Intel Skylake / Kaby Lake

[Azuma Hazuki]

Sounds like some marketing guy blowing smoke out his...er...circular rear exhaust port to me. Vague as hell.

 

[frozentundra123456]

frozen even when AMD had a horrible architecture with Bulldozer they had decent updates every 12-18 months with Piledriver, Steamroller and Excavator. I think AMD will keep pushing the Zen core at the same rate with reasonably good IPC improvements of 5-10% every 12 months. I think will see Zen 2 based Pinnacle Ridge in Q1 2018 on an improved 14LPP process and a GF 7nm based Zen 3 in H1 2019. It will be much easier to make significant improvements for the first couple of generations with a brand new architecture especially with such a unique CCX design. Improving memory support for high speed DDR4, improving fabric speeds, improving cache performance are some of the most obvious areas for Zen improvements.

Obviously you missed (or are simply ignoring) my point. It is easier to improve from a low level than from a high level. Even with all those improvements, the construction cores never came close to intel in ipc and efficiency. I don't really fault intel for lack of improvement in ipc and clock speed. What I do fault them for is not increasing core counts and making hyperthreading more available on the mainstream platform.

 

[LTC8K6]

This chip looks like a dud, Kaby Lake Refresh-U is a better product. I think CNL-U will not see the light of day.

Is there any reason to trust that result? I think a dual core Phenom chip beats it...

 

[cbn]

I vote for a 10nm+ unlocked desktop i3, for $74. C'mon Intel, you know that you want to.

They're not going to do that, that would be a 50% price cut or so from the current unlocked i3s.

Well, Intel approached that with the G4560 Pentium. (ie, a 2C/4T for about half the price of the Core i3....though it lacks AVX and has a bit lower clockspeed)

But whatever Intel does sell for $74 at the 10nm+ generation should be quite nice.

20th Anniversary Celeron?

 

[jpiniero]

It's 14nm++.

So what exactly is "performance" in Intel's chart? Max clock speed? And it's not like 10 nm is worse than 14 nm+ in that respect. I imagine the focus would be more on the 5.2 W models, which should be very popular with OEMs for fanless laptops.

 

[Arachnotronic]

So what exactly is "performance" in Intel's chart? Max clock speed? And it's not like 10 nm is worse than 14 nm+ in that respect. I imagine the focus would be more on the 5.2 W models, which should be very popular with OEMs for fanless laptops.

Drive current at a given level of leakage.

 

[PaulIntellini]

14++ performance is based on relaxed gate pitch of 84nm.

 

[mikk]

This chip looks like a dud, Kaby Lake Refresh-U is a better product. I think CNL-U will not see the light of day.

ES in the wild is a good indication for its existence, especially a recent entry. It would be very unusual for Intel if they canceled a product so late in development. I guess the Gen10 entry from gfxbench is also from CNL-U because both were clocked at 2.4 Ghz. There is no indication that CNL-U is cancelled. Given that Intel is doing only SoC variants for CNL it makes perfect sense that KBL-Y and KBL-U will be replaced by CNL, there is no Coffe Lake replacement for those. CNL-U also should have a much more capable GT2 than KBL-R.

 

[Ajay]

frozen even when AMD had a horrible architecture with Bulldozer they had decent updates every 12-18 months with Piledriver, Steamroller and Excavator. I think AMD will keep pushing the Zen core at the same rate with reasonably good IPC improvements of 5-10% every 12 months. I think will see Zen 2 based Pinnacle Ridge in Q1 2018 on an improved 14LPP process and a GF 7nm based Zen 3 in H1 2019. It will be much easier to make significant improvements for the first couple of generations with a brand new architecture especially with such a unique CCX design. Improving memory support for high speed DDR4, improving fabric speeds, improving cache performance are some of the most obvious areas for Zen improvements.

So, what...there aren't enough Ryzen threads for you to post this info in? Maybe you could start thread on GFL vs Intel's process tech instead just jumping into a thread on Intel processors and related aspects. I know having AMD back in the game is exciting. I'd rather keep this thread on track.

 

[IntelUser2000]

This chip looks like a dud, Kaby Lake Refresh-U is a better product. I think CNL-U will not see the light of day.

It's stupid to skip a process, which is why its never done. It'll come in limited quantities. Progress is linear, and doesn't happen out of nowhere. Especially now when companies are struggling so mightily so scale down.

Also, on the lower power variants, performance won't be transistor-limited as its power limited. So whatever little regression in performance 10nm has over 14++ will be negligible.

By the way, Intel's density claims are in practice, misleading. Their actual products are at a significant density disadvantage over the competition. Despite their claim in one of the slides that 14nm is 2.5x the density of 22nm, their Core line was at traditional slightly less than 2x. Their logic scale better than traditional, because it used to be 0.55x for caches and 0.65x for logic. On 14nm, it was ~0.55 for both. Slightly better, not as good as they say.

Now, Airmont core was significantly smaller. However, the SoC itself was of a much less shrink.

 

[Arachnotronic]

ES in the wild is a good indication for its existence, especially a recent entry. It would be very unusual for Intel if they canceled a product so late in development. I guess the Gen10 entry from gfxbench is also from CNL-U because both were clocked at 2.4 Ghz. There is no indication that CNL-U is cancelled. Given that Intel is doing only SoC variants for CNL it makes perfect sense that KBL-Y and KBL-U will be replaced by CNL, there is no Coffe Lake replacement for those. CNL-U also should have a much more capable GT2 than KBL-R.

The way I see it:

KBL-U 2+2 -> KBL-U 4+2
KBL-U 2+3e -> CFL-U 4+3e
KBL-H 4+2 -> CFL-H 6+2
KBL-Y 2+2 -> CNL-Y 2+2

 

[Sweepr]

Kaby Lake-U Refresh vs Cannon Lake-U will be an interesting way to compare 14nm++ to 10nm. First one has twice the number of CPU cores in a better performing proccess, the latter will probably pack much more powerful/capable graphics as mikk said.

BTW, Intel indirectly confirmed 15W 4C/8T KBL-U Refresh will Turbo to 4.0 GHz. Subject to change, but nice to see it won't regress in performance in applications that rely on ST performance.

http://intelstudios.edgesuite.net/170328_itmd/pdf/2017_TMD_MurthyRenduchintala_FINAL.pdf

Based on SYSmark* 2014 v1.5 (Windows Desktop Application Performance). Comparing 6th Gen (Skylake): i7-6600U, PL1=15W TDP, 2C4T, Turbo up to 3.4GHz, Memory: 2x4GB DDR4-2133, vs. 7th Gen (Kaby Lake): i7-7600U, PL1=15W TDP, 2C4T, Turbo up to 3.9GHz, Memory: 2x4GB DDR4-2133, vs. Estimates for 8th Gen (Kaby Lake U42): PL1=15W and PL2=44W TDP, 4C8T, Turbo up to 4.0GHz. Additional config details: Storage: Intel SSD, Display Resolution: 1920x1080, OS: Windows* 10 TH2. Note: Kaby Lake U42 performance estimates are Pre-Silicon, apply to top bin, and are subject to change. Pre-Si projections have +/- 7% margin of error.

 

[mikk]

The way I see it:

KBL-U 2+2 -> KBL-U 4+2
KBL-U 2+3e -> CFL-U 4+3e
KBL-H 4+2 -> CFL-H 6+2
KBL-Y 2+2 -> CNL-Y 2+2

KBL-U 4+2 is a new option, it's not a replacement for 2+2 U.

 

[Bouowmx]

4 cores with power limit 44 W, nice, but only for a minute.
For sustained use, 15 W permits 4 cores at approximately 2.0-2.5 GHz (estimating from Core i5-7200U) with idle integrated GPU.

 

[Sweepr]

Looking back, Intel's first 15W i7 using the 14nm process was a 2C/4T clocked at 2.6-3.2 GHz (5600U). With 14nm++ we're talking about twice the number of cores (4C/8T) and Turbo clocks of up to 4.0 GHz for Kaby Lake-U Refresh. Wonder if we will see the same progress with 10nm.

 

[Sweepr]

BenchLife just leaked details about their new Optane product.

Intel Optane SSD 900P: Read speeds up to 2,500MB/s

Intel Optane SSD 900P sequential read speed of up to 2,500MB / s, while the maximum sequential write is 2,000MB / s; 4K random reading part of the maximum to 550K IOPS, and 4K random write is the maximum to 500K IOPS The Although the capacity is different, but the Intel Optane SSD 900P and Intel Optane SSD DC P4800X performance is roughly the same.

https://benchlife.info/intel-optane-ssd-900p-03312017

 

[lolfail9001]

BenchLife just leaked details about their new Optane product.

So, basically retail version of P4800X. We'll see, maybe folks forking out for 960 PRO will consider adding that one to their collection.

 

[Nothingness]

BenchLife just leaked details about their new Optane product.

Intel Optane SSD 900P: Read speeds up to 2,500MB/s

https://benchlife.info/intel-optane-ssd-900p-03312017

Hum I must be missing something because some Samsung NVMe drives are getting more than 3 GB/s sequential read.

 

[Edrick]

Hum I must be missing something because some Samsung NVMe drives are getting more than 3 GB/s sequential read.

Exactly what I was thinking.

 

[LTC8K6]

Exactly what I was thinking.

The story behind those numbers is quite interesting, though. If true, it's probably going to be much better.
https://arstechnica.com/information...ptane-ssd-375gb-that-you-can-also-use-as-ram/

The P4800X can do 550,000 read IOPS and 500,000 write IOPS, but critically, Intel says it achieves this even at low queue depths. The spec sheet figure has a queue depth of 16, and the company says that a queue depth of about 8 tends to be about the limit seen in the real world.

Moreover, Intel says that the latency of each I/O operation remains low even under heavy load. 99.999 percent of operations have a read or write latency below 60 or 100 microseconds (respectively) with a queue depth of 1, rising to 150 or 200 microseconds with a queue depth of 16. Under a comparable load, Intel's own P3700 NAND SSD can only serve 99 percent of operations with a latency below about 2,800 microseconds.

Likewise, under sustained write workloads, the P4800X retains its low latency for reads, whereas the read latency of the P3700 NAND steadily deteriorates as the write bandwidth increases.

 

[dullard]

Hum I must be missing something because some Samsung NVMe drives are getting more than 3 GB/s sequential read.

While 3 GB/s is nothing to sneeze at, that isn't the main advantage of Optane. The main advantage is if you want a responsive system. Opening a program may need several hundred small files. That doesn't require much bandwidth. But at several milliseconds per small file (at least, sometimes far more), even the fastest SSD will take a while to load most programs that are more that just a simple executable. But at only several microseconds per small file, Optane should feel instantaneous for most tasks even when loading complex programs with thousands of small files.

Same goes when saving or loading any data that is scattered amongst many small files.

 

[IntelUser2000]

While 3 GB/s is nothing to sneeze at, that isn't the main advantage of Optane. The main advantage is if you want a responsive system. Opening a program may need several hundred small files. That doesn't require much bandwidth.

Same goes when saving or loading any data that is scattered amongst many small files.

Even regarding bandwidth, when do you actually see peak bandwidth in actual usage? It'll often fall far short of that.

The drives are rated at 550K/500K read/write IOPS in random 4K, and the sequential peak is at 2.5GB/s, 2GB/s. The random 4K performance is identical in writes and nearly identical in reads. That tells you'll see those theoretical peaks reaching far more often with these drives.

The reason that the random performance saturates peak sequential is because of low latency as well. For the P4800X datacenter drive, it achieves ~90K IOPS at QD1 4K. That's 11.1us. At its peak at QD11, it achieves ~500K IOPS or 22us. In comparison, one of the fastest datacenter SSD drives, the DC P3700 does less than 10K at QD1 for >100us latency and 100K at QD16 for 160us latency. The Optane drive is 8-10x the performance. Amazingly, even that is held back significantly by being on the NVMe bus, which is a very fast interface for an SSD.

 

[JoeRambo]

And what is also important is behavior under mixed loads. Once flash is under a certain percentage of writes, read latency/bw is going down fast due to massive write block size, leading to unpredictable latencies for reads. Byte addressability is amazing stuff, and will shine once connected as DIMM.

I view 1st gen Optane like legendary X25-M SSD drives - quantum leap over disastrous flash discrediting "controllers" and "products" from others. But it also had its drawbacks like write speeds were limited to 40MB/s on smaller drives.

2nd and futher gen Optane (esp with V-NAND like manufacturing treatment?) could be revolutionary. ( ofc flash won't go away, just like rust hasn't, it will have uses and niches and advantages )

 

[Nothingness]

While 3 GB/s is nothing to sneeze at, that isn't the main advantage of Optane. The main advantage is if you want a responsive system. Opening a program may need several hundred small files. That doesn't require much bandwidth. But at several milliseconds per small file (at least, sometimes far more), even the fastest SSD will take a while to load most programs that are more that just a simple executable. But at only several microseconds per small file, Optane should feel instantaneous for most tasks even when loading complex programs with thousands of small files.

Same goes when saving or loading any data that is scattered amongst many small files.

Agreed. I was just suprised Sweepr highlighted the BW rather than other characteristics

 

[Sweepr]

INTEL XEON E3-1200 V6 SERIES VALUE COMPARISON

We think the general sweet spot is around the Intel Xeon E3-1240 V6

CPU. It provides Hyper-threading as well as giving access to the overall Intel Xeon E3-1200 V6 platform at a reasonable price. The performance delta between the E3-1280 V6 and the E3-1270 V6 should be relatively minor for an overall increase of almost double, or an impact of 4-13% of total system costs. For those looking at low-cost appliances, the Intel Xeon E3-1220/ E3-1225 V6 CPUs still reign supreme. We looked at the overall hardware costs but there is one more important component to the equation, software license costs. For applications that are licensed on a per-core basis, license costs can easily eclipse hardware costs. For those applications, the Intel Xeon E3-1280 V6 may be worth every penny for the speed bump over lower-end models.

https://www.servethehome.com/intel-xeon-e3-1200-v6-series-value-comparison