Intel Skylake / Kaby Lake

Page 390 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Why? I would think Intel is capable of re-architecturing the iGPU to increase perf and perf/watt. Or do you think they will be lazy and just go with a 'dumb' shrink?

I don't expect anything significant.

End of Moore's Law scaling means it takes lot more effort and resources to do the same thing as before. With Pascal Nvidia had to significantly alter parts of the chip to raise clocks, despite it being "2 nodes ahead". Before, they would have used that effort and time for something else because much simpler process porting would net the same thing.

Can't be attributed solely to laziness when everyone is struggling. I don't believe in the conspiracy claims they are holding back either. We must simply accept computer chips are in the same advancement rate as every other tech sector not subject to Moore's Law like rapid scaling.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
Right.

Wasn't Cannonlake the one with 48EUs? Is that changed too or it was always Icelake?

Early on in development GPUs had room for noticeable gains at same process and power. Now it's nearly all dependent on power and thermal constraints. The last generation that brought big performance-wise changes were Gen 7. Gen 8 and 9 advances were in corner cases scenarios and overall minimal. I do not believe it'll change in Gen 10.

So assume Cannonlake brings 30% gain. Icelake for a other 30%?


Even with a 25% or 20% EU increase from Broadwell and Skylake Intel was able to gain +20-25% better performance in the mobile space, with Kabylake they gained another 20-25% performance without any EU increase. That was only a result of a higher Turbo and 14nm+. That's why I think your expectation is ridiculously small if GT2 is really a 48 EU version. Assuming they are going with twice the EU count Intel could drastically reduce its GPU clock which is usually far more power efficient than a smaller high clocking GPU design. Beside the power limitation a big factor is bandwidth limitation by the way. Also I think there is lots of room for µarch improvements on Intels side because Nvidia has a far better GPU design than Intel as well as AMD.


We can't say for certain that it comes with 48 EUs, especially why is it named Gen9 in Sisoft. Initially I thought about a misnamed CFL-U which is coming in a GT3 config but this model should come with a Quadcore and not Dualcore. But even for CNL we can never know if Intel changed their mind and is building a CNL GT3 config even if the Roadmap and CNL Linux platform enabling stated CNL U 2+2. Hopefully GT2 really comes with 48 EUs.
 

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
Something I noticed from the BenchLife article about Skylake-X being pushed forward:

skylake-x.jpg


Has Intel ever mentioned 'LCC' next to their HEDT product codenames? I wonder if there's a 'MCC' version (up to 18C/36T) coming later. :D
 
  • Like
Reactions: Drazick and Ajay

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
Can't be attributed solely to laziness when everyone is struggling. I don't believe in the conspiracy claims they are holding back either. We must simply accept computer chips are in the same advancement rate as every other tech sector not subject to Moore's Law like rapid scaling.

Sorry, I meant 'lazy' as in not willing to spend the $$s needed. Perhaps the ROI is just terrible.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
I don't think it's a fake. May be buggy, but imho only the latency test could be affected. Other entries looks legit.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Also I think there is lots of room for µarch improvements on Intels side because Nvidia has a far better GPU design than Intel as well as AMD.

Certainly this is true. Nvidia is on par with Intel perf/watt even though they are using discrete components and Intel has everything integrated. Iris Pro was a major disappointment because of that.

But,

Like I said Ivy Bridge with 22nm process only brought 30% improvements on the 15W U levels. Haswell did another 30%. So I keep my expectations conservative. If anything, in the semi-world keeping it conservative is realistic because it happens so often.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Sorry, I meant 'lazy' as in not willing to spend the $$s needed. Perhaps the ROI is just terrible.

Yep. That's it. But "ROI" as in not just money, but people, tools, effort.

The most amazing thing nowadays is how much effort it takes to do less than before. Not more, less. Because its on a nano-scale, making flat transistors must be very difficult. But FinFET and TriGate makes it 3D. No longer flat. But then the gains are less than we got in the Golden sans-High K, sans-FinFET, sans-Strained Silicon days.

We've yet to hit the hard size limits of shrinking. You know why node names aren't anywhere related to actual sizes nowadays? Because the minimum size determines the difficulty of production, meaning its connected to yield and functioning chips. So with 45nm Intel went High-K because the dielectric was too small, and they wanted to make it bigger, but achieve "same" benefits as shrinking without High-K. On 22nm Intel went TriGate to avoid scaling down gate length but try to get the same benefit as if they did scale gate length. It's not lying, but a practical means to solve what's an increasing issue.

The architectural advancements are mostly based on concept known for decades. The biggest difference from now and 30 years ago is that now we can produce them. Architectures cost more to do less too.

An article I read compares the advancement in computers to advancements in Aeronautics.

In 1903, Wright Brothers demonstrated the first powered flight. The engine was probably weak as the same one some people mount on their bicycles. The whole aircraft was made of materials probably less durable than modern parachutes. In a span of 70 years however, airplanes advanced to a point where they could ferry 200-300 passengers 20k feet or so up in the air at 500 miles an hour. All with interior comfort exceeding Greyhound travels(based on my experience ;)).

40 years later, we basically have the same technology as that first commercial jet. In fact, the two biggest airplane manufacturers, Boeing and Airbus had so much problems with A380 and 787 Dreamliner, but historically the two airplanes are at standstill in terms of technological prowess.

So at some point, Airplane advancement full-on halted. I'm pretty sure people thought we'd be flying regularly to Mars by now and travelling around the globe would take 2-3 hours.

Reason? Too noisy. Too polluting. Too expensive. Too expensive for customers. Sure, ROI. :)

For some who read my comments about competition being an ultimate limiter. Idontcare mentioned this as well. We may be able to take it to the next level, if we all cooperate. Companies worry so much about competition nowadays that they reduced revealing circuit advancements at ISSCC. Litigation. Patent trolls. Heck, stop fighting and work together!

The problem isn't just technological nowadays. Social and Economic problems are causing problems in advancements. Moral too.
 
Last edited:

Nothingness

Platinum Member
Jul 3, 2013
2,371
713
136
I don't think it's a fake. May be buggy, but imho only the latency test could be affected. Other entries looks legit.
The cache sizes look odd: 48KB L1 D size; and why would they put 12 MB of L3 cache on what looks like a U CPU (as it uses LPDDR)? Bandwidth looks very high too, unless the 4 GB of memory are some form of on-chip RAM? Intriguing :)
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
The cache sizes look odd: 48KB L1 D size; and why would they put 12 MB of L3 cache on what looks like a U CPU (as it uses LPDDR)? Bandwidth looks very high too, unless the 4 GB of memory are some form of on-chip RAM? Intriguing :)

I hope its as big of a change as it looks. I am glad(if its true) PC finally catches up to mobile in DRAM standards. I mean, Intel used to be a leader in all kinds of standards. Apple comes and just blows them away. That's what change in focus for the industry does I guess. PC to mobile.

Micron was talking about how it was all the investment and interest in the DRAM ecosystem that prevented alternative technologies from being successful. By the time it went from paper to production, DRAM advanced enough to cancel whatever chance the alternative had.
 
  • Like
Reactions: Drazick

Ajay

Lifer
Jan 8, 2001
15,332
7,792
136
For some who read my comments about competition being an ultimate limiter. Idontcare mentioned this as well. We may be able to take it to the next level, if we all cooperate. Companies worry so much about competition nowadays that they reduced revealing circuit advancements at ISSCC. Litigation. Patent trolls. Heck, stop fighting and work together!

Great post! Yes, given the challenges of advanced node development in this nano-scale era, we need an effort more akin to the ISS* rather than the ISSCC. So long as companies can achieve high profitability without greater collaboration, I don't suppose they will change. EUV, despite industry funding, turned out to be *much* harder to implement than was expected. As EUV is introduced, semicon fabs will experience greater exposure to quantum effects requiring more complex physical xtor geometries and unimaginable precision. If it weren't for the steady demand of less advanced nodes for a broad range of semiconductor devices, consolidation in this industry would be proceeding at an even faster rate.


* International Space Station
 

thepaleobiker

Member
Feb 22, 2017
149
45
61
@VirtualLarry heads up! the Kaby Lake Pentium G4560 is back in stock at low prices..... https://slickdeals.net/f/10045684-i...-3-5ghz-8-0gt-s-3mb-lga-1151-cpu-retail-49-99

The price is waaay below MSRP - will other Intel CPUs drop in price? Anyone have their ear to the ground on this one? Thanks in advance.

SuperBiiz has Intel Pentium G4560 Dual-Core LGA 1151 Kaby Lake Processor on sale for $54.99 - $5 w/ coupon code SPRINGBREAK = $49.99. Shipping starts at $4.79. Thanks Rupchik

Features:
  • Socket: LGA 1151
  • No. of cores: 2
  • No. of threads: 4
  • TDP: 54W
  • Base clock frequency: 3.5 GHz
  • Intel HD Graphics 610
Regards,
Vish
 
  • Like
Reactions: Sweepr

Sweepr

Diamond Member
May 12, 2006
5,148
1,142
131
Hiroshige Goto's word about Kaby Lake-G:

Intel will introduce "Kaby Lake-G" which integrates stacked DRAM "HBM 2" into CPU package within the year. Like the conventional eDRAM version CPU, the DRAM of the HBM 2 is sealed in the CPU package. However, there are some major differences. As already reported in the rumors, the GPU core is not Intel 's built - in core (it also has an embedded core), a third - party discrete GPU die.

http://pc.watch.impress.co.jp/docs/column/kaigai/1054618.html
 
  • Like
Reactions: Drazick and psolord

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
HBM 2 could give a nice boost for the CPU, it should be way faster than their edram. I wonder if Intel is going to use HBM 2 instead edram for its Iris Pro SKUs in the future.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
HBM 2 could give a nice boost for the CPU, it should be way faster than their edram. I wonder if Intel is going to use HBM 2 instead edram for its Iris Pro SKUs in the future.

I can't imagine that being the case.

HBM is mostly a bandwidth play. Their eDRAM is built on their high performance logic process(at least high performance compared to DRAM processes). Even then, the effects on CPU performance can be minimal at times. PCWatch states Intel's original intention for the eDRAM eventually being included in caches. HMC in Knights Landing shows the latency to be somewhat worse than DDR4 memory. eDRAM has half the latency of DDR3. Latency, is what determines performance in CPU memory, much more so than bandwidth.

Also, the PCWatch tells us Radeon cores are going to be used for KBL-G. That means death of Iris Pro, at least for PCs(perhaps they will keep it around for niche server applications). The reports of Ice Lake using Gen cores may indicate for low power applications, Gen isn't too bad and it has the advantage of being on the die rather than off die with EMIB.

Interestingly, PCWatch also says its cost of the interposer that kept HBM from being adopted, and EMIB will allow that cost to be lowered so it can start penetrating more mainstream markets. It means the high end iGPU is still going to be on the Intel side, just using Radeon cores. I am not sure of their theory it'll use 256GB/s bandwidth version. That's identical to RX 480. Considering that the whole chip has only 100W to play with, I can't imagine anything higher than RX 460 cores in there. 128GB/s will suffice.
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
Considering that the whole chip has only 100W to play with, I can't imagine anything higher than RX 460 cores in there.
I refrained from posting this last time we spoke about KBL-G (about which it seems I'll have to taste some crow) since I felt it would be too much about AMD, but since it's back in the spotlight, Polaris 11 in it's mobile incarnation is a 35W TDP chip. Also as an interesting fact, the entire PS4 Pro uses between 100-150W while gaming, and that's 2304 GCN cores in there among many other things. Power considerations alone - they could fit much more than RX 460 at efficient clocks (900 vs. 1100-1200Mhz).

4 KBL cores would only need 25-35W to offer good base clocks, the rest could all be diverted to GPU budget. I still remain very skeptic abut this though.
 
  • Like
Reactions: Drazick

Zucker2k

Golden Member
Feb 15, 2006
1,810
1,159
136
Snip.

For some who read my comments about competition being an ultimate limiter. Idontcare mentioned this as well. We may be able to take it to the next level, if we all cooperate. Companies worry so much about competition nowadays that they reduced revealing circuit advancements at ISSCC. Litigation. Patent trolls. Heck, stop fighting and work together!

The problem isn't just technological nowadays. Social and Economic problems are causing problems in advancements. Moral too.
Competition far outweighs cooperation, in the long term, especially in technology. Think of all the forks in technological development, which resulted in mass adoptions and rejections, with the adoptions further serving as stepping stones to even greater technological advancement. Without these crucial forks, we could have hit insurmountable walls many times over. So competition, just as in Darwinian theory on evolution, is crucial to development. In technology though, there's the added advantage of being able to go back to a fork in the evolution cycle and going with an alternative idea or implementation should one hit a wall on their current branch.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
I refrained from posting this last time we spoke about KBL-G (about which it seems I'll have to taste some crow) since I felt it would be too much about AMD, but since it's back in the spotlight, Polaris 11 in it's mobile incarnation is a 35W TDP chip. Also as an interesting fact, the entire PS4 Pro uses between 100-150W while gaming, and that's 2304 GCN cores in there among many other things. Power considerations alone - they could fit much more than RX 460 at efficient clocks (900 vs. 1100-1200Mhz).

4 KBL cores would only need 25-35W to offer good base clocks, the rest could all be diverted to GPU budget. I still remain very skeptic abut this though.

You can't take the power usage in gaming to TDP scenarios. TDP is far more demanding than most usage scenarios. Its a realistic maximum power use over a longer period of time. It would have to stand scrutiny in programs like Furmark, or whatever GPU burn programs are out there.

That said, you do have a point. I think they may be able to put a full-chip Polaris 11. Unless there's something in between though, its a big jump to Polaris 10.
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
You can't take the power usage in gaming to TDP scenarios. TDP is far more demanding than most usage scenarios. Its a realistic maximum power use over a longer period of time. It would have to stand scrutiny in programs like Furmark, or whatever GPU burn programs are out there.
I was going to both agree with you (on a general note) and disagree for Intel's specific TDP implementation (after Turbo time limit, cTDP defines max package power, no exception) but then it hit me: wonder how tight integration would be on such a chip, whether Intel's power management would still have real time control on power usage, or if they would have to consider some sort of ceiling for the GPU part.
 
  • Like
Reactions: Drazick