The Intel Atom Thread

[amd6502]

Yes that IPC is shockingly high on these atoms.

https://www.theregister.co.uk/2017/12/11/intel_beefs_up_lowend_line_with_gemini_lake_cpus/

From the GB4 numbers it looks like these atom quads might be dual channel memory; if so, better performance at the slight expense of battery life if running multiple sticks of memory. Stoney and Mullins/Beema only are single channel.

At 6W Stoney (E2-9000e) looses badly:
http://browser.geekbench.com/v4/cpu/compare/983142?baseline=5315637

At 10W Stoney (A9-9400) looses badly only in multicore:
http://browser.geekbench.com/v4/cpu/compare/3734696?baseline=5315637

 

[bakyt115]

Yes that IPC is shockingly high on these atoms.

https://www.theregister.co.uk/2017/12/11/intel_beefs_up_lowend_line_with_gemini_lake_cpus/

From the GB4 numbers it looks like these atom quads have dual channel memory; better performance at the slight expense of battery life if running multiple sticks of memory. Stoney and Mullins/Beema only are single channel.

also new atom has 2 more decoders.

 

[amd6502]

also new atom has 2 more decoders.

I'm not so sure anymore if the atom results are from dual channel, maybe GB is picking up transient bandwidth from the hefty 4MB L2 cache. Multicore bandwidth of 13.5 GB/sec seems very large; single core is similar to Stoney result, ~7.5.

 

[IntelUser2000]

From the GB4 numbers it looks like these atom quads might be dual channel memory; if so, better performance at the slight expense of battery life if running multiple sticks of memory. Stoney and Mullins/Beema only are single channel.

Uh, they had dual channel support since the Bay Trail days back in 2013.

 

[jpiniero]

Is Atom too at the end of life? (I heard Xeon Phi is discontinued).

Could be... Seems like Intel isn't going to be doing much to the mainstream Core anymore beyond shrinks, and at 10 and especially 7 would start to get really tiny. Probably not worth it at some point to continue to do the R&D on Atom cores.

 

[IntelUser2000]

Could be... Seems like Intel isn't going to be doing much to the mainstream Core anymore beyond shrinks, and at 10 and especially 7 would start to get really tiny. Probably not worth it at some point to continue to do the R&D on Atom cores.

Of course they'll continue. They use the cores on the cable modems, and on server. The name Atom is still used in the embedded space, and server chips: https://ark.intel.com/products/codename/80644/Apollo-Lake

It's only gone on the consumer side.

Weren't they also saying only certain versions of the Xeon Phi were cancelled?

 

[jpiniero]

Weren't they also saying only certain versions of the Xeon Phi were cancelled?

The Icelake-SP Phi replaces the high DP Phi; but the DL one is probably going to get replaced by Nervana anyway.

 

[dark zero]

Eh? I think you misread that, 3391 > 2058...

Oh... i see that.... Still, seems that Intel has the edge now.

Wondering if AMD delivers now a refresh or a node shrink to Stoney Ridge or Mullins to 12 nm... With better GPU.

Also... Seems that Spreadtrum made a mistake... They should went with Goldmont which is ARM A72 levels directly instead of Airmont which is near A57 levels...

 

[amd6502]

Wondering if AMD delivers now a refresh or a node shrink to Stoney Ridge or Mullins to 12 nm... With better GPU.

Probably not. It very much looks like Puma and Excavator are at the end of life. Nosta Seronx made some good points about why a 22FDX Stoney+Mullins hybrid would be worth it. 22FDX could also later be shrunk to 12FDX. But with how cheap Atoms have been priced historically, AMD will probably pass on such a project.

I think they should, because this market has great potential for ultra small form factor desktops, as well as the whole low end laptop market (including eMMC netbooks and tablets), where low power usage is the top priority (above performance in priority). https://www.theinquirer.net/inquire...heerful-gemini-lake-celeron-and-pentium-chips

 

[Arachnotronic]

Probably not. It very much looks like Puma and Excavator are at the end of life. Nosta Seronx made some good points about why a 22FDX Stoney+Mullins hybrid would be worth it. 22FDX could also later be shrunk to 12FDX. But with how cheap Atoms have been priced historically, AMD will probably pass on such a project.

After Stoney Ridge Refresh, it looks like the AMD ultra-cheap roadmap goes dark:

 

[Dayman1225]

Added some more microarchitecture information on Gemini Lake from Intel, as well as confirmation that this is 14nm, not 14+/14++. - AnandTech


Intel Launches New Pentium Silver and Celeron Atom Processors: Gemini Lake is Here:

After a few back and forth emails with Intel, we can confirm that the Gemini Lake family is built on 14nm, rather than 14+ or 14++ which are newer variations of Intel's 14nm process technologies. Intel also confirmed the microarchitecture is a three-wide decode, the same as Goldmont, with the following execution port layout:

There are three integer ALUs, one jump port, one load port, one store port, and two general purpose FP/vector ports, so we can dispatch up to 8 instructions in a single cycle.
We're requesting more information.

 

[IntelUser2000]

The Icelake-SP Phi replaces the high DP Phi; but the DL one is probably going to get replaced by Nervana anyway.

Yes, I know this. Their wording implied not all Xeon Phi chips were cancelled.

Added some more microarchitecture information on Gemini Lake from Intel, as well as confirmation that this is 14nm, not 14+/14++. - AnandTech

Some are also saying its still a 3-wide issue core, just that its expanded elsewhere. Its the end performance that matters, so really, it doesn't matter.

 

[IntelUser2000]

Could be... Seems like Intel isn't going to be doing much to the mainstream Core anymore beyond shrinks, and at 10 and especially 7 would start to get really tiny. Probably not worth it at some point to continue to do the R&D on Atom cores.

Another reason Atom(or more properly small cores) investment should continue is because it defends their PC space. Using x86 translation with Windows 10 and ARM might turn out better than Windows RT, but its the Atom-based solutions that will likely turn out to be the best defense.

x86 on Snapdragon 835 makes it quite a bit slower than native ARM performance, meaning Apollo Lake devices are 30-50% faster. https://www.notebookcheck.net/Asus-...petition-of-the-Netbook-Dilemma.271884.0.html

Translation of an ISA is way better than emulation, but won't make it perfect in terms of compatibility. Compatibility problems exist even among different Windows OS versions. It's merely stepping stone for native porting of applications to happen.

Microsoft tends to forget their OSes are also the problem. If they decided to adopt the good parts of Android and iOS and added tricks of their own, they might have had a chance. But by wanting to do it completely differently, almost for the sake of it, they go nowhere.

 

[jpiniero]

BTW, didn't notice this but Gemini Lake supports LPDDR4, which no other x86 processor shipping does.

 

[Dayman1225]

BTW, didn't notice this but Gemini Lake supports LPDDR4, which no other x86 processor shipping does.

Apollo Lake supports it too iirc

 

[IntelUser2000]

BTW, didn't notice this but Gemini Lake supports LPDDR4, which no other x86 processor shipping does.

Yes it does. Though as Dayman mentions, Apollo Lake supports it up to LPDDR4-2400.

They are still slower considering. Apple supported LPDDR4-3200 in its first usage of LPDDR4.

 

[Dayman1225]

Yes it does. Though as Dayman mentions, Apollo Lake supports it up to LPDDR4-2400.

They are still slower considering. Apple supported LPDDR4-3200 in its first usage of LPDDR4.

We know Intel has the IP for "5333mbps LPDDR5/4x/4" so I imagine ATOM gets 2400mhz rated because of A: Bottom tier memory controller B: not needing a huge amount of bandwidth

 

[IntelUser2000]

They are simply late due to poor planning. Apple A series were once below Atom. In a year or two, their SoCs will be so ahead it'll take Icelake generation just to catch up to A11.

 

[Nothingness]

Courtesy of https://www.realworldtech.com/forum/?threadid=173524&curpostid=173524

https://software.intel.com/sites/de...4-ia-32-architectures-optimization-manual.pdf

Goldmont+ changes:

• Widen previous generation Atom processor back-end pipeline to 4-wide allocation to 4-wide retire, while maintaining 3-wide fetch and decode pipeline.
• Enhanced branch prediction unit.
• 64KB shared second level pre-decode cache (16KB in Goldmont microarchitecture).
• Larger reservation station and ROB entries to support large out-of-order window.
• Wider integer execution unit. New dedicated JEU port with support for faster branch redirection.
• Radix-1024 floating point divider for fast scalar/packed single, double and extended precision floating point divides.
• Improved AES-NI instruction latency and throughput.
• Larger load and store buffers. Improved store-to-load forwarding latency store data from register.
• Shared instruction and data second level TLB. Paging Cache Enhancements (PxE/ePxE caches).
• Modular system design with four cores sharing up to 4MB L2 cache.
• Support for Read Processor ID (RDP) new instruction.

 

[IntelUser2000]

Courtesy of https://www.realworldtech.com/forum/?threadid=173524&curpostid=173524

https://software.intel.com/sites/de...4-ia-32-architectures-optimization-manual.pdf

Goldmont+ changes:

Nice. Thanks for the information.

Goldmont cores are even faster than expected. Pentium N4200 gets ~1600 on Windows, but gets 1700-1800 with Android. N5000 with Goldmont Plus cores get ~2000 in Windows. We may be able to see 2200 on Android.

If Icelake and successors do not become much faster, are we going to see "Atom" based cores close up to the regular Core chips? We may at least see successors to Goldmont Plus cores be branded into chips that go firmly into Pentium territory, maybe all Pentium and Celeron chips can be using them.

By the way I estimate Goldmont Plus based cores to have perf/clock on par with then much vaunted Athlon 64s. In that aspect, we've come far.

 

[Dayman1225]

Wow they changed it quite a lot, makes we wonder why they called it Goldmont+ and not an entirely new core.

 

[Dayman1225]

Goldmont Plus detailed, large improvements, setting the stage for a 32-core model - WikiChip

 

[NTMBK]

Wow they changed it quite a lot, makes we wonder why they called it Goldmont+ and not an entirely new core.

Meanwhile the exact same Skylake CPU gets three different rebrands.

 

[NTMBK]

Interesting that they moved to Jaguar style quadraplexes. I bet that this CPU would be pretty good for a PS5, if it was strapped to a big Radeon.

EDIT: Though they'd need to add AVX support.

 

[bullzz]

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