AMD CPU speculation... and expert conjecture

[Cazalan]

All of which supports the idea they are going to work out the bugs in steamroller before they launch a half baked solution.

And even with a small share, its a growing market with high margins. It seems unlikely they would drop it so easily. Especially since the Bulldozer architecture was built with servers in mind. It doesn't even require huge investments since so much of the desktop tech is fairly easily ported.

AMD shed 30% of their employees so something had to give. They returned to profitability through high end APU sales. Servers may have higher margins but they also have higher costs at every level. Design, fabrication, testing, platform support. It's rather difficult for a 3Billion company to compete with the likes of Intel, IBM, Oracle at the server level.

Besides for us consumers isn't it better they're focusing more intently on gaming?

 

[FALC0N]

Cazalan, all the more reason for them to wait for steamroller to mature.

 

[szatkus]

The reason I believe the delay theory is no x86 server upgrade on the current roadmap either. I might be willing to buy AMD giving up on the high end desktop market, but the lucrative server market also? I doubt that. And the development of those two parts is closely related. The roadmap I saw also says piledriver on am3 and server through 2015. It doesn't exclude a die shrunk piledriver as a stopgap.

Why shrink Piledriver when they already have Steamroller and probably Excavator?

 

[FALC0N]

Might be cost savings reusing the Richland tooling. I doubt it, but they didn't rule it out.

 

[jdwii]

True, jdwii. But the management of AMD and the market as a whole have changed in that time.

And need I point out that the "rush the server and enthusiast parts out first" strategy did not go well for either Barcelona or bulldozer. Perhaps a lesson learned?

Well in that case it was more about the actual design flaw vs anything else. Not to mention that is why BD was skipped for the APU. As for the phenom 1 issue well that would of happened regardless if they made a single core vs a quad core. If anything Amd seems to be releasing the APU first and then the CPU. Which still makes me think we will see a future 8 core processor from Amd if they feel like the improvement was worth it.

Unless again Amd wants their processors to have less features when their compared to Intel. With the 8350fx you get close performance compared to the I5 in current games but faster multithreading performance its a good choice for some.
Plus Amd always has to offer more or at least different features per dollar that is the only way they will ever survive with Intel's bigger name.

However for the masses the need Performance per watt and performance per dollar to be key and that is where 80% of the market is. However the high-end PC market is growing not shrinking which means Amd still should make a higher performing CPU for those customers.

Ask your self how many parts do they need to sell to make a profit? Is that profit and marketing(if the product is good) worth the overall cost. I say yes, yes it is.

Who control's the prices anyways in the CPU market? Intel! Why? They have higher performing parts so Amd has to position themselves to remain competitive. Plus their APU's and CPU's are considerably bigger in terms of die space and what their performance offers from that die space which equal's cost.
Not only does Amd have to position themselves in the lower-end market(less money per product) they also have to deal with bigger die size making them more money to make then Intel!

Look at a sandy bridge CPU die space(also built on 32nm) and compare that to PD and then look at performance per die space its not good.

 

[de5_Roy]

calxeda ran out of moniez T_T
http://vr-zone.com/articles/chip-maker-calxeda-shutters/67439.html
now amd has a better chance with seattle. hopefully calxeda won't be bought out by qualcomm. or google.

 

[juanrga]

Not faster, just scheduled differently. Essentially, Windows treats the CMT cores as HTT cores, using them LAST, resulting in more aggressive turbo clocks.

Scheduled differently to run faster by optimizing resources: 5-15%

I also note that under full load, since all the cores are used anyway, you wouldn't see any performance improvements out of this. Likewise, single-core loads use only a single core, and thus the performance penalty of CMT doesn't occur. So its unlikely that this type of benchmarking would be affected by the Windows scheduler. Nice try though.

Full load also benefits from the new schedulers which maps interdependent threads to cores in the same module, eliminating performance penalties from moving L2 data intermodule.

And that the benchmark was single core doesn't mean that only one core is being used. The OS and background tasks continue running.

If AMD re-designed how its memory subsystem works for Kaveri, I'm relatively sure we'd know about it by now. Your totally BSing on this one.

AMD already disclosed that Kaveri comes with hUMA, which requires a new IMC. Moreover, I know that the L2 cache is ~20% faster now. I will reveal the source for the claim in the update of my Kaveri article. I am soliciting permission.

And the same as intel CPU's, VIA CPU's, and ARM CPU's.

Not exactly the same. Intel chips have a better IMC and extract more bandwidth from 1600MHz, thus when upgraded to 2133MHz you don't see real performance gains in ordinary workloads. This is not the case with FX/Trinity. Also memory bandwidth requirements for a weak CPU are not the same that for a fast CPU. The faster CPU needs to be feed faster as well.

So...in Memory benchmarks, using faster RAM increases scores? Wow, what a concept.

Therefore you believe that Intel is putting faster 2400MHz memory in the new Broadwell-EP Xeons, because servers are running "Memory benchmarks" 24/7. LOL

Your numbers are WAY off.

JPEG Decrompress Multicore was a 31.75662219% increase in performance compared to Trinity, one of AMD's best increases.

Kaveri ES was 47% faster (IPC) than Trinity in JPEG Decompress Multicore. I am not sure why you continue in denial mode.

And those ES numbers will improve with final Kaveri silicon (which has higher clocks) running on W8.1 and with stock memory (rumoured ot be 2400MHz).

 

[juanrga]

Funny to see the same discussion over and over again in this thread.

AGAIN:
======

As I said before, AMD is focusing in dual/quad-core APUs because that cover >90% of the market. Chips such as the FX-8350 have only a market share among gamers 0.3% or something as that.

I showed that FX are not selling well. People here denied the data. A pair of days ago AMD reported (link given before) that FM2 platform accounts for more than 80% of incomes. Now split the remaining <20% between rest of platforms and you can see that AM3+ is minority product. This explain why AMD is not upgrading the platform and all the exciting stuff is coming to the new FM2+.

I said no Steamroller FX for 2014. People didn't believe it and said me "wait for the desktop roadmap".

The 2014 roadmap said no Steamroller FX for 2014. People didn't believe it and some said "wait for the update of the roadmap", whereas others said that it was "not official roadmap".

The 2015 roadmap said no Steamroller FX for 2015. People didn't believe it and said it was "fake".

The 2015 roadmap has been confirmed and says no Steamroller FX for 2015. People ignores it and continue speculating about if AMD will release FX Steamroller 8 cores in 2014. LOL

Servers. More of the same. AMD has said very clearly that ARM will win over x86. Therefore they are migrating to ARM. AMD has said very clearly that Warsaw is released for customers that will be slow on migrating to ARM servers. Warsaw is an upgrade path to former Opteron customers.

AMD is moving from (CPU+dGPU) to (APU+dGPU) and finally to only APU solution.

AMD has plans for an ultra-high performance APU with 8 cores, stacked RAM, and about 10TFLOP performance. CPUs and dGPUs will disappear in some few years (five?) by the simple reason that there is no way you can sustain that high level of performance using an old CPU+dGPU configuration. The laws of physics say so. AMD know this. Intel knows this. Nvidia know this. Self-proclaimed experts and engineers here ignore the laws of physics.

Heterogeneous System Architecture. I wonder why people still doubt that HSA will not succeed. HSA is not AMD thing only. HSA is an "HSA foundation" thing, which is founded by some of the bigger players in the industry: Samsung, Qualcomm, ARM,...

Not only HSA software will be ready for Kaveri launch, but benchmark suites with HSA enabled will be ready as well.

Intel will be fighting HSA with its "neo-heterogeneous" approach. And Nvidia will be fighting with its own heterogeneous ARM+CUDA approach and also with POWER+CUDA in the OpenPower Consortium.

Therefore everyone: Intel, Nvidia AMD, ARM, Samsung, Qualcomm, IBM... knows that the future is heterogeneous. What is so difficult to accept?

 

[szatkus]

With these results you can prove anything.

http://browser.primatelabs.com/geekbench3/compare/264788?baseline=223722
http://browser.primatelabs.com/geekbench3/compare/279514?baseline=223722

Actually Kaveri is regress:
http://browser.primatelabs.com/geekbench3/compare/96279?baseline=223722

Whatever you want.

 

[Ranth]

Juan
"Self-proclaimed experts and engineers here ignore the laws of physics."......
Okay.... let's see:
"CPUs and dGPUs will disappear in some few years (five?)"
So we are gonna get ~8 cores and next gen high end graphics onto a smaller die and keeping the tdp at what ~100w? And at the same time keeping the same performance? Call me impressed!

 

[gamerk316]

calxeda ran out of moniez T_T
http://vr-zone.com/articles/chip-maker-calxeda-shutters/67439.html
now amd has a better chance with seattle. hopefully calxeda won't be bought out by qualcomm. or google.

Somehow, not shocking to me. Should be interesting to see who picks up the pieces.

 

[jdwii]

"As I said before, AMD is focusing in dual/quad-core APUs because that cover >90% of the market. Chips such as the FX-8350 have only a market share among gamers 0.3% or something as that."

Show us how many people buy a I5 or I7 and come back and tell us that market share. Then come back and explain to me why Amd made a 290X.

 

[gamerk316]

"As I said before, AMD is focusing in dual/quad-core APUs because that cover >90% of the market. Chips such as the FX-8350 have only a market share among gamers 0.3% or something as that."

Show us how many people buy a I5 or I7 and come back and tell us that market share. Then come back and explain to me why Amd made a 290X.

They made cards like the 290x because the margins are HUGE. For APU's, the margins are TINY. So even though the market for cards like the 290x is small, they account for a great deal of the total profit, due to the sky high margins they can charge.

Switching targets, I think the folding of Calxeda should be the final proof that customers don't want ARM servers. Hopefully, AMD gets the message...

 

[juanrga]

Since you ignore what I write and ask me for stuff that I said more than once, I will ignore your questions.

Here a new 2015 roadmap updated with AMD new processors

Some people is speculating that the new cores are unnamed because AMD is studying the switch to ARM, because in 2015 the cross-licensing of x86 with Intel finishes. Don't take that too seriously, not still at least ;-)

 

[noob2222]

^^ so what your implying is AMD is leaving DT to intel and GPUs to nvidia to compete with samsung, apple, qualcomm and the rest of the ARM gang?

 

[juanrga]

Some screenshoots of Haswell i3. Our local gamerguru predicted the i3 will be faster than quad, hexa, and octocores because games cannot scale well beyond two treads (he said us). Look at this

hum no wait, because here Intel quad cores are faster contrary to what he said us

and here the AMD octocore is faster than any i3/i5

the same than here

BF4 and posterior games will increase the gap between octo, hexa, quad cores and i3s.

Some non-gaming benchmarks, because PCs can be used for more than gaming

 

[de5_Roy]

Switching targets, I think the folding of Calxeda should be the final proof that customers don't want ARM servers. Hopefully, AMD gets the message...

32-bit arm servers, specifically. amd will be making socs with 64-bit cores, skipping 32bit entirely. i think customers want to see how 64bit pans out before spending their cash.

 

[de5_Roy]

@juanrga: why do your so-called amd roadmap have poorly edited out (i bet it was done with ms paint) 'hsa foundation'? suspicious, to say the least. the edits are not visible unless you zoom in. i am not saying if amd won't release new fx cpus in 2015, just the poorly executed edits are making me question it's worth.

@noob: if amd's x86 cross-licensing does end in 2015 or around that time, amd has very small amount of time to come up with a rock solid range of cpu/soc products for post-x86 life. without their cpu/soc dept., they'll have only the gfx division, which they neglect and abuse rather frequently. even in gfx, they'd have to grab feasible amount of marketshare in high-margin segments like the ones r9 290x is battling in, and in pro/workstation and servers. niche segments like 64-bit arm microservers have tons of untapped potential but that segment is now non-existent and it's marketshare will be up for grabs right the moment it comes into being. what's worse, nothing's stopping big bad intel from buying an arm v8 license (not core license) and compete with others.

 

[noob2222]

@juan

Look where richland is in all of those charts.

 

[Cazalan]

Switching targets, I think the folding of Calxeda should be the final proof that customers don't want ARM servers. Hopefully, AMD gets the message...

32-bit arm servers, specifically. amd will be making socs with 64-bit cores, skipping 32bit entirely. i think customers want to see how 64bit pans out before spending their cash.

It also shows that converting people isn't easy. The CPU is just one aspect of a sever ecosystem. The virtualized storage arrays are a big aspect. All that software has to be ported.

 

[Cazalan]

@noob: if amd's x86 cross-licensing does end in 2015 or around that time, amd has very small amount of time to come up with a rock solid range of cpu/soc products for post-x86 life.

Just a technicality. Intel has to renew or they would lose X86_64 capability.

 

[de5_Roy]

It also shows that converting people isn't easy. The CPU is just one aspect of a sever ecosystem. The virtualized storage arrays are a big aspect. All that software has to be ported.

definitely.

Just a technicality. Intel has to renew or they would lose X86_64 capability.

ah. can't intel one-sidedly renew x86-64?

 

[-Fran-]

ah. can't intel one-sidedly renew x86-64?

It's called "cross licensing" for a reason

The question is another one I'd say... Is x86_64 big enough of a reason for Intel to keep telling AMD to use X86?

At first glance, with Wintels participation in the market being as big as it is, I'd say yes.

Cheers!

 

[FALC0N]

ah. can't intel one-sidedly renew x86-64?

It's called "cross licensing" for a reason

The question is another one I'd say... Is x86_64 big enough of a reason for Intel to keep telling AMD to use X86?

At first glance, with Wintels participation in the market being as big as it is, I'd say yes.

Cheers!

Intel would never risk the type of anti-trust scrutiny that would come with it.

 

[noob2222]

I was actually considering one to replace my llano htpc, but not for $175.