clock speeds. 6nm used the same tools, but provides more density and thus about 12% increase in performance or even lower power consumption. I fully expect 6nm to bring in additional 300MHz frequencies, hoping for up to 500MHz, but ultimately it doesn't matter, they can probably squeeze out 2-4% ipc improvement without clock speed boost.

I don't really want to be "that guy", but honestly this is one of the few active Intel threads left on the forum. Can't we discuss AMD things elsewhere?

 

[mikk]

Fact is on AMD's first 7nm implementation they beat their max clocks over Ryzen 2000 by 400MHz with the 3950x reaching 4.7GHz.

Only on paper. Ryzen 3000 with OC tops out at 4200-4300 Mhz on air with a huge efficiency drop. It's a low clock for Intels 14nm standard. I'm sure they could have reached this target for desktop CPUs on 10nm this year considering they can go up to 4100 Mhz on a mobile low power version.

 

[jpiniero]

I'm sure they could have reached this target for desktop CPUs on 10nm this year considering they can go up to 4100 Mhz on a mobile low power version.

4.1 is almost certainly the max for Icelake, and we're talking one core only, and it's likely only one 28W model you will never see on the market. If yields were better I imagine something like the 1065G7 would turbo to 4.1 and not 3.9, but no more.

 

[mikk]

4.1 is almost certainly the max for Icelake, and we're talking one core only, and it's likely only one 28W model you will never see on the market. If yields were better I imagine something like the 1065G7 would turbo to 4.1 and not 3.9, but no more.

For mobile surely yes. As for desktop I'm sure they could go a little higher than this.

 

[Guru]

Only on paper. Ryzen 3000 with OC tops out at 4200-4300 Mhz on air with a huge efficiency drop. It's a low clock for Intels 14nm standard. I'm sure they could have reached this target for desktop CPUs on 10nm this year considering they can go up to 4100 Mhz on a mobile low power version.

LOL, who cares about overclocking, you get more performance by boosting 1-2 cores anyways. Plus I've seen Ryzen 3000 OC up to 4.4GHz.

Again who cares about all core OC, the top chip has 12 cores anyways, its impossible to oc all 12 to 5GHz, it would waste so much power anyways.

Fact is their boost clocks are way higher than Ryzen 2000 and it will be even higher with 6nm, FACT!

 

[Guru]

Only on paper. Ryzen 3000 with OC tops out at 4200-4300 Mhz on air with a huge efficiency drop. It's a low clock for Intels 14nm standard. I'm sure they could have reached this target for desktop CPUs on 10nm this year considering they can go up to 4100 Mhz on a mobile low power version.

We saw what Intel brings in the 10 years with no competition, the same old 4 cores for 10 years straight with minor clock improvements and new mobo's each year. So no, there is nothing to expect from Intel, except more milking from consumers, that is historical fact!

Their 10nm is useless, their CEO just gave a speech saying they are skipping 10nm for the most part and going straight to 7nm in 2021. Even their next desktop cpu's are forecasted to be on 10nm, I don't expect any 10nm desktop part, not unless its a 4 core apu with igp for low cost laptops or stuff.

 

[mikk]

LOL, who cares about overclocking, you get more performance by boosting 1-2 cores anyways. Plus I've seen Ryzen 3000 OC up to 4.4GHz.

I have seen Ryzen 3000 who couldn't even go to 4.2 Ghz, AMD blatantly advertised clock speeds they can't reach in reality. Fact is Intel has different standards in the desktop field. For AMD 4200-4300 Mhz is an improvement, for Intel it's bad if you compare it to Coffeelake.

 

[Abwx]

For mobile surely yes. As for desktop I'm sure they could go a little higher than this.

How much higher.?

They can undoubtly extract as much as 10% higher frequency, wich would put the ceiling at 4.5GHz, yet at such a rate they need more than 10% better IPC to not make the new products look average if not mediocre, so far that s the improvement of ICL, quite good but still not enough considering the eventual frequency losses.

Besides their 14nm has been refined as much as possible, and likely that at 4.5GHz it still better than their 10nm when it comes to perf/watt, you pointed the huge efficency loss of TSMC s process when getting to the frequency ceiling, it is likely no different for Intel s 10nm, hence the focus on mobile in the waiting of some process enhancements, it should take a good year after first 10nm are produced for the process being valuably improved.

 

[Guru]

I have seen Ryzen 3000 who couldn't even go to 4.2 Ghz, AMD blatantly advertised clock speeds they can't reach in reality. Fact is Intel has different standards in the desktop field. For AMD 4200-4300 Mhz is an improvement, for Intel it's bad if you compare it to Coffeelake.

At absurd power with less cores. Fact is 3900x with 12 cores and 24 threads uses less power than an 8 core Intel processor in the 9900k, we are not even talking about the factory OC'ed 9900kf.

And AMD don't need to beat Intel with 5GHz clock speeds, when they are beating intel in IPC and pretty much every single multithreaded test there is with less power, more cores, cheaper and much better value.

But they will probably beat Intel at frequency as well with 6nm, while Intel is moving away from 10nm and going to 7nm in 2021, so realistically 2021 on paper, probably very late to actually have products on 7nm, by which time TSCM will have 5nm, which is one of their biggest shrinks yet, even though it's technically only 1 or 2nm from 7/6nm.

Its ridiculous and inefficient for AMD to run 12 cores at 5GHz, that is a heater, not a cpu. They can run 4.7GHz at 2 cores, 4.5GHz at 4 cores, etc....

 

[DrMrLordX]

factory OC'ed 9900kf.

9900KS. 9900KF(C) is the iGPU-less 9900K.

 

[Errorfatal]

We saw what Intel brings in the 10 years with no competition, the same old 4 cores for 10 years straight with minor clock improvements and new mobo's each year. So no, there is nothing to expect from Intel, except more milking from consumers, that is historical fact!

Their 10nm is useless, their CEO just gave a speech saying they are skipping 10nm for the most part and going straight to 7nm in 2021. Even their next desktop cpu's are forecasted to be on 10nm, I don't expect any 10nm desktop part, not unless its a 4 core apu with igp for low cost laptops or stuff.

10nm useless? doesnt the new 4core beat a 12 core 3900x while using 1/10th the watts for single core? Sure maybe it is a hoax. Is it intel's fault they had zero competition? Why should they make new products when the only thing holding us all back is graphics cards (which intel is now going to bless us with ) no intel fanboi here, i loved amd when they could beat intel. I would be all over the new amd chips if they used less power. x570 uses a extra 40w, the new zen chips idle incredibly high for what they are. Excited to see the refresh, im holding out until at least christmas. My 9900k does everything i needed it to now purchased almost 1 year ago for the same price as a 3900x.

 

[DrMrLordX]

10nm useless?

It's not so much useless as it is extremely low-yield. Most of us won't be able to buy 10nm Intel products for awhile due to limited availability.

 

[Errorfatal]

It's not so much useless as it is extremely low-yield. Most of us won't be able to buy 10nm Intel products for awhile due to limited availability.

Understandable, Imagined there would be laptops at xmas with the 10nm for all of us that wish? How limited will it be and why?

 

[DrMrLordX]

Understandable, Imagined there would be laptops at xmas with the 10nm for all of us that wish? How limited will it be and why?

The laptops will be available. We don't know how limited the quantity. The why: each wafer will only yield a certain % of usable dice. Yields may be as low as 15%. Intel isn't talking about that right now. Also, a lot of the wafers will be sacrificed for IceLake-SP (server processor).