AMD is already designing products on the 14nm node

[patrickjp93]

Other then iGPU which AMD is fully capable of, Intel isn't going to improve much on the CPU side in skylake and even less in cannonlake. Yeah there will be some improvements but Intel is mainly focused on Efficency and the IGPU now and not CPU performance.

Not true. Intel is seeing some returning competition from IBM's Power8/9 chips, and Skylake needs a strong IPC improvement to keep ahead.

[Silent_Xaxal]

It's a server/supercomputer space only deal right now. That said, innovations there trickle down to consumer hardware eventually.

 

All that considered, what would be the likelihood of IBM competing in the consumer space again?

[vm'N]

Everything is lined up in AMD's favor to once again go for the performance throne.

What?

I haven't seen anything yet, that would confirm AMD would go after the performance throne.

AMD have already mentioned how they would avoid directly competing against Intel, and how they have been "over-concentration in consumer PCs".

[Homicidium]

What?

I haven't seen anything yet, that would confirm AMD would go after the performance throne.

AMD have already mentioned how they would avoid directly competing against Intel, and how they have been "over-concentration in consumer PCs".

 

I probably guess that he was talking more about the GPU's than the CPU's, however with the advancements that are currently leaked could indicate some very good performance gains and a possible reentering into the enthusiasts market again. I certainly hope so!

[Faa]

Intel is seeing pressure from the Power 8 and Power 9 chips from IBM, 6 and 12-core 5GHz beasts that they are.

A 6 core Power 8 matches a Xeon 15 core, their IPC is miles ahead of Intels but they consume so much power and are significantly larger. A 12 core power 8 would perform as good as 40-50 FX cores.

 

 

Top that with a 14nm FinFET design and even the biggest Intel fanboy has to admit AMD could come back swinging very hard. 

Lets be a bit realistic; Bulldozer took them 6 years and it literally couldn't compete with Conroe an architecture from what 2005 in terms of IPC/corecount-vs-corecount with a 1-2GHz higher clock speed. Stepping down to a lower process when Intel would be on 14nm as well just means nothing, BD (32nm) couldn't even hold a candle against sandy bridge.

They have outperform Conroe, Nehalem, Sandy Bridge, Ivy Bridge, Haswell and Skylake in what 2-3 years time including development? 

 

 

i feel sort of bad for saying this but its sort of the fault of tech reviewers like linus saying oh its an amazing price to performance card and you cant get better anywhere else

He said that when the 970 was launched, back then it was 50$. You have to admit, AMD doesn't always have the best price/performance. 

[Opcode]

What?

I haven't seen anything yet, that would confirm AMD would go after the performance throne.

AMD have already mentioned how they would avoid directly competing against Intel, and how they have been "over-concentration in consumer PCs".

AMD won't and cannot make such claims as they don't have working silicon in hand.

 

Lets be a bit realistic; Bulldozer took them 6 years and it literally couldn't compete with Conroe an architecture from what 2005 in terms of IPC/corecount-vs-corecount with a 1-2GHz higher clock speed. Stepping down to a lower process when Intel would be on 14nm as well just means nothing, BD (32nm) couldn't even hold a candle against sandy bridge.

They have outperform Conroe, Nehalem, Sandy Bridge, Ivy Bridge, Haswell and Skylake in what 2-3 years time including development? 

What has to be understood is Zen is an entirely new architecture developed from scratch by Jim Keller. With someone who's reputable for developing some of the fastest architectures around while coupled with being on the same node as Intel. AMD has a shot at bringing more performance at similar or less power consumption. Especially factoring in the power improvement technologies that AMD has (resonant clock mesh). Will Zen out perform it's competitor Skylake? We don't know but they do have everything working in their favor to give them the chance to. Even being within a ~10% margin of Skylake performance will be a huge win for AMD.

[SSOB]

Cool, finally some movement.

[Faa]

AMD won't and cannot make such claims as they don't have working silicon in hand.

Lol the manufacturing process is not the reason why they can't make such claims or limit them from taking the performance throne. You don't need 22nm or lower to outperform Haswell.

[patrickjp93]

Lol the manufacturing process is not the reason why they can't make such claims or limit them from taking the performance throne. You don't need 22nm or lower to outperform Haswell.

IBM does. Their Power8 architecture barely beats Haswell and is on the 22nm FinFET process.

[Faa]

IBM does. Their Power8 architecture barely beats Haswell and is on the 22nm FinFET process.

Lol? 

Intel CPU's are nowhere in the league of IBM in terms of singlethreaded performance just nowhere. If you're denying that then you're just a fanboy. IBM just doesn't have the market share Intel does because it's all about performance per watt for datacenters.

[Opcode]

Lol the manufacturing process is not the reason why they can't make such claims or limit them from taking the performance throne. You don't need 22nm or lower to outperform Haswell.

What does process node have to do with performance at all? To fill you in on some facts the architecture is responsible entirely for performance. AMD could out perform Haswell on 32nm if they came out with a good enough architecture. The process node is semi-critical for power consumption and while being on the same node it may help AMD's new architecture compete in a performance per watt perspective as well.

[Faa]

What does process node have to do with performance at all? To fill you in on some facts the architecture is responsible entirely for performance. AMD could out perform Haswell on 32nm if they came out with a good enough architecture. 

Thats what I'm saying you tool with "you dont need a 22nm or lower CPU to outperform Haswell". Do you ever read people's comment you quote? You quote people and you repeat the exact same thing, saw you doing that a lot.

 

 

I was implying if Carrizo was built on 14nm FinFET it would been a strong competitor to Intel's Broadwell in performance per watt across the board. AMD has a bright future especially with the two massive guns under their belt both Zen and K12 being developed by and in partial with the legendary Jim Keller. Top that with a 14nm FinFET design and even the biggest Intel fanboy has to admit AMD could come back swinging very hard. Everything is lined up in AMD's favor to once again go for the performance throne. 

 

You gave us two reasons why AMD can take the performance throne; Jim Keller & 14nm. So I assumed you thought a dieshrink gives a performance boost.

 

[Opcode]

Thats what I'm saying you tool with "you dont need a 22nm or lower CPU to outperform Haswell". Do you ever read people's comment you quote? You quote people and you repeat the exact same thing, saw you doing that a lot.

I'm only correcting you about what you think you know. Somewhere along your assumptions you dug yourself a hole.

 

You gave us two reasons why AMD can take the performance throne; Jim Keller & 14nm. So I assumed you thought a dieshrink gives a performance boost.

Die shrinks bring performance per watt. I even made it specifically clear in the post you've quoted.

[patrickjp93]

Intel CPU's are nowhere in the league of IBM in terms of singlethreaded performance just nowhere. If you're denying that then you're just a fanboy. IBM just doesn't have the market share Intel does because it's all about performance per watt for datacenters.

You do realize before Power8 IBM was trailing Intel by miles, right? Also, quote flagship to flagship. Currently you're running Intel's bottom of the barrel Ivy Bridge against IBM's flagship of the current generation which is meant to compete against Haswell's flagship.

[vm'N]

Lol? 

Intel CPU's are nowhere in the league of IBM in terms of singlethreaded performance just nowhere. If you're denying that then you're just a fanboy. IBM just doesn't have the market share Intel does because it's all about performance per watt for datacenters.

Power8's advantage is not single-threaded performance.

Power8's great advantage is how well it scales with SMT8.

Now, if you meant core performance that is something different..

[patrickjp93]

Power8's advantage is not single-threaded performance.

Power8's great advantage is how well it scales with SMT8.

Now, if you meant core performance that is something different..

And as I pointed out before he's comparing the lowest clocked IB Xeon against IBM's flagship at 5GHz. I think Skylake will be interesting solely because IBm looks to be making a comeback.

[vm'N]

And as I pointed out before he's comparing the lowest clocked IB Xeon against IBM's flagship at 5GHz. I think Skylake will be interesting solely because IBm looks to be making a comeback.

The power-architecture seems to go after some very specific markets, where the extreme parallel processing power can be utilized effectively.

Skylake will increase the data-parallism, but I doubt they would go after the processing parallism that the power-architecture strive for.

[patrickjp93]

The power-architecture seems to go after some very specific markets, where the extreme parallel processing power can be utilized effectively.

Skylake will increase the data-parallism, but I doubt they would go after the processing parallism that the power-architecture strive for.

Forgive me but what is the difference to you in terms of processing and data parallelism. The whole point of data parallelism is to process multiple independent data simultaneously.

[vm'N]

Forgive me but what is the difference to you in terms of processing and data parallelism. The whole point of data parallelism is to process multiple independent data simultaneously.

Data-parallelism would be using the vector, where processing parallelism (or task-parallelism) would be scaling to more threads.

[vm'N]

nvm

[Thony]

Hmm already

Intel will release their 10nm chips in late 2016-mid 2017 probably.

Can't we just have a look into what's gonna happen by 2020 instead of waiting ?! Time machine please !

[patrickjp93]

Data-parallelism would be using the vector, where processing parallelism (or task-parallelism) would be scaling to more threads.

Except Intel has the advantage here too offering 15-core chips with hyperthreading. Soon we'll have 18-core chips. Power 8 maxes out at 12, though I'm sure that 5Ghz clock helps.

[vm'N]

Except Intel has the advantage here too offering 15-core chips with hyperthreading. Soon we'll have 18-core chips. Power 8 maxes out at 12, though I'm sure that 5Ghz clock helps.

Except that Power8 feature SMT8, where Intel only feature SMT2.

And considering how well the power8 scales with threads, gives them the advantage.

As in Faa's example, the power system have 192 threads, where the Intel system "only" have 48 threads, giving the same amount of cores.

[patrickjp93]

Except that Power8 feature SMT8, where Intel only feature SMT2.

And considering how well the power8 scales with threads, gives them the advantage.

As in Faa's example, the power system have 192 threads, where the Intel system "only" have 48 threads, giving the same amount of cores.

Yeah but without a copious number of ALUs and FPUs, those extra threads will just be stuck idling. I can't imagine the hardware-level context switching to be able to actually keep those 8 threads running per core. Also, 8x12 is 96. Where are you getting 192 from?

[vm'N]

Yeah but without a copious number of ALUs and FPUs, those extra threads will just be stuck idling. I can't imagine the hardware-level context switching to be able to actually keep those 8 threads running per core. Also, 8x12 is 96. Where are you getting 192 from?

Its backend is so wide, that it will be able to dispatch instruction to various units.

They just have to control the instruction flow of the various threads running.

Faa's example is "2s/24c/192t"

---

However the individual instruction flow will have limited throughput after a certain amount of threads.

Edited January 26, 2015 by vm'N