AMD CPU speculation... and expert conjecture

[juanrga]

LOL

 

[Cazalan]

GeekBench3 is my bible so let me know when any ARM systems can get to this level.

http://browser.primatelabs.com/geekbench3/491996

 

[de5_Roy]

TSMC Delivers First Fully Functional 16FinFET Networking Processor
http://www.techpowerup.com/205658/tsmc-delivers-first-fully-functional-16finfet-networking-processor.html

It(the process) has twice the gate density of TSMC's 28HPM process, and operates more than 40% faster at the same total power, or reduces total power over 60% at the same speed.

promising.

 

[cdrkf]

ARM scales upward poorly.

Not because the core cannot do the work, but because by the time you get ARM to do what x86 can on the high end, you now need all kinds of additional hardware for things like OoO, and other stuff.

What is being highly misplaced is comparing cherry picked benchmarks for an in order execution chip designed for phones and tablets against an OoO capable chip for laptops.

Consider this, AnTuTu is not aimed at OoO execution, because no ARM chips to this point will do OoO with perhaps exception to 80W server chips, and even then, you are no longer in much better territory, which is what I said all along. If you want ARM to compete on that end it will not be at 20w. If their cores excel at microserver functions, that is fine, they are designed for it. However do not hold your breath watching for a wall street firm to sink big money into ARM servers...

As for HPC, everyone knows that you need just enough cores to baby sit a massive bank of GPUs and keep them fed well...let us not kid ourselves shall we?

It's my understanding that A57 is an OoO design though?

 

[vmN]

It's my understanding that A57 is an OoO design though?

Yes. More information can be found here: http://www.arm.com/products/processors/cortex-a/cortex-a57-processor.php

 

[juanrga]

It's my understanding that A57 is an OoO design though?

Yes, the A57 used in AMD's Seattle is OoO, and the former 32bit Cortex A15 is OoO, and Apple Cyclone is OoO and Qualcomm Krait is OoO... and AMD K12 is OoO. He was spreading misinformation once again, here my above "LOL".

 

[juanrga]

don't hold your breath, it will take some time to reach Intel Xeon E7-4890 v2 @ 2.80 GHz
4 processors, 60 cores, 64 threads level. K1 only does 3470 in GB 3

http://cdn.wccftech.com/wp-content/uploads/2014/05/NVIDIA-Tegra-K1-Xaomi-Mi-Performance-1.jpg

The SoC+DRAM consumes less than 7W, and knowing that TDP != power consumption we can assume that CPU alone is rated at about 3W or so. The mentioned multi-socket Xeon high-end server is rated at 600W TDP (only the CPUs). Nvidia has made a good work and Denver is both fast and efficient, but cannot do miracles and beat a 600W server-class system.

It is worth mentioning that Denver is not aimed at servers. Nvidia project for servers is Boulder, which is aimed to compete against Xeons and Opterons.

Whereas Boulder arrives, we have other announced server-class ARM products that beat Xeons E5/E7. AMD K12 will compete against those ARM products, thus I expect K12 to be very fast.

 

[gamerk316]

When we find out that ARM scaled up won't beat X86, I will be in full "I told you so" mode. Shame 8350 is under NDA and can't give us a reveal timeline.

 

[gamerk316]

^^ To be fair though, you do have a lot more latency problems on the i3. And for non-gaming general use applications, FX is better.

That being said, I still want to see what an i3 @ 4.5GHz could do...

 

[-Fran-]

^^ To be fair though, you do have a lot more latency problems on the i3. And for non-gaming general use applications, FX is better.

That being said, I still want to see what an i3 @ 4.5GHz could do...

Don't waste your time gamerk, he's a troll.

Cheers!

 

[jdwii]

When we find out that ARM scaled up won't beat X86, I will be in full "I told you so" mode. Shame 8350 is under NDA and can't give us a reveal timeline.

TBH i would be quite impressed if Amd could deliver K12 and offer speeds that juan is talking about and i'm big enough to tell juan hey you were right. But something tells me like everything else he claims its going to be off. It did indeed take 8 years for project denver and some of the team responsible for that design came from Amd themselves(although before 2011). That being said it is possible to see a product from Amd that is greater than project denver on the Arm side and given Amd's remarkable low performance per clock(i hate the false IPC term that is thrown around here) i wouldn't say juan isn't to far off when he claims the K12 will have higher "IPC" compared to zen.
Edit
I would also like to stick up for juan and say he never claimed Arm would beat Intel's X86 in terms of performance besides in server applications in a heavy threaded environment, although we all suspect that is what he thinks.

 

[jdwii]

6GB of VRAM!!
http://www.dsogaming.com/news/middle-earth-shadow-of-mordor-pc-settings-unveiled-ultra-textures-require-6gb-of-vram/
Holly jebus

 

[blackkstar]

6GB of VRAM!!
http://www.dsogaming.com/news/middle-earth-shadow-of-mordor-pc-settings-unveiled-ultra-textures-require-6gb-of-vram/
Holly jebus

Yeah, but it's bethesda, lol. Look how demanding skyrim was on some systems and how crappy it looked.

We don't need to discuss Juan. He cherrypicks benchmarks to prove his point. He already did it when he claimed i5 was on par with 2m/4c Steamroller APU. As FX said, he will find benchmarks that show ARM as best as possible and x86 as worse as possible and then look at 4 benchmarks and make broad generalizations about how the ARM is better because he has 4 benchmarks that "prove" his point. When no one only uses 4 programs.

It's difficult enough to benchmark and draw a meaningful comparison between Piledriver and IB/Haswell because both x86 CPUs have massive strengths and weaknesses. Comparing ARM and x86 just amplifies that problem and the only real way to come to meaningful conclusions is to analyze the architecture.

 

[Cazalan]

TSMC Delivers First Fully Functional 16FinFET Networking Processor
http://www.techpowerup.com/205658/tsmc-delivers-first-fully-functional-16finfet-networking-processor.html

It(the process) has twice the gate density of TSMC's 28HPM process, and operates more than 40% faster at the same total power, or reduces total power over 60% at the same speed.

promising.

Good signs they're on track for 2016 shipments. Things are heating up in the fab space. It took Intel like 18 months from early risk production to volume 14nm production. The added complexities of triple patterning lithography.

Wonder how AMDs 16 core custom ARMv8 will do against a 32 core stock A57.

 

[h2323]

According to AMD, Kaveri A10 is 68% faster than Haswell i5-K on BF4 @1080p medium settings, both using integrated graphics. According to Charlie, Broadwell GPU will be 40% faster than Haswell. Adding that desktop K series will be upgraded to Iris Pro level graphics one obtains that Broadwell i5-K will run BF4 faster than Kaveri. In fact, the Haswell i5-R already offers more than 80% of the performance of Kaveri in BF4.

But let us be conservative and assume that Broadwell i5-R will only run BF4 as fast as Kaveri. Skylake graphics will bring another 40% gain over Broadwell... The claim that Skylake will only be on par with Kaveri is nonsense. It is the same kind of nonsense coming from those believing that Zen will be faster than Skylake on CPU terms.

The people that pretends that the decisions behind KL CPU are marketing are plain wrong. The transition from multicore to manycore is motivated by well-known scientific/technical stuff which is independent of brands and, of course, of marketing. What was really hilarious was reading the "KL doesnt even have an IGP". This proves again this people didn't understand anything of what was said there.

I also find interesting that some people believes that 200--250W APUs will be used for running linux games on laptops. I want see that foolish mobile stuff

As you know Intel's GPU performance claims have traditionally been optimistic. Kaveri's GPU is limited by a severe bottleneck. Carrizo will likely have a bottleneck as well.

There is zero information regarding the performance of Skylake. There is zero information regarding the performance of Zen. Zen is a from the ground up CPU, it would be unwise to peg it's performance down based on a chart of CPU performance increase from AMD's current design and then compare that to Intel's CPU performance increase along a similar timeline. But that is all you have to go on other than an extremely small amount of information that you twist from AMD. You have no clue what the new cores will be, do not write like you have inside knowledge of this, you don't.

What's amusing is watching you troll the internet saying things about Zen, K12, Skylake with little to nill information. Twisting every little news story to fit your world view.

Readers her can go to the CPU forums on semi-accurate and see the same thing as he does here, with less freedom though, they slap Juan around when he gets to crazy.

 

[gamerk316]

As you know Intel's GPU performance claims have traditionally been optimistic. Kaveri's GPU is limited by a severe bottleneck. Carrizo will likely have a bottleneck as well.

I'll say it again: Though there is a memory bandwidth bottleneck, I do not think it is nearly as server as people think it is.

 

[de5_Roy]

As you know Intel's GPU performance claims have traditionally been optimistic. Kaveri's GPU is limited by a severe bottleneck. Carrizo will likely have a bottleneck as well.

I'll say it again: Though there is a memory bandwidth bottleneck, I do not think it is nearly as server as people think it is.

i wouldn't call shouting something like <1st promo slide>"300% moar powerful....
....
...
..
<the last, the one that never gets published on a tech website - the disclaimer slide> ... "than hd2000" optimistic.

if carrizo gets edram/stacked ram treatment, it will have a lot of room to flex it's igpu muscles.

 

[-Fran-]

As you know Intel's GPU performance claims have traditionally been optimistic. Kaveri's GPU is limited by a severe bottleneck. Carrizo will likely have a bottleneck as well.

I'll say it again: Though there is a memory bandwidth bottleneck, I do not think it is nearly as server as people think it is.

It is a big bottleneck. If you go just by specs of the iGPU inside the APUs and compare them to their dGPU counterparts with GDDR5, the difference is always around 30-40% (IIRC and can't find direct comparisons either =/).

Take a look at IrisPro from Intel. When the L4-as-RAM is not saturated, it is lightning fast (for an iGPU).

Cheers!

 

[juanrga]

Almost everyone in the industry applaud AMD strategy of migrating from x86 to ARM to stole market server to Intel. We also know that half-dozen of people that dislike/hate AMD will continue posting nonsense and lies in forums (like their "ARM cannot scale up because isn't CISC", "Denver is quad-core", "ARM cannot do OoO", "ARM only can run Android"...) and they will continue attacking and lying about everyone who don't accept their silly viewpoints.

Of course, no amount of your trolling/lying will change the facts but you can think otherwise :-D

 

[cdrkf]

As you know Intel's GPU performance claims have traditionally been optimistic. Kaveri's GPU is limited by a severe bottleneck. Carrizo will likely have a bottleneck as well.

I'll say it again: Though there is a memory bandwidth bottleneck, I do not think it is nearly as server as people think it is.

Simple test for that is compare Kaveri benchmarks against a HD 7750. The iGPU in Kaveri has the same shader config, with a few updated features (newer GCN variant). The only *disadvantage* is that Kaveri is using the IMC and DDR3 memory. A GDDR5 equipped HD7750 is notably faster (would have to double check numbers). That is the speed up faster memory will give essentially- and it's not insignificant

 

[gamerk316]

Almost everyone in the industry applaud AMD strategy of migrating from x86 to ARM to stole market server to Intel. We also know that half-dozen of people that dislike/hate AMD will continue posting nonsense and lies in forums (like their "ARM cannot scale up because isn't CISC", "Denver is quad-core", "ARM cannot do OoO", "ARM only can run Android"...) and they will continue attacking and lying about everyone who don't accept their silly viewpoints.

Of course, no amount of your trolling/lying will change the facts but you can think otherwise :-D

No Juan, our primary argument is that by the time you scale up ARM to include what X86 has, you are going to end up with a CPU that is simply slower, hotter, and all around not as good as a comparable X86 chip. Intel's spend 30 years optimizing the design; you aren't going to throw those same blocks into an ARM chip and suddenly end up with optimized performance.

That, and software. Even if ARM performance, AMD has to make a case for developers to develop the necessary software, and for end users to convert their entire infrastructure over to ARM.

Point being, ARM needs to be better then X86 in sever for it to be successful, and I doubt, at the same TDP, it reaches 80% of X86 performance.

 

[cdrkf]

Almost everyone in the industry applaud AMD strategy of migrating from x86 to ARM to stole market server to Intel. We also know that half-dozen of people that dislike/hate AMD will continue posting nonsense and lies in forums (like their "ARM cannot scale up because isn't CISC", "Denver is quad-core", "ARM cannot do OoO", "ARM only can run Android"...) and they will continue attacking and lying about everyone who don't accept their silly viewpoints.

Of course, no amount of your trolling/lying will change the facts but you can think otherwise :-D

No Juan, our primary argument is that by the time you scale up ARM to include what X86 has, you are going to end up with a CPU that is simply slower, hotter, and all around not as good as a comparable X86 chip. Intel's spend 30 years optimizing the design; you aren't going to throw those same blocks into an ARM chip and suddenly end up with optimized performance.

That, and software. Even if ARM performance, AMD has to make a case for developers to develop the necessary software, and for end users to convert their entire infrastructure over to ARM.

Point being, ARM needs to be better then X86 in sever for it to be successful, and I doubt, at the same TDP, it reaches 80% of X86 performance.

I wonder if your underestimating ARM a bit here, along with AMD for that matter. Keller is convinced there are some fundamental advantages of using ARM V8 over x86 (a big one being that ARM V8 is more modern, I think x86 and x86-64 suffer from allot of baggage due to it's age). That doesn't mean to say I think that at the same TDP an ARM chip automatically demolishes the best Intel has to offer, but I can still see there being some surprises. As with all thing it will undoubtedly come down to *specific use cases* and *particular applications*. Some areas Intel will be insurmountable, others ARM will probably provide a significant advantage.

It's true that gaining traction with a new product like this will be an issue, still for it to be a big win for AMD it doesn't need to take over the server market. If they can find a few specific areas which are big wins and focus on those, they'll have a shot at getting a few additional % of the server market, and that could make a real positive impact to them as a company.

 

[gamerk316]

As has been shown: ISA doesn't matter when it comes to performance. So at best, you'll get ARM equaling X86, which won't be enough to get people to switch over, due to lack of SW.

Oh wait a sec, ARM won't have AVX or SSE. So even if ARM is competitive at the ISA level, it will lose due to CPU extensions.

Seriously, AMD couldn't beat intel when making a x86-64 chip, and they're the ones who created the bloody ISA in the first place. So why should we suddenly believe AMD could take a foreign CPU architecture, add in 20 years worth of HW improvements, minus the CPU extensions, and somehow come out with a chip that out-competes a high end X86-64 chip from Intel? Because when you put it that way, you realize how silly a prospect what AMD is trying to do really is.

 

[juanrga]

Regarding performance

HD 7750 DDR3 version << Kaveri iGP << HD 7750 GDDR5 version

 

[blackkstar]

Lots of RISC designs have tried to compete with x86 and have not been able to keep up. I realize x86 itself is an ugly mess and x86 ASM over Power or SPARC or whatever is for people who like really good challenges, but people have tried throughout history to unseat x86 with RISC and it never works, at least in the long run.

Also, ARM was designed specifically to not target high performance x86 so they didn't have to take Intel on directly. Much like Via's plans with their CPUs even though they are x86.

Historically I don't see ARM being this magic bullet. In fact I recall hearing similar arguments that x86 is in trouble when PowerPC was being pushed in Macs and there were a lot of fanboys clamoring at the fact that PPC was infinitely better.