[WCCF] AMD 14/16 nm Opteron and FX Processors with up to 20 Cores in 2016-2017

[SOFTengCOMPelec]

You know, there was a plan of 16-core Seattle. But it was scrapped. God knows why.

Yes, that seems a bit crazy.

Surely it can't cost that much to create an additional mask ?

But I suppose it could be, yield issues and/or they did not see the engineering efforts/costs as being worth it.
Such a pity, because it could potentially compete with the very high core count Intel Xeon-EPs, because they sell for huge amounts of money, so even if it was slower, it should still be able to sell, at a reasonable (profitable) price.

I'm surprised that 16 core Arm chips are not already on the consumer market. In theory a decent Linux operating system could be put together, which for some applications could use all 16 cores, so that it might be faster than an Intel Quad, such as video editing.
Arm cores may not be fast enough for that yet. But sooner or later, they will probably get there.
With current software, it would MOSTLY not usefully utilise the 16 cores, so would be significantly slower than current Intel Quads. But in time, the Linux distribution, may be ale to be improved with software which does utilise all 16 cores, and overcomes this performance barrier.

N.B. This post and my previous post, contradict each other. So I have to agree with my first post, and say that it is VERY problematic for consumer software to usefully use 16 cores, for most things.
But servers would love 16 cores, and use them just fine, with much of today's software.

 

[Enigmoid]

I'm not really impressed with core count. 20 cores sounds good but when you realize that 16 core opteron (6300) generally loses to 8 core SB its important to realize the performance per core matters a lot.

 

[SOFTengCOMPelec]

I'm not really impressed with core count. 20 cores sounds good but when you realize that 16 core opteron (6300) generally loses to 8 core SB its important to realize the performance per core matters a lot.

Yes, the overall "real-life" performance is what really matters.
Because some software mainly uses few or only one core, the single core performance is very important.

A current quad core Haswell cpu, is considerably faster, even in heavily multi-threaded tasks, than the Intel 8 core Atom (C2750), by about a factor of 2 or 3, depending on what software is used.

i.e. it would need at least two C2750's (making 16 cores in total), just to keep up with a 4 core Intel I7-4771 (or similar).

BUT they do use very little power. The C2750 only has a TDP of 20 or 22 Watts (from memory), so it potentially can use less energy than the standard (I7) high clocked Haswells.

 

[pTmdfx]

Such a pity, because it could potentially compete with the very high core count Intel Xeon-EPs, because they sell for huge amounts of money, so even if it was slower, it should still be able to sell, at a reasonable (profitable) price.

This would be really a wishful thinking when those Xeon sells not because only of their high core count, but also the core itself and the platform package associated. Otherwise, wouldn't those many-small-core solutions from Oracle dominant the world already, instead of Xeon's deep root in the mainstream server market?

 

[inf64]

Well, 4Ghz seems a little too high.
From my understanding, high frequency is achieved by sacrificing IPC.

Not true. Prime example: IvyBridge/Haswell generations. Haswell clocks way past 4Ghz (factory clock) while having highest IPC in x86 world.
AMD can do a core that can have Haswell-like IPC and clock close to 4Ghz. It's only a question whether that core will be competitive in 2016 or whenever it launches.

 

[Ajay]

Indeed there is a new x86 core (ARM K12 "sister core") and it will operate closer to ~4Ghz range according to Keller.

Do you have a link to this? I must have missed this interview. THX!

 

[inf64]

Do you have a link to this? I must have missed this interview. THX!

Sure I do. Core innovation update @ YT.
I'm not sure when the "closer to 4Ghz range" remark comes in but 55 minute mark and onward are interesting since Keller mentions what AMD will be doing in both ARM and x86 parts.

 

[pTmdfx]

Sure I do. Core innovation update @ YT.
I'm not sure when the "closer to 4Ghz range" remark comes in but 55 minute mark and onward are interesting since Keller mentions what AMD will be doing in both ARM and x86 parts.

http://www.eetasia.com/ART_8800697976_499489_NT_538687e6.HTM
EETimes

The veteran processor architect [Jim Keller] got "a whole lot of new ideas working with both" ARM and x86 cores. But he gave only a few oblique hints on what AMD's custom chips might look like.

They will be closer in frequency to the 4GHz of AMD's latest Kaveri x86 SoC than today's 2GHz ARM chips, he said. For Skybridge, AMD tweaked its new Puma x86 cores to look more like an ARM V8 core.

But how exactly this indirect statement was backed is uncertain. Moreover, the later statement isn't quite precise with regard to what was really told on the Core Innovation Day event. So I assume there is a chance of this being the author's own digest.

 

[SOFTengCOMPelec]

This would be really a wishful thinking when those Xeon sells not because only of their high core count, but also the core itself and the platform package associated. Otherwise, wouldn't those many-small-core solutions from Oracle dominant the world already, instead of Xeon's deep root in the mainstream server market?

Yes, I agree.
Intel make huge efforts, at creating/maintaining /keeping their immense server cpu lead, over the competition (what little is left, these days).

I suppose it would need a very big player, with a huge budget, to crack the market for real.
AMD who have been somewhat struggling in recent years, and the separate (relatively small, or small divisions of very big companies) Arm suppliers, don't necessarily have the financial muscle and will power to do it.
It would need a (very big) company like Apple, Samsung, or some other business, to make a major marketing decision to go for the highly profitable server cpu market, in a big way.

Even if they did, the Arm cores are not necessarily up to the challenge, as there would still be the issue of the fact that most pre-existing software (with servers) is for X86, and conversion would be a major headache, and a big IPC/clock-speed disadvantage (which might gradually ease in time, as Arm progresses).

 

[Conker10k]

I don't think its gonna be a cut and paste more cores this time for AMD. Most likely they are going to rework the cpu design itself for improved ipc. Have to see what happens. DDR4 will be around also so it will be interesting to see what that will do also.