AMD Kaveri official reviews

[Musho]

7850k with 290x dual graphics thanks to mantle! This could be _very_ interesting. Due to less of a CPU bottleneck (thanks to mantle as well) and dual graphics, this might actually be faster than Intel highend + 290x. Due to the big speed differences of the two GPUs, I'm thinking this isn't using AFR. Check it out! What do you guys think?

http://m.youtube.com/watch?v=Nj0_6Rb...%3DNj0_6RbCHA4

[Hornet331]

No matter how much overhead you remove, a 7850 will always be a bottleneck for a CF high end card setup.

This is pointless as it gets, not that they will show much of a cpu bench anyway, as already shown they run 5400*1920 (aka 10 mega pixels) that even puts dual 290x into sub 60fps regions permanently.

[zalbard]

Right, nothing impressive here

Well, at least AMD keeps improving at a faster rate than Intel.

[Lightman]

How about more native usb3 and much better sata 6gbs performance. The AHCI performance is very weak, not to mention RAID performance. AMD always has buggy chipsets.

Does AMD support RAID trim passthrough?

RAID trim is available on A88X chipsets.

[eRacer]

AMD Board: http://www.scan.co.uk/products/gigab...6%29-vga-dvi-d

Intel Board: http://www.scan.co.uk/products/msi-h...vi-d-hdmi-matx

Pretty clear price difference there.

Why is the AMD board missing HDMI, DisplayPort, a PCIe x16 slot, 2 memory slots and 2 SATAIII ports? No wonder why there is a price difference.

[Piledriver]

let me guess, not one opencl benchmark, you know the reason of APU, HSA, etc... etc...

[paulbagz]

So there's rumours afloat that the Dual Graphics mode will only work with DDR3 based cards.

wut.

-PB

[Darakian]

let me guess, not one opencl benchmark, you know the reason of APU, HSA, etc... etc...

Look up. There are quite a few and it does quite well.

[[XC] Lead Head]

Well, at least AMD keeps improving at a faster rate than Intel.

Not really. Kaveri for most work loads is only marginally faster than piledriver based CPUs. Its only saving grace is HSA workloads which pretty much do not exist right now. Athlon 64 was a success because even if the AMD64 instruction set didn't take off, they had a super fast x86 CPU anyways. If HSA doesn't take off, they're stuck with a pretty low performance CPU. Unless the next revision (excavator is it?) is a complete miracle, AMD should probably just scrap this architecture. Actually, they should just shutter their whole CPU business all together and rename themselves back to ATi - and this is coming from a person who is currently has a PC with an FX-6300 in it.

[xVeinx]

I'd guess the next true architecture "change" would be in 2016 at the soonest. They have excavator in 2015, then something beyond in 2016. I'm not sure how many engineers they have working on the next generation architecture, but they need a lot if they want to continue on the CPU front--unless they intend to switch over to GPU and ARM and call it a day...

[blindbox]

Not really. Kaveri for most work loads is only marginally faster than piledriver based CPUs. Its only saving grace is HSA workloads which pretty much do not exist right now. Athlon 64 was a success because even if the AMD64 instruction set didn't take off, they had a super fast x86 CPU anyways. If HSA doesn't take off, they're stuck with a pretty low performance CPU. Unless the next revision (excavator is it?) is a complete miracle, AMD should probably just scrap this architecture. Actually, they should just shutter their whole CPU business all together and rename themselves back to ATi - and this is coming from a person who is currently has a PC with an FX-6300 in it.

I sometimes wonder what's stopping them from going back to Stars architecture and just keep increasing the clocks on that, and keep evolving it. It had better IPC and better perf/watt. Was there some sort of an architectural wall? Stars architecture had only one single revision. Deneb is simply a die-shrink and Thuban is simply extra cores. There's bound to be things that weren't disclosed to the public but I wonder what would have happened if all that time, 7 years, was spent on improving Stars rather than making bulldozer and improving it.

EDIT: Here's the reason why I'm asking. http://www.hardwarecanucks.com/forum...review-16.html . Steamroller might have improved by 8% compared to piledriver, but that is still not enough to even beat the Stars architecture. And I just checked, Stars is 7 years old. Intel learned something from Netburst. They completely scrapped Netburst when they realized their other architecture is much better.

EDIT2: I should also add that a theoretical 32nm 8-core Stars is smaller than 8-core Piledriver (not so sure about steamroller). Hell, in fact, AMD has MCM tech ready from the servers, they could just slap two 4-core Stars and be done with it.

[demonkevy666]

wasn't llano a stars techinically speaking on 32nm I don't recall a whole lot of those going over 3.7ghz on overclocks.

[blindbox]

wasn't llano a stars techinically speaking on 32nm I don't recall those a whole lot going over 3.7ghz on over clocks.

Well, mainly because there's probably isn't that much architectural improvements, and HSA was in its infancy. It is after all, their first time integrating the processor and the GPU together. The main reason Bulldozer got very high clocks is due to a tech they bought from a company. (sorry, I'm bad with details on this one, and it may, or may not affect anything at all. We don't know if AMD actually integrated that tech into bulldozer. Also, either they licensed the tech, or they bought the company itself. I think it's the latter). I wonder what happens if AMD used that tech to improve Stars instead.

EDIT: Found it. Seems to be licensed. http://www.bit-tech.net/news/hardwar...s-piledriver/1. Seems to be more focused on decreasing power consumption rather than increasing clock speed, as well, so it's unrelated to bulldozer's clock speed. Not to forget, Thuban hits 4 GHz very, very frequently.

[demonkevy666]

Well, mainly because there's probably isn't that much architectural improvements, and HSA was in its infancy. It is after all, their first time integrating the processor and the GPU together. The main reason Bulldozer got very high clocks is due to a tech they bought from a company. (sorry, I'm bad with details on this one, and it may, or may not affect anything at all. We don't know if AMD actually integrated that tech into bulldozer. Also, either they licensed the tech, or they bought the company itself. I think it's the latter). I wonder what happens if AMD used that tech to improve Stars instead.

EDIT: Found it. Seems to be licensed. http://www.bit-tech.net/news/hardwar...s-piledriver/1. Seems to be more focused on decreasing power consumption rather than increasing clock speed, as well, so it's unrelated to bulldozer's clock speed. Not to forget, Thuban hits 4 GHz very, very frequently.

thuban has a low K die electric layer

I do believe it was too late for piledriver to have that implemented

Richland is suppose to have RCM, but I can't find any confirmation on that.

steamroller cores is suppose to have it too,

[lizardmech]

How can they justify releasing more bulldozer based chips? This is the 3rd iteration and it's still behind their previous arch. How many stars cores can you actually fit per bulldozer module? Isn't it something like 2+ entire stars cores per module? Netburst was never even this bad, at least the CPUs still got faster overall with new models years after the initial release.

[[XC] Lead Head]

I sometimes wonder what's stopping them from going back to Stars architecture and just keep increasing the clocks on that, and keep evolving it. It had better IPC and better perf/watt. Was there some sort of an architectural wall? Stars architecture had only one single revision. Deneb is simply a die-shrink and Thuban is simply extra cores. There's bound to be things that weren't disclosed to the public but I wonder what would have happened if all that time, 7 years, was spent on improving Stars rather than making bulldozer and improving it.

EDIT: Here's the reason why I'm asking. http://www.hardwarecanucks.com/forum...review-16.html . Steamroller might have improved by 8% compared to piledriver, but that is still not enough to even beat the Stars architecture. And I just checked, Stars is 7 years old. Intel learned something from Netburst. They completely scrapped Netburst when they realized their other architecture is much better.

EDIT2: I should also add that a theoretical 32nm 8-core Stars is smaller than 8-core Piledriver (not so sure about steamroller). Hell, in fact, AMD has MCM tech ready from the servers, they could just slap two 4-core Stars and be done with it.

Incompetent management, probably. Bulldozer, Piledriver and Steamroller were under development for years before their release. They had years worth of simulation and engineering samples to show that performance was a joke and worse than the previous architecture. In the time they spent wasting with steamroller and piledriver, they could have released an updated K10 core that may have actually been competitive with Intel's current offerings. My buddy has an older A6 1.6GHz Llano laptop. K10Stat let him push that thing to near 2.8GHz stable, and that honestly gave my i7-2630QM laptop purchased around the same time a good run for its money.

[blindbox]

Incompetent management, probably. Bulldozer, Piledriver and Steamroller were under development for years before their release. They had years worth of simulation and engineering samples to show that performance was a joke and worse than the previous architecture. In the time they spent wasting with steamroller and piledriver, they could have released an updated K10 core that may have actually been competitive with Intel's current offerings. My buddy has an older A6 1.6GHz Llano laptop. K10Stat let him push that thing to near 2.8GHz stable, and that honestly gave my i7-2630QM laptop purchased around the same time a good run for its money.

Yes, I would've came into the conclusion of incompetent management as well, but I want to avoid myself saying that because someone would jump onto me and say 'You don't know what you're talking about!'. That's one I have to thank you for
Now I can only dream of a matured, improved K10.

I think we're seeing better things under the new management. I love seeing the faster iteration of products from AMD, just that I think if they just drop Steamroller and go for Stars again. It might be too late, and I know it's a hard decision to make. Personally, I would just drop steamroller already. Not sure how my engineering experience could translate well into semiconductor design, but from my experiences so far (I work in PCB design and programming), is when you're so knee-deep in design and you don't really like how things are going, start over from scratch, redo from the experience gained, and just do better the next time. That's just me though, and I don't have the pressure of answering to the CEO when my CEO is an electronics engineer himself.

[FlanK3r]

Its no way...K10 Stars are yout yet. Tomshardware proof, Piledriver core is better this days 2012-2013. Deneb was average in same performance as Athlon APU, FX-43xx with L3 cache was better. After OC the same (Deneb OC vs FX-43XX OC).
It not much possible improve K10 in new architecture for 32/28nm process with some higher gain.

[Vassili]

I guess L3 cache would come in handy, maybe Carrizo?
I'm gonna buy one, just to know how it feels in the real world.

Here is another interesting test, it's about memory scaling:
http://semiaccurate.com/2014/01/15/e...h-amds-kaveri/

And one about the performance per watt/scaling:
http://semiaccurate.com/2014/01/14/d...e-amds-kaveri/

Edit:
Just a crazy idea but how hard would it be to compile/make Linux kernel HSA compatible?

[Darakian]

Edit:
Just a crazy idea but how hard would it be to compile/make Linux kernel HSA compatible?

Short answer: Very.
Long answer: It's more than just making the kernel an HSA thing. The kernel really just exposes hardware to applications which then use it, so I'm guessing you're asking for either GPU cores/clusters/whatevers to be exposed as general compute processors or for the linux libraries to use the GPU as a general compute thingy allowing generic apps to use the gpu for compute. The former would be a complete kernel rewrite with a new model for GPU drivers. The later is happening.
The LLVM compiler group is working on extracting parallel code from serial code during the compile process, a feature that will help a lot with respect to GPUs.
https://en.wikipedia.org/wiki/Llvm
Oracle and AMD are working on making the JVM gpu aware and transparent which "should" make all java code gpu accelerated without any recoding work. According to AMD's slides this will come in java version 9.
So long answer, very hard but they're working on it

[god_43]

amd cant design their way out of a paper bag. i second the motion that amd close up cpu shop, at this point its just embarrassing. Steamroller was supposed to be bulldozer done right? i dont like this new amd, we dont even know what comes after XV? so now we must wait to see if xv is amds ace in the hole? yeah that's not going to happen.

[FlanK3r]

no, but Im happy for last two Bulldozers. Vishera really helped much in games in aplications. Most of increase clock to clock were fixed bugs, but still...

clock for clock: http://www.hardware.fr/articles/880-...ver-4-ghz.html

SR is again better to piledriver core. Only it needed higher clocks. If HSA will be more and more supported in future...Maybe there is hope for heteregenous systems.

[dasa]

so if mantle will be able to effectively offload work to the igp that would normally be done by the cpu i wonder could a10-7850 become a competitive option for a gaming system when its currently priced the about same as i5 4570 which is around twice as fast

[Sparky]

None of you seriously want AMD to close up the CPU shop. Yeah, sure, we'd do so well with Intel being the only CPU maker in town

[vardirox]

HSA Acceleration





Transcode Acceleration




One thing is clear about the AMD Kaveri: when it accelerates, it is capable of ludicrous speed against the competition.




http://www.guru3d.com/articles_pages...review,13.html