Intel Skylake / Kaby Lake

[IntelUser2000]

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.

Their TDP specs pan out quite accurately in reality. It would have to be accurate, or eventually people will figure it out. You have notable exceptions like Core M, where it blows power usage limits but still call it "4.5W". Especially the Broadwell version. That was a major disappointment.

Looking at how much Intel struggles with Core M, Apple absolutely blew them away. They are at a totally different class. Atom power levels with Core performance. It's all Intel fans wanted for years but you get it from a totally different source.

 

[Sweepr]

Some thoughts about the latest round of leaks from this week:

- Skylake-X schedule moved up, now launching at Computex according to BenchLife

Skylake-X is probably the most exciting HEDT product coming from Intel since the original Core i7, and I'm sure they want to make an impact here. One friend told me some applications/games are seeing 10-15% performance per clock increase compared to original Skylake, that's probably because of the new cache architecture and quad-channel DDR4. Expect the entry level 6C/12T to become a very attractive all-around choice for productivity and high-refresh/multi GPU gaming choice.

Also, unless I'm mising something, this is the first time Intel mentions 'LCC' next to one of their HEDT codenames. I wouldn't be surprised if they are planning a special 'MCC' version of it with even higher core count. BenchLife hinted 10C SKL-X would not be their X SKU and than an italian blog reiterated the rumors about 12C SKL-X.

- Coffee Lake-S moved up to Q4-2017

This one come from the same article as above. I read Q4-2017 here the same way as Q4-2016 for Kaby Lake. You could buy one, but official launch was shortly after, during CES. Still, that's a lot sooner than we though. Other reports, except a 2016 DigiTimes article, meantioned CFL-S/CFL-H in Q2-2018 so we're talking 3-6+ months earlier than expected. Bringing 6C to mainstream will allow them to increase core/thread count on every segment. Don't be surprised if we get 4C/8T or even 6C/6T Core i5 and a 4C/4T Core i3. This will close the gap in productivity with the competition in a dramatic way and if they manage to retain or perhaps even increase 1-4 thread Turbo with the new 14nm++ process gaming performance should be stellar.

- Cannon Lake SiSoftware entry shows a 48 EUs integrated GPU (twice as big as Kaby Lake's)

While it's only a single entry, and there's some detail wrong - Cannon Lake is listed as Gen 9 when we know (previous GFXBench leak) it's a Gen 10 product - the prospect of Intel doubling graphics resources is exciting. That means significant performance boost for small/thin laptops where yu can't have a discrete GPU (except exotic external dGPU solutions). Don't forget:
- Broadwell, Skylake and Kaby Lake all packed 24 EUs in their GT2 configuration
- Bigger GPUs operating at lower clocks are usually more efficient than small ones at high clocks
- Skylake-Y as a 2C+GT2 is tiny (only 98.5 mm²!), now imagine that on a much denser 10nm process
- At 10nm even if they double the iGPU for every tier they will still have small dies
- Leaving the higher core count products to 14nm++ (Kaby Lake Refresh and Coffee Lake-S/H) makes perfect sense because of their performance profile
- If past is any indication Ice Lake-S could end up with the an updated 48 EUs iGPU as well
- Apple and the whole industry is pushing graphics performance in mobile

- Ice Lake shows up at Geekbench

This one caught me by surprise and so far no one answered my question. Is it expected to have functional ES for chips so far from launch? Unless Intel pulls a Broadwell/Skylake here and Ice Lake actually replaces Cannon Lake for mobile later in the same year (2018). Perhaps a CES 2019 launch for desktop Ice Lake is not far fetched?

The leak also shows 48KB L1 Data Cache, 12MB L3 and apparently LPDDR4/4x support. If GB isn't misreading the specs, and assuming this is their Ice Lake-U SKU, that's 3x the amount of L3 cache as their current designs (Core i7-7500U has 4MB L3). Cache size don't match any previous Intel product, even Skylake-SP/Skylake-X. I guess core segmentation (client x server) already began.

Intel launches "Kaby Lake-G" that integrates HBM 2 and AMD GPU die within the year

I'm taking this one with a grain of salt, but the prospect of having a 14nm++ (?) Skylake cores next to a Radeon GPU is interesting from a technical point of view. Would HBM 2 improve CPU performance? Regarding graphics performance, don't forget Apple is using a highly binned version and lower clocks to reach their 35W TDP target, bread and butter mobile Polaris 11 for gaming laptops is actually 75W TDP. According to BenchLife there's two SKUs (65W and 100W TDP), that's not a lot of power to play with, but I have little doubt it will still outperform an improved version of Iris Pro 580 (48 EUs + eDRAM, Gen 9). We need to confirm if this is real and how it works first.

 

[raghu78]

Coffeelake could turn out to be Intel's best CPU lineup in a decade given the very high transistor performance from 14nm++ and also due to the fact that AMD are finally providing competition to Intel in the mainstream desktop CPU market. If Intel comes out with a unlocked 6C/12T at 4 Ghz base and 4.8-5.0 Ghz turbo at USD 350 which can easily overclock to 5.2 Ghz average OC on air it would be tremendous. I think Intel should relax their TDP to 100w for core i7 6C/12T and aim for 4.2 Ghz base. That would mean 50% more cores and threads at the same base clocks as Kabylake 7700k for the same price. That would be amaaaazing. AMD will have Zen 2 based Pinnacle Ridge out by late Q1 2018 and it should be a good contest. AMD knows the weaknesses of Zen like slow fabric speeds, high L3 cache latency and thus they can improve on those to narrow IPC gap while GF improves the process to hit higher clocks. A 8C/16T Zen 2 with 4 Ghz base / 4.5 Ghz turbo at USD 399 would be nice. I think with the launch of Coffeelake AMD will have to trim the price range of their 8C/16T lineup to occupy USD 299 to USD 399 as I doubt they can get more than USD 399 given the advent of mainstream Coffeelake 6C/12T at USD 350. All in all I expect gamers and enthusiasts to finally enjoy outstanding multithreaded performance at mainstream prices.

 

[coercitiv]

If Intel comes out with a unlocked 6C/12T at 4 Ghz base and 4.8-5.0 Ghz turbo

Wow, hold your horses. That's 50% more cores, we should be grateful if 6c/12t only loses a few hundred Mhz compared to 7700K. I doubt mainstream TDP will go over 100W.

 

[Bouowmx]

"Unlocked": K processors already exceed 100 W when overclocked.

Anyway, 4.5-4.8 GHz stock Turbo Boost is a bit more realistic.

 

[raghu78]

"Unlocked": K processors already exceed 100 W when overclocked.

Anyway, 4.5-4.8 GHz stock Turbo Boost is a bit more realistic.

14nm++ provides 12.5% higher transistor performance than 14nm+ (Kabylake) and 26% higher performance than 14nm (Broadwell). With 7700k already having 4.5 Ghz max turbo I think Coffeelake can definitely hit 4.8-5.0 Ghz for max single core turbo. btw i was talking about Intel increasing TDP for Coffeelake 6c/12t to 100w for stock config. Overclocked we know any processor will draw more power and quite exponentially if you increase CPU voltage.

 

[ozzy702]

Any hopes that Coffeelake 6c will work in a z170 mobo?

 

[IntelUser2000]

Hiroshige Goto says in his article the reason for eDRAM was because they were pursuing HMC at the time which was also thought to be a higher end version compared to HBM so they needed a mainstream alternative. That came out to be eDRAM.

Now eDRAM is there also for on-die integration but that makes it unsuitable for large capacities since the densities are low compared to DRAM. So it becomes a stepping stone for, HBM. So at some point they decided they will go HBM and that year was supposed to be 2018(2019 in shipping parts). But the AMD die supposedly hastened the introduction. For competitive reasons? Ah, maybe Apple.

He also says that Kabylake-G having an on-package AMD GPU is not only said by mainboard manufacturers but software developers as well. Its very real. Now the question remaining is what benefit AMD gets from it.

The power management part can't be that complex. It's as complex as Pentium D chips were. Kabylake plus AMD graphics just connected by a PCI Express x8 interface. Now, PCI Express x8 may mean a derivative like OPI and DMI. It may not be, for TTM reasons. If they want it out as fast as possible, using a standard x8 interface would do it. Remember with eDRAM they went for custom OPI interface and that took time. Of course, they could have expanded the OPI interface, but that would have to be in planning for a while because only GT3e and GT4e parts have the OPI interface on die.

He also says eventual goal for Intel is to have HBM interface on CPU-die, after JEDEC ratified HBM to be able to use in CPUs. But that's not KBL-G, so the HBM one is on the GPU-portion. The latency won't be stellar, possibly worse than DRAM. It'll work for graphics though.

This also means the die could possibly exist already. Signs really point to Polaris 11.

 

[coercitiv]

Now the question remaining is what benefit AMD gets from it. [...]

This also means the die could possibly exist already. Signs really point to Polaris 11.

Well, we could get a few puns out of the way, like "AMD Inside" or "The future is fusion". Stupid giggles aside, AFAIK the die does not exist, since no current modern AMD die has an embedded HBM controller.

 

[jpiniero]

Now the question remaining is what benefit AMD gets from it.

It's simple.. AMD gets money. It does scream Apple though, and they tend to get what they want.

I do think the GPU part almost has to be Vega. The CPU is probably Kaby Lake Refresh; and the current U line only has 12 3.0 lanes available; so if KBR is the same x16's not going to work.

 

[IntelUser2000]

Coercitiv: Yes you are right. Interestingly, multiple sites reported on Polaris basically being ready for HBM.

I do think the GPU part almost has to be Vega. The CPU is probably Kaby Lake Refresh; and the current U line only has 12 3.0 lanes available; so if KBR is the same x16's not going to work.

We aren't talking U series chips here. It's a Broadwell C replacement aka 5755C/5775C. Cannonlake and even Icelake uses Gen cores. It's the Iris Pro that got canned starting with Kabylake.

Also, early leaks still indicate GT2 Gen GPU die existing in KBL-G. Integration of different heterogeneous architectures are almost a generation behind because it takes time to put them together. You might remember AMD's slide with Fusion saying iGPU uses previous generation cores for this reason.

Vega is this year's product. So is KBL-G. Signs of quick integration suggest Polaris is more likely.

Yeah, Apple seems like one of the reasons. The socket is significantly larger and also BGA.

 

[Qwertilot]

Apple certainly makes it all seem like a vaguely plausible idea. They've got vast money, plenty of resources to help make this sort of thing happen and love single chip solutions.

 

[jpiniero]

We aren't talking U series chips here. It's a Broadwell C replacement aka 5755C/5775C. Cannonlake and even Icelake uses Gen cores. It's the Iris Pro that got canned starting with Kabylake.

KBR is quad core, so I don't see why they couldn't use it.

Also, early leaks still indicate GT2 Gen GPU die existing in KBL-G. Integration of different heterogeneous architectures are almost a generation behind because it takes time to put them together. You might remember AMD's slide with Fusion saying iGPU uses previous generation cores for this reason.

It's not heterogeneous though. The RTG GPU and HBM2 are fused together but they are not connected directly with the CPU. The GPU is in the package but as it says it's connected via PCIe x8.

 

[Arachnotronic]

Apple certainly makes it all seem like a vaguely plausible idea. They've got vast money, plenty of resources to help make this sort of thing happen and love single chip solutions.

This really seems like Apple pushing Intel and AMD to build it a "new" Iris Pro. After all, the 15" MBP already uses Intel CPU + AMD dGPU, why not fuse them together on a single package to reduce board space?

 

[Sweepr]

Well, late 2016 Macbook Pro already got a full Polaris 11 (1024 SPs). If that's an Apple product (likely), expect something equal or better. The fact that there's a 100W SKU as well hints at something a bit more powerful IMHO.

The 65W TDP SKU is probably the one they will choose. With single package integration and process improvements (14nm vs 14nm++; refined Polaris or Vega based GPU) they will be able to reduce board space and average power consumption relative to the current 45W Skylake-H + 35W Polaris 11 solution while (hopefully) improving performance.

 

[mikk]

- Ice Lake shows up at Geekbench

This one caught me by surprise and so far no one answered my question. Is it expected to have functional ES for chips so far from launch? Unless Intel pulls a Broadwell/Skylake here and Ice Lake actually replaces Cannon Lake for mobile later in the same year (2018). Perhaps a CES 2019 launch for desktop Ice Lake is not far fetched?

Not sure if this is a fully functional chip because performance looks subpar with only 1 active CPU core. In September 2011 Intel demonstrated a running Haswell system at IDF: https://www.youtube.com/watch?v=EkRmcV1bnLs&t=26s
https://www.youtube.com/watch?v=EA-H6yuARps

In June 2013 Intel officially launched Haswell, 1 year and 9 months after the first official demonstration. To me it is plausible that they could have first silicon running 2 years in advance.

 

[jpiniero]

I actually think the KB-G might be for a new Mac "Mini".

 

[IntelUser2000]

KBR is quad core, so I don't see why they couldn't use it.

Actually, the KBL-G implementation doesn't make it fit for mobile. Especially the -U that you were talking about. We know at least there Gen graphics are being used.

The package is at a large 58.5mm x 31mm. The TDP is 65/100W. Also the idle power will be quite a bit higher. It won't stop some DTR manufacturers from putting them in a "mobile" devices but it officially won't be.

The point of putting such a high end iGPU solution bars entry into most mobile solutions.

It's not heterogeneous though. The RTG GPU and HBM2 are fused together but they are not connected directly with the CPU. The GPU is in the package but as it says it's connected via PCIe x8.

I could have worded that better. I meant integrating two packages into a single one takes time. Its two discrete components. It can't be further different. The two companies will have their own development schedules. Even inside AMD the CPU development goes separately from RTG. The "integration" happens after the basics are done.

 

[jpiniero]

Yeah, but KB-G has been rumored for some time, so it's been in the works. Obviously putting two separate dies together is easier on validation than say integrating Radeon IP into a Intel CPU.

Apple could put the 65W model in a future 15" MBP but yeah it seems more likely that it is for a future Mini or iMac.

 

[Arachnotronic]

Yeah, but KB-G has been rumored for some time, so it's been in the works. Obviously putting two separate dies together is easier on validation than say integrating Radeon IP into a Intel CPU.

Apple could put the 65W model in a future 15" MBP but yeah it seems more likely that it is for a future Mini or iMac.

I don't think Apple would have Intel do a special chip/integration for a Mac Mini. 65W one seems suitable for Macbook Pro, 100W one for an iMac.

 

[coercitiv]

65W one seems suitable for Macbook Pro

I can understand how this logic works: in a unit like Macbook Pro the new KBL-G would be allowed to use up to 35-45W (current CPU) + 35W (current GPU), so in theory it's up to ~70W. (when dGPU is being used the CPU is unlikely to reach TDP, so one can shave a few watts for combined usage). The max 90W system power consumption reported by notebookcheck falls in line with this estimate. This type of config with 45W TDP CPU and 35W TDP GPU is actually often seen in "multimedia" notebooks, which are also the last category of mobile devices able to fit into thin chassis (thin by comparing with workstation/gaming units).

However, I find 65W TDP to be a bit stretching for Macboook Pro or other "thin" mobile PCs. It may be within parameters for what a decent "silent" dual fan cooling system is able to dissipate, but it reaches today's limit while positioning the chips closer, which affects cooling performance (less time to react for the fans). I would expect a mobile oriented chip to aim for lower TDP, something like 50W. This would be a very nice middle ground, since cTDP would also allow targeting 35W and 65W for special cases.

 

[Sweepr]

Skylake-SP motherboards showing up

https://exxactcorp.com/index.php/product/prod_detail/2263
https://exxactcorp.com/index.php/product/prod_detail/2264

Also confirms up to 205W TDP.


Intel Cannonlake Patches Posted For Mesa

''Hitting the Mesa mailing list on Friday were patches for enabling Intel Cannonlake "Gen 10" graphics within Mesa. Bringing up this new generation of Intel support within the Intel OpenGL Mesa driver was just over 3,700 lines of new code. But this support isn't yet finalized and is still relying upon many "Gen 9" code-paths. ''

http://www.phoronix.com/scan.php?page=news_item&px=Intel-Cannonlake-Mesa

 

[mikk]

CHIPSET(0x3184, glk, "Intel(R) HD Graphics (Geminilake)")
CHIPSET(0x3185, glk_2x6, "Intel(R) HD Graphics (Geminilake 2x6")

+CHIPSET(0x5A49, cnl_2x8, "Intel(R) HD Graphics (Cannonlake 2x8 GT0.5)")
+CHIPSET(0x5A4A, cnl_2x8, "Intel(R) HD Graphics (Cannonlake 2x8 GT0.5)")
+CHIPSET(0x5A41, cnl_3x8, "Intel(R) HD Graphics (Cannonlake 3x8 GT1)")
+CHIPSET(0x5A42, cnl_3x8, "Intel(R) HD Graphics (Cannonlake 3x8 GT1)")
+CHIPSET(0x5A44, cnl_3x8, "Intel(R) HD Graphics (Cannonlake 3x8 GT1)")
+CHIPSET(0x5A59, cnl_4x8, "Intel(R) HD Graphics (Cannonlake 4x8 GT1.5)")
+CHIPSET(0x5A5A, cnl_4x8, "Intel(R) HD Graphics (Cannonlake 4x8 GT1.5)")
+CHIPSET(0x5A5C, cnl_4x8, "Intel(R) HD Graphics (Cannonlake 4x8 GT1.5)")
+CHIPSET(0x5A50, cnl_5x8, "Intel(R) HD Graphics (Cannonlake 5x8 GT2)")
+CHIPSET(0x5A51, cnl_5x8, "Intel(R) HD Graphics (Cannonlake 5x8 GT2)")
+CHIPSET(0x5A52, cnl_5x8, "Intel(R) HD Graphics (Cannonlake 5x8 GT2)")
+CHIPSET(0x5A54, cnl_5x8, "Intel(R) HD Graphics (Cannonlake 5x8 GT2)")

https://patchwork.freedesktop.org/patch/150606/



Some may remember that the same entries for CNL were seen in the windows driver for CNL (only that GT3 and GT4 has been removed for the new Linux entry)
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...-update-vr-zone.2418879/page-10#post-37185583


So I wonder if this is somehow connected to the EU count. For Geminilake it seems logical because EU count should be identical to Broxton, 18 EUs: http://ranker.sisoftware.net/show_r...efdbe9d0e7d3ebcdbf82b294f194a999bfccf1c9&l=en

And the lower variant with 12 EUs just like Broxton makes sense because it is a 2x6 version as well: https://lists.freedesktop.org/archives/mesa-dev/2016-November/134606.html

As for CNL it wouldn't match the Sisoft entry of 48 EUs, however it was a buggy entry with Gen9.

 

[mikk]

Intel 300 series desktop processor motherboard chip code for the Cannon Lake PCH-H, on the current know, will still be maintained at the LGA1151 pin

https://benchlife.info/intel-z370-q370-b350-h310-pch-for-coffe-lake-04162017/

As expected, Coffee Lake still uses LGA 1151.

Intel will introduce the Coffee Lake-S platform for desktop computers in the fourth quarter of 2017.

Q4 2017 would be nice.

 

[Sweepr]

https://patchwork.freedesktop.org/patch/150606/

Some may remember that the same entries for CNL were seen in the windows driver for CNL (only that GT3 and GT4 has been removed for the new Linux entry)
http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...-update-vr-zone.2418879/page-10#post-37185583

So I wonder if this is somehow connected to the EU count. For Geminilake it seems logical because EU count should be identical to Broxton, 18 EUs: http://ranker.sisoftware.net/show_r...efdbe9d0e7d3ebcdbf82b294f194a999bfccf1c9&l=en

And the lower variant with 12 EUs just like Broxton makes sense because it is a 2x6 version as well: https://lists.freedesktop.org/archives/mesa-dev/2016-November/134606.html

As for CNL it wouldn't match the Sisoft entry of 48 EUs, however it was a buggy entry with Gen9.

So this confirms Intel is going big with Cannon Lake's integrated graphics. Cheaper Celeron/Pentium (GT1) gets 24 EUs like current Kaby Lake (GT2), while the Core lineup gets 40 EUs (GT2). New architecture on top.