The Intel Atom Thread

[bullzz]

goldmont plus seems to be well above a73 performance range. shame they dont have a soc to go against snapdragon 845.

 

[raghu78]

goldmont plus seems to be well above a73 performance range. shame they dont have a soc to go against snapdragon 845.

SD845 thrashes Gemini Lake in all areas - ST,MT and GPU even when running at lower TDP.

https://browser.geekbench.com/v4/cpu/5751914
https://www.fool.com/investing/2017/12/27/does-qualcomm-incs-snapdragon-845-deliver.aspx

 

[Arachnotronic]

goldmont plus seems to be well above a73 performance range. shame they dont have a soc to go against snapdragon 845.

Goldmont Plus is another good step for Atom, but the CPU is still behind A75 and in terms of the full SoC for smartphone use, Gemini Lake is primitive compared to Snapdragon 845 (or even 835 for that matter).

If Intel wants to go into phones again, they'd need a much better uncore to have a viable product.

 

[raghu78]

Goldmont Plus is another good step for Atom, but the CPU is still behind A75 and in terms of the full SoC for smartphone use, Gemini Lake is primitive compared to Snapdragon 845 (or even 835 for that matter).

If Intel wants to go into phones again, they'd need a much better uncore to have a viable product.

Its funny and tragic that Intel has done no major briefings to the press or released performance information given that Gemini Lake with Goldmont Plus cores is such a big improvement over the earlier Apollo Lake generation with Goldmont cores. Its really unexplicable how Intel focusses so much of its marketing / PR on the big cores family like KBL/CFL which have the same core and IPC as SKL but does not bother to do any real detailed technical / performance disclosures for such a big improvement with the small core family.

 

[NTMBK]

Its funny and tragic that Intel has done no major briefings to the press or released performance information given that Gemini Lake with Goldmont Plus cores is such a big improvement over the earlier Apollo Lake generation with Goldmont cores. Its really unexplicable how Intel focusses so much of its marketing / PR on the big cores family like KBL/CFL which have the same core and IPC as SKL but does not bother to do any real detailed technical / performance disclosures for such a big improvement with the small core family.

Small core is aimed at the cheap craptop market. No point marketing that to enthusiasts, they want us buying i5s and i7s. Not to mention that it would drive home to investors that they have given up on mobile.

 

[Dayman1225]

CPU comparison chart for low power cores 1.3 (new #GoldmontPlus infos) - InstLatX64

 

[Thala]

SD845 thrashes Gemini Lake in all areas - ST,MT and GPU even when running at lower TDP.

https://browser.geekbench.com/v4/cpu/5751914
https://www.fool.com/investing/2017/12/27/does-qualcomm-incs-snapdragon-845-deliver.aspx

Problem is not only that 4-wide Gemini Lake is crushed by 3-wide Cortex A75 but already by 2-wide Cortex A73. Your links above refer to the Pentium N5000, which has a turbo of 2.7 GHz while most smartphone Snapdragons 835 do not run faster than 2.4GHz.
Now keep in mind that ARM will reveal Cortex A75 successor in 2018, where ARM is expecting another 20% performance boost - possibly by also going 4-wide.

 

[mikk]

SD845 thrashes Gemini Lake in all areas - ST,MT and GPU even when running at lower TDP.

https://browser.geekbench.com/v4/cpu/5751914
https://www.fool.com/investing/2017/12/27/does-qualcomm-incs-snapdragon-845-deliver.aspx

This is another Windows vs Android comparison which is flawed for Geekbench in favour to Android. And Geekbench is a well known best case bench for Android.

 

[dark zero]

Problem is not only that 4-wide Gemini Lake is crushed by 3-wide Cortex A75 but already by 2-wide Cortex A73. Your links above refer to the Pentium N5000, which has a turbo of 2.7 GHz while most smartphone Snapdragons 835 do not run faster than 2.4GHz.
Now keep in mind that ARM will reveal Cortex A75 successor in 2018, where ARM is expecting another 20% performance boost - possibly by also going 4-wide.

Even at A72 the performance difference is not that high...
Here is an A72 core at 2.7 Ghz... from Mediatek, which is the worst of the ones who uses Big Cores
https://browser.geekbench.com/v4/cpu/4589302

And here is from a Pentium N5000
https://browser.geekbench.com/v4/cpu/5751914

The difference is not that high...

At least, all of them are on Core 2 Levels.

 

[ksec]

http://browser.geekbench.com/v4/cpu/compare/5744016?baseline=5751914

Judging from this it offer similar performance to my ageing but capable Sandy Bridge, but with HEVC decode and slightly better graphics. My only problem is,

No USB 3.1 support ?
$161 ? Are we going to see MB with N5000 soldered in for less then $100?

 

[Nothingness]

This is another Windows vs Android comparison which is flawed for Geekbench in favour to Android. And Geekbench is a well known best case bench for Android.

The difference between Android and Windows is about 10% on N4200. So SD845 would stll have the lead if that difference is similar on N5000.

 

[french toast]

..all well and good, but at what power consumption is this performance coming at?
Goldmont was demonstrably faster than silvermont..but seemed to be less efficient overall...hence why intel ran out of the smartphone race with its tail between its legs when announcing it.

But goldmont plus is going to be a better bet for Windows even though A75 is certainly going to be a better processor overall, due to the emulation tax...if only all apps were compiled for universal apps.
Then we could get some proper comparisons between architecture.

 

[Bouowmx]

$161 ? Are we going to see MB with N5000 soldered in for less then $100?

For DIY, ASRock has the motherboards.

For the current generation:
Celeron J3455 costs 67 USD: https://pcpartpicker.com/product/TC...-j3455-up-to-23ghz-micro-atx?history_days=365
Pentium J4205 costs 102: https://pcpartpicker.com/product/YB...ini-itx-motherboardcpu-combo?history_days=365

 

[Thala]

The difference between Android and Windows is about 10% on N4200. So SD845 would stll have the lead if that difference is similar on N5000.

There is not a Windows tax per se, since there are no OS calls at all in Geekbench, which could impact results. It is rather a different compiler used for the Windows build. I would have preferred if they had used the same compilers and options.
On the other hand, there is currently no GCC compiler available, which generates Windows PEs for ARM. So a theoretical native Geekbench for Windows ARM would most likely use MSVC as well.

 

[Brunnis]

One of the first tests of a Gemini Lake product, the Asrock J4105-ITX, which is equipped with a Celeron J4105: Link (translated from german)

The performance increase from Apollo Lake is pretty massive in Cinebench R15, which is the only test they ran:

OpenGL: 21.19 FPS
ST: 73
MT: 263

Compare this to my own tests of the Apollo Lake (Pentium J4205) based Asrock J4205-ITX, which I posted in this thread in November 2016:

OpenGL: 21.57 FPS
ST: 52
MT 193

The CPU part of the Gemini Lake Celeron is ~40% faster than the Apollo Lake Pentium! It's also interesting that power consumption does not appear to have increased, at least on the system level. From what I can gather this is slightly better than the venerable old Core 2 Quad Q6600 @ stock. Given that they both run at basically the same frequencies, it seems Goldmont Plus has slightly better IPC than Conroe/Kentsfield. Obviously with the help from much faster RAM than what was available at the time, but still. More results are needed, but this architecture seems pretty impressive.

EDIT: My old tests can be seen here.

I'm really tempted to get the NUC7PJYH with the Pentium J5005. Should be a pretty awesome NUC for media and emulation. The native HDMI 2.0 should solve a lot of the problems people had with display output.

Out of pure academic interest, it would be interesting to see how Gemini Lake compares to modern ARM chips in terms of performance and power consumption. My guess is that performance is comparable to the latest high-end offerings (excluding Apple...), but the power consumption comparison is harder to guess the outcome of.

 

[mikk]

This is a major increase, much bigger than Airmont and Goldmont brought us, it's puzzling that Intel calls it Goldmont+ which implies it's a small upgrade.

 

[LTC8K6]

https://en.wikipedia.org/wiki/Goldmont_Plus#Design

 

[you2]

oops....

 

[evident]

looks promising

 

[Brunnis]

I have a NUC7PJYH (Pentium J5005) on order now. If everything goes according to plan, I'll be testing it with dual channel DDR4-2400 14-14-14 and comparing it to my old Pentium J4205 results. It will probably take a couple of weeks until I have results up, though.

 

[Bouowmx]

Thank you.

I'd also like to know performance using single-channel memory. Even with single-channel, the Goldmont Plus has more memory bandwidth than Kentsfield Core 2 Quad with dual-channel DDR2:

• 2-channel DDR2-1066: 17.1 GB/s
• 1-channel DDR4-2400: 19.2 GB/s

 

[IntelUser2000]

OpenGL: 21.19 FPS
ST: 73
MT: 263


MT 193

The CPU part of the Gemini Lake Celeron is ~40% faster than the Apollo Lake Pentium! It's also interesting that power consumption does not appear to have increased, at least on the system level. From what I can gather this is slightly better than the venerable old Core 2 Quad Q6600 @ stock. Given that they both run at basically the same frequencies, it seems Goldmont Plus has slightly better IPC than Conroe/Kentsfield.

The important part in responsiveness is always with single thread. Results also exist with the fastest J5005.

https://www.cpu-monkey.com/en/compare_cpu-intel_celeron_j4105-841-vs-intel_pentium_j5005-844

With Cinebench R15, the Celeron J4105 performs close to Q9650 and Pentium Silver J5005 beats Q9650 by 8%.

https://www.techspot.com/article/1039-ten-years-intel-cpu-compared/page2.html

Results show that the J5005,

-Comes within 20% of the Core M3 chips in single thread
-Is on par with 7200U
-Due to 4 cores it also beats those two significantly

Geekbench seems to somewhat understate the gains?

One more such jump like this and they can close the gap with Skylake derivatives.

Edit: Updated comparisons against Q9650

 

[Brunnis]

The Cinebench results mostly point towards a 40-50% IPC increase, which is almost strangely high... I'm going to temper my expectations and prepare for not all tests showing the same gigantic increases. Going by the one available Passmark result, single thread performance is up 30% from my own Passmark result of my J4205. Pretty good, but indicates an IPC increase of ~20%, which "feels" more realistic.

 

[IntelUser2000]

The Cinebench results mostly point towards a 40-50% IPC increase, which is almost strangely high... I'm going to temper my expectations and prepare for not all tests showing the same gigantic increases. Going by the one available Passmark result, single thread performance is up 30% from my own Passmark result of my J4205. Pretty good, but indicates an IPC increase of ~20%, which "feels" more realistic.

So looking at your scores for J4205, it gets 0.68 points in ST Cinebench R11.5. On cpumonkey, J4205 gets 0.61. I felt the cpumonkey result may be on the low side. If you take 0.68 points as J4205 score, normalizing it to clocks we get something like 35% improvement. It does better in R15, because you also got the same R15 score. On there, J5005 is more than 45% faster per clock.

20% may be on the low side, and 30% may be more realistic.

Rendering applications like Cinebench tend to have more ILP than general purpose code, so widening the core can gain more than average compared to typical applications like desktop usage benchmarks and responsiveness. Still, nowadays there's enough rendering-like code that we can't discount them entirely. You can also go the other extreme where the code is so old that the gains may be in the range of 10%. That's not a good representation of current usage scenario either.

Goldmont Plus also has Radix-1024 divider which probably helps Cinebench as it has to do with general purpose SIMD. R15 is newer than R11.5 which may show greater advantages on a newer architecture. R11.5 may be a better "average" benchmark than R15 because using older code, it represents general un-optimized applications.

Comparisons against more recent architectures like Sandy Bridge based and even Kabylake shows that there's a greater gap between them and J5005 in R15 as opposed to using R11.5. Reinforcing my beliefs above.

 

[dark zero]

Seeing how Intel is doing Goldmont +, makes me wonder why they didn't continued with the mobile project.... it has ARM A75 performance....