Should Intel design a CPU core with a perfomance level in between Core and Atom?

[Exophase]

Why would Atom even need that? only software I know of that at least makes a little bit sense to run on an Atom is Handbrake. But else? Waste of die pace and energy.

Even having an AVX2 implementation that implements 256-bit operations over 2 uops would be of some value. This could still alleviate some decode bottlenecks and a 1 load/cycle gather can be better than the alternative. There wouldn't be too much of an overhead in implementing this.

The other benefit is that if you're doing something like writing intrinsics or using ASM you can write an AVX2 target as your best optimized case and maybe not bother with a separate SSE2 target anymore. As it stands most stuff out now doesn't use auto-dispatch like in ICC so won't target something like AVX2 unless it's done explicitly, and given software incompatibility a lot of people won't bother. So then the CPUs that do support it miss out. Lack of support on Core-based Celeron and Pentiums is the biggest problem, but Atoms aren't kept out of the PC space by any means either.

 

[Insert_Nickname]

Even having an AVX2 implementation that implements 256-bit operations over 2 uops would be of some value. This could still alleviate some decode bottlenecks and a 1 load/cycle gather can be better than the alternative. There wouldn't be too much of an overhead in implementing this.

What is truly ironic is that AMD is doing precisely that with their newer cat cores. Its limited to AVX (no support for AVX2) however.

 

[cbn]

For the E5 Xeons, here are two examples:

E5 2699 v3
E5 2670 v3 (this one is used in Facebook's Open Compute Servers)

E5 2699 v3 is a 18C/36T Haswell E5 Xeon with 145W TDP and 2.3 Ghz base clock (3.6 Ghz 1C turbo).

E5 2670 v3 is a 12C/24T Haswell E5 Xeom with 120W TDP and 2.3 Ghz base clock (3.1 Ghz 1C turbo)

2.3 Ghz (Haswell IPC) for base clock is not much in the way of single thread.....but it is too much for even a max clocked atom cpu core to handle.

For something like a VDI Server, I'll bet they can get more cores powered up at a single thread level of performance equivalent to that 2.3 Ghz Haswell with a smaller core.

A core smaller than Core, but larger than atom.

Sure the 1C turbo single thread performance equivalent to Haswell at 3.6 Ghz (E5-2699 v3) or 3.1 Ghz (E5 2670 v3) won't be there. However, for high core count VDI server (running a lot of OS instances ) I have a hard time imagining this would be a problem. (Disclaimer: I don't manage VDI servers)

 

[Essence_of_War]

Atom cores are not used in Xeon Phi. They may have started with that design but it's not even close after modifications.

http://www.realworldtech.com/knights-landing-details/

I understand what you're trying to say, that the modifications make it a significant deviation from "stock" Atom. I'm aware of many of these modifications, and while I didn't carefully enumerate them in my reply to beginner, I think your reply was pedantic to the point of being even more wrong by claiming that Xeon Phi doesn't use Atom cores.

 

[ElFenix]

Core
Atom
Phi
Quark
Itanium
Various controller cores, such as the xl170

i love how this thread started with an invalid premise, and when pointed out we just continue to argue about whatever it is people want to argue about.

intel doesn't need a midsize core between atom and the laptop/standard desktop core. intel needs to update atom.

 

[cbn]

Core
Atom
Phi
Quark
Itanium
Various controller cores, such as the xl170

i love how this thread started with an invalid premise, and when pointed out we just continue to argue about whatever it is people want to argue about.

intel doesn't need a midsize core between atom and the laptop/standard desktop core. intel needs to update atom.

None of those cores listed by Phynaz is a "medium (x86) core" though (sitting in between the 3.5x single performance gap of the current atom and core designs).

And if Intel updates Atom (making it almost twice as powerful in single thread), what takes the place of atom?

Quark?

So Core, massively improved atom, quark (improved to airmont atom levels)?

 

[NTMBK]

Upgrade Atom to make it into a competent laptop and tablet processor, and stop trying to shove x86 into phones. Make a great chip with ARM cores and an Intel modem on an Intel process.

 

[cbn]

Upgrade Atom to make it into a competent laptop and tablet processor, and stop trying to shove x86 into phones. Make a great chip with ARM cores and an Intel modem on an Intel process.

I think the current atom level of performance is pretty decent, but one problem that still exists (for example) is that Cherry Trail is not quite power efficient enough to use the CPU and GPU at the same time without throttling.

So while making atom more powerful would be a good idea (for various reasons), there still needs to be a very power efficient core to address the problems that already exist in mobile.

I don't see why it needs to be ARM though.

 

[kimmel]

I understand what you're trying to say, that the modifications make it a significant deviation from "stock" Atom. I'm aware of many of these modifications, and while I didn't carefully enumerate them in my reply to beginner, I think your reply was pedantic to the point of being even more wrong by claiming that Xeon Phi doesn't use Atom cores.

Being slightly more specific, there aren't Atom cores in KNL, there are Atom cores in KNC. It's not pedantic. They are virtually indistinguishable after the modifications. I highly doubt even what remains of the common logic has the same layout between KNL and Silvermont.

If I actually wanted to be pedantic I could have said that all CPUs have an ALU so all CPUs are the same CPU core. Or perhaps, all cars and trucks have wheels therefore all cars are trucks.

 

[jpiniero]

Upgrade Atom to make it into a competent laptop and tablet processor, and stop trying to shove x86 into phones. Make a great chip with ARM cores and an Intel modem on an Intel process.

Phones are where the growth is. And they want to put Core in there, not Atom. I won't say it'll work; but one device with a dock that replaces everything else (desktop, laptop, tablet) would be pretty appealing. Problem is that Windows mobile is rather crappy, and only Apple at this point is making money on phones. But when you look at how bad the future looks like for the PC they almost have to try.

 

[cbn]

So Core, atom (massively improved so it becomes a medium core), Quark (the new atom).

Then Morph Core comes later with these three cores as a comparison point?

 

[cbn]

Phones are where the growth is. And they want to put Core in there, not Atom. I won't say it'll work; but one device with a dock that replaces everything else (desktop, laptop, tablet) would be pretty appealing. Problem is that Windows mobile is rather crappy, and only Apple at this point is making money on phones. But when you look at how bad the future looks like for the PC they almost have to try.

If we look at a chip like SoFIA, the atom cores are clocked at only 1.4 Ghz.

So atom still has a ways to go in order to unlock its full potential.

Maybe 10nm helps that. EDIT: 10nm should definitely help that.

But for Core (in a phone), I don't know if 10nm will help so much.

But more pressing than that I think is the high core count E5/E7 servers

 

[ElFenix]

None of those cores listed by Phynaz is a "medium (x86) core" though (sitting in between the 3.5x single performance gap of the current atom and core designs).

right, but that still disproves your assumption in the OP.

further, so far as thermals are concerned, there isn't enough gap to shove another product line in there. the 'big core' laptop/desktop line is designed for lower and lower thermals as it is.

it's atom that needs improvement, not a (7th or 8th) new product line

 

[cbn]

further, so far as thermals are concerned, there isn't enough gap to shove another product line in there. the 'big core' laptop/desktop line is designed for lower and lower thermals as it is.

The gap is huge.

3.5x difference in single thread performance.

And just because Intel is using Core in Core M and 2.3 Ghz E5 Xeons doesn't mean it is ideal.

Example: Broadwell Xeon is supposed to be only 24 cores. This despite 14nm being a 1.5 node jump from 22nm (obviously the xtor budget is there for a much larger increase in cores, but the power is not).

 

[jpiniero]

Example: Broadwell Xeon is supposed to be only 24 cores. This despite 14nm being a 1.5 node jump from 22nm (obviously the xtor budget is there for a much larger increase in cores, but the power is not).

You have to factor in costs and yield when talking about Broadwell Xeon being "only" 24 cores.

 

[Phynaz]

There is no way Intel can design a mobile cpu core that is worth a damn without it cannabalizing their high margin overpriced offerings. So they wont. Atom will remain useless and Core M will remain overpriced. Even if it costs them their business at some point.

You realize that if Core M were overpriced the market would reject them?

You may not like the price, but they do appear to be priced correctly.

 

[Phynaz]

So Core, atom (massively improved so it becomes a medium core), Quark (the new atom).

Then Morph Core comes later with these three cores as a comparison point?

Atom as it stands is already the medium processor. It sits between the large and small x86 cores.

 

[cbn]

You have to factor in costs and yield when talking about Broadwell Xeon being "only" 24 cores.

Haswell EP/EX is 662 mm2, so die size/yields is not the problem.

Power definitely is though.

 

[Exophase]

What is truly ironic is that AMD is doing precisely that with their newer cat cores. Its limited to AVX (no support for AVX2) however.

VIA's latest Nano supports AVX2, and while I don't know for sure that it uses 128-bit units I'd be really amazed if it doesn't.

 

[cbn]

Atom as it stands is already the medium processor. It sits between the large and small x86 cores.

IMO the gap between these cores needs to be smaller.

Say ~2:1 single thread ratio between Core and whatever you want to call the medium core. And ~2:1 single thread ratio between whatever you want to call the medium core and the small core.

This instead of the 3.5x gap between Core and atom.

P.S. Assuming the new mid sized core was designed for a single thread exactly in between the current versions of Core and atom, that would put it somewhat ahead of the base clock single thread of the Haswell Xeons I mentioned earlier in this thread.

 

[Phynaz]

So you really want a 4th x86 core from Intel.

 

[cbn]

Another processor line that I think could benefit from a mid sized Intel x86 core (designed for a single thread in between the current Core and Atom CPUs) would be the Xeon-D line (which has eight Broadwell cores with base clock of 2.0 Ghz and 2.8 Ghz turbo on the top SKU).

 

[cbn]

So you really want a 4th x86 core from Intel.

Maybe they can drop what is presently known as Quark? (which as far as I understand is just a die shrunk Intel x86 processor from ~20 years ago).

Example would be the Intel Edison board (which at one time used Quark, but now uses atom)

The Intel Edison is a tiny computer offered by Intel as a development system for wearable devices. The system was initially announced to be the same size and shape as an SD card and contain a dual-core Intel Quark x86 CPU at 400 MHz communicating via Bluetooth and Wi-Fi. A later announcement changed the CPU to a 22 nm Silvermont dual-core Intel Atom CPU.

So that would get us back to three x86 cores.....but three spaced closer together.

The purpose of this would be to fit modern niches better.

With that mentioned, I do admit Intel Curie (which uses quark) sounds like it could make for some very interesting devices:

 

[Thala]

With that mentioned, I do admit Intel Curie (which uses quark) sounds like it could make for some very interesting devices

Why do you think so? Ever compared Quark to likes of Cortex-A5 or Cortex-A7? With the ARM CPUs you will get a full fledged MMU, optionally NEON, Snoop Control Unit and caches supporting MOESI protocol (for Multi-Core) and more with better perf/power ratio.
I am sure Quark will be discontinued for not beeing competitive.
This happens while ARM is upgrading their Cortex-R CPU line with multi-core capabilities and in some cases with MMUs.

 

[Essence_of_War]

Another processor line that I think could benefit from a mid sized Intel x86 core (designed for a single thread in between the current Core and Atom CPUs) would be the Xeon-D line (which has eight Broadwell cores with base clock of 2.0 Ghz and 2.8 Ghz turbo on the top SKU).

I don't understand what Xeon-d is missing that you think a new core arch would solve. It already crushes all x86 and arm offerings in its thermal envelope. It appears to be near the ideal perf/watt for broadwell. Why would we expect a new arch to improve on this?