David Kanter: AMD's ARM core will be 10% faster than their x86 one, ditch Bulldozer

[NTMBK]

> The ISA advantage will be greatly reduced in the top-end side of the performance spectrum, but will not vanish.
> Keller mentioned during Core Day conference that his K12 core will have a "bigger engine" than its x86 sister
> thanks to the advantages of ARMv8 over x86-64, which allows to spend more transistors on compute.

I happen to know the differences between those two designs. I'm not really sure it's going to translate into a significant performance delta. My guess is maybe 10%.

http://www.realworldtech.com/forum/?threadid=142175&curpostid=142759

I can't really share anything unfortunately. However, I can say that they will be ditching Bulldozer and its lineage.

http://www.realworldtech.com/forum/?threadid=142175&curpostid=142792

 

[meloz]

And immediately I have some issues.

• Nowhere does he say ARMv8 core will have a 10% advantage over this hypothetical AMD64 core. He talks of a 10% delta, for all we know this net performance could be in favor of AMD64 even though from a performance/watt perspective -and maybe even performance/die area perspective- the ARMv8 design might make far more sense.
  
  
• I hope this does not turn into yet another "x86(sic) is dying" warcry. A look at Haswell benchmarks shows us the true performance of a modern AMD64 core. And this is with software compiled for AMD64, when people compile for Haswell (AVX2 support) the gap grows larger. Intel have nothing to fear from this upcoming ARMv8 crapware.
  
  
• The 10% delta is a sad reflection on just how much AMD have butchered their own AMD64 core. The latest Kaveri benchmarks are laughable, even after all these years they can barely keep up with a Sandy Bridge.
  
  
• We also do not know which AMD64 core this ARMv8 is being compared against, it could be some crippled 'cost-optimized' future AMD64 APU primarily meant for consoles and such.

Aside, is there any way to list all comments on RWT chronologically, or do you have to always give Kanter his one click per comment?

 

[witeken]

That whole thread is quite interesting to read. There's also some talk about the "x86 tax".

Aside, is there any way to list all comments on RWT chronologically, or do you have to always give Kanter his one click per comment?

Unfortunately it seems not.

 

[NTMBK]

Nowhere does he say ARMv8 core will have a 10% advantage over this hypothetical AMD64 core. He talks of a 10% delta, for all we know this net performance could be in favor of AMD64 even though from a performance/watt perspective -and maybe even performance/die area perspective- the ARMv8 design might make far more sense.

He was responding to a poster who brought up Jim Keller's "bigger engine" comments, where he stated that K12 would have higher performance than its x86 sibling. Kanter was just providing his prediction based on information he has seen (and sadly is not yet able to share).

I hope this does not turn into yet another "x86(sic) is dying" warcry.

That certainly wasn't my intention Heck, just go read the whole of that thread on RWT to see why I don't want an ARM-vs-x86 flamewar! I just wished to share some information about AMD's new CPUs. They have kept their cards very close to their chest, so any snippet is interesting. For instance I think this is the first time I have heard it confirmed that they are abandoning the BD lineage.

The 10% delta is a sad reflection on just how much AMD have butchered their own AMD64 core. The latest Kaveri benchmarks are laughable, even after all these years they can barely keep up with a Sandy Bridge.

Presumably this is why they are ditching the Bulldozer lineage.

We also do not know which AMD64 core this ARMv8 is being compared against, it could be some crippled 'cost-optimized' future AMD64 APU primarily meant for consoles and such.

It is being compared against it's "sister core". AMD are developing a new ARM and x86 core simultaneously, to be pin compatible with the same sockets, and sharing other design similarities. (Though it is still unclear just how much in common they have.)

 

[ShintaiDK]

Isnt it just underlining that AMD gambles fully on ARM and hope its a market that can keep them going. Since the delta in x86 is getting bigger and bigger vs Intel.

 

[ShintaiDK]

That whole thread is quite interesting to read. There's also some talk about the "x86 tax".


Unfortunately it seems not.

There is a short version of it as well:

ARM is full of legacy crap as well. Not to mention the fact that an ARMv8 requires 3-4 different decoders. I know a few people who have had the pleasure of designing custom ARM cores, and according to them 'ARMv8 decode is just as terrible as x86'

http://www.realworldtech.com/forum/?threadid=142175&curpostid=142291

 

[itsmydamnation]

for an arm server they don't have to support the old arm instruciton sets. theres no need, there is no software yet.

 

[NTMBK]

Isnt it just underlining that AMD gambles fully on ARM and hope its a market that can keep them going. Since the delta in x86 is getting bigger and bigger vs Intel.

Not really. It is simply a comparison of the relative performance they expect to get on the same platform (i.e. same memory bandwidth, same socket size, same TDP limits) with their two related designs. The fact that they are designing a new, non-Bulldozer derived x86 core for servers in the first place is pretty strong evidence that they are not gambling solely on ARM.

 

[ShintaiDK]

Not really. It is simply a comparison of the relative performance they expect to get on the same platform (i.e. same memory bandwidth, same socket size, same TDP limits) with their two related designs. The fact that they are designing a new, non-Bulldozer derived x86 core for servers in the first place is pretty strong evidence that they are not gambling solely on ARM.

The question was essentially. Who got the biggest R&D budget of the 2. x86 or ARM? If its ARM, then AMD put its bets on ARM for its future. While x86 would grow ever more insignificant for them. We also have to look realisticly on their CPU division sales. And they have no x86 server share worth writing home about and their regular PC sales is in an ever declining cyclus. So isnt it logical to make the bet on something with more equal competition from other fabless companies with a fraction of Intels revenue? Specially when IDMs are moving further ahead of the fabless.

Also sharing the same sockets and infrastruture isnt ideal. Either one or both of them have to suffer in a certain degree from it. Also why Intel never got IA64 and x86 unified on the same socket.

 

[NTMBK]

The question was essentially. Who got the biggest R&D budget of the 2. x86 or ARM? If its ARM, then AMD put its bets on ARM for its future.

It's certainly an interesting question. I wonder how much of the R&D spending is common to the two? Things like the platform design will be common across both, obviously, and the uncore/on die fabric design is likely to be the same too. Will the memory controller be the same? A fast DDR4 controller is a fast DDR4 controller, will it need to be different for the two? Will research into stacked memory pay off for both designs? (I would suspect so, in some way.) Can research into better cache hierarchies be reused between both- in terms of fundamental design concepts, if not in precise implementation? I'm looking forward to finding out.

 

[meloz]

He was responding to a poster who brought up Jim Keller's "bigger engine" comments, where he stated that K12 would have higher performance than its x86 sibling. Kanter was just providing his prediction based on information he has seen (and sadly is not yet able to share).

That's not how I read it. Maybe I misread it. That website is hellish to read because of how they format their comments, but I'll try to quote the relevant bits:

> > > > > The ISA advantage will be greatly reduced in the top-end
> > > > > side of the performance spectrum, but will not vanish.
> > > > > Keller mentioned during Core Day conference that his K12
> > > > > core will have a "bigger engine" than its x86 sister
> > > > > thanks to the advantages of ARMv8 over x86-64, which allows to spend more transistors on compute.
> > > >
> > > > I happen to know the differences between those two designs. I'm not really sure it's
> > > > going to translate into a significant performance delta. My guess is maybe 10%.

So they are talking of reducing the gap and right now AMD64 is ahead. Thus Kanter's delta comment could only mean they still expect AMD64 to be ahead. Maybe someone can post there and ask Kanter to clarify.

Are you taking the 'more transistors on compute' statement to infer that ARMv8 will be ahead? That is not necessarily true, as we have already seen with AMD. AMD delivered a very nice abortion with Bulldozer, it was worse than its own predecessor in single-threaded performance, to say nothing of Intel CPUs.

That certainly wasn't my intention Heck, just go read the whole of that thread on RWT to see why I don't want an ARM-vs-x86 flamewar!

I would read, but that website is so hard to follow with so many clicks required. Pity, as the content is usually thought provoking.

It is being compared against it's "sister core". AMD are developing a new ARM and x86 core simultaneously, to be pin compatible with the same sockets, and sharing other design similarities. (Though it is still unclear just how much in common they have.)

And this is why I am not at all impressed even if they reach within 10% of AMD64 with their ARMv8.

Is it really a great achievement for ARMv8 when AMD admit that they are crippling their future AMD64 core ("K12 core will have a "bigger engine" than its x86 sister"). The design is already lopsided in ARMv8's favor because AMD know they cannot design a 'Haswell' and thus must rely on shared-cost ARM designs in future. The AMD64 'sister core' is being built to satisfy console market, it seems to me.

There also seems to be some unspoken implication in Keller's statement that AMD64 is inherently bottlenecked or inferior, but if someone wants to make a great AMD64 core it can be done. Just look at Haswell! Even when software is compiled for AMD64 -and not Haswell- it still delivers terrific performance/watt.

After making their name with AMD64 they are now intentionally sabotaging their most famous 'brand' CPU just to make the new hot thing (ARMv8) look good. Pathetic. May be I am wrong and it is not intentional sabotage, AMD are naturally that incompetent. That would explain Bulldozer come to think of it.

Anyway, market will not judge AMD's ARMv8 product against their own impotent AMD64 cores, the reference will always be Intel's CPU. Unless AMD can give stronger performance/watt while also improving single-threaded performance they will continue to falter.

 

[witeken]

Meloz, what is ambiguous about

"Keller mentioned during Core Day conference that his K12 core will have a "bigger engine" than its x86 sister."

 

[meloz]

Meloz, what is ambiguous about

"Keller mentioned during Core Day conference that his K12 core will have a "bigger engine" than its x86 sister."

Bigger engine (whatever that means) does not necessarily translate into bigger performance.

 

[NTMBK]

So they are talking of reducing the gap and right now AMD64 is ahead. Thus Kanter's delta comment could only mean they still expect AMD64 to be ahead. Maybe someone can post there and ask Kanter to clarify.

The gap that poster was referring to was the supposed gap in efficiency between ARM and x86 ISAs, in the context of phones. (Comparing Apple's A7 vs Silvermont, etc.) The way I read that comment was that when you move up to higher performance levels, i.e. server instead of mobile, the supposed "cost" of x86 is reduced.

...I think.

Are you taking the 'more transistors on compute' statement to infer that ARMv8 will be ahead? That is not necessarily true, as we have already seen with AMD. AMD delivered a very nice abortion with Bulldozer, it was worse than its own predecessor in single-threaded performance, to say nothing of Intel CPUs.

Ah, but there you are comparing across different generations or between AMD and Intel. Within the same generation (and with similarly designed cores), I would expect there to be a more straightforward relationship. But I agree that its certainly not guaranteed!

And this is why I am not at all impressed even if they reach within 10% of AMD64 with their ARMv8.

Is it really a great achievement for ARMv8 when AMD admit that they are crippling their future AMD64 core ("K12 core will have a "bigger engine" than its x86 sister"). The design is already lopsided in ARMv8's favor because AMD know they cannot design a 'Haswell' and thus must rely on shared-cost ARM designs in future. The AMD64 'sister core' is being built to satisfy console market, it seems to me.

I don't view it as a crippling- more that by having a smaller decode portion of the chip (and more straightforward logic paths? IDK, not a CPU designer) they can fit more performance into the same number of transistors/same die size/same TDP/same socket.

But here's hoping that it's a proper successor to the Phenom II/competitor to Skylake, not just a slightly souped up Bobcat.

There also seems to be some unspoken implication in Keller's statement that AMD64 is inherently bottlenecked or inferior, but if someone wants to make a great AMD64 core it can be done. Just look at Haswell! Even when software is compiled for AMD64 -and not Haswell- it still delivers terrific performance/watt.

You can certainly build very powerful AMD64 cores, no question. But could the same die size/R&D costs/power budget give you a slightly more powerful ARM core? *shrug* Of course it's irrelevant to me, I don't want to give up the last 20 years of PC games for an extra 10%.

After making their name with AMD64 they are now intentionally sabotaging their most famous 'brand' CPU just to make the new hot thing (ARMv8) look good. Pathetic. May be I am wrong and it is not intentional sabotage, AMD are naturally that incompetent. That would explain Bulldozer come to think of it.

The thing is, "AMD64" isn't really the brand people buy- and considering Intel tried to call it EMT64, and Microsoft settled on x64, and others call it x86-64, it's not that strong a brand. Their strong brands were Athlon and Opteron (though they have completely mismanaged those, too...) Still, I think it's kind of cool that they've called the new core K12, it's a nice callback to the K8 and K10.

Anyway, market will not judge AMD's ARMv8 product against their own impotent AMD64 cores, the reference will always be Intel's CPU. Unless AMD can give stronger performance/watt while also improving single-threaded performance they will continue to falter.

Yeah, the true competition will always be Intel. They have a large gap to close.

 

[pTmdfx]

Bigger engine (whatever that means) does not necessarily translate into bigger performance.

Yep, though it generally is, even if it is just marginally better. Clock range is uncertain, by the way. If we assume bigger means wider architecture with more execution units and deeper structures (eh, as more die area is being occupied), it is possible of a design to sacrifice some of the frequency headroom, or to not design for optimally clocking higher than a certain point, in order to implement such a thing efficiently and easily. Whether or not the core pursues SMT will also affect the design choice.

Another interesting bit I can think of is the width of the vector cluster - AMD may implement full 256-bit vector units in this generation of core, as XV seems to have AVX2 supported according to its related GCC patch. In the meantime, NEON is still sticking to 128-bit in the foreseeable future. If there is really such a plan, the larger vector unit of the x86 core may draw some die/power budget away from the integer complex.

 

[pTmdfx]

So they are talking of reducing the gap and right now AMD64 is ahead. Thus Kanter's delta comment could only mean they still expect AMD64 to be ahead. Maybe someone can post there and ask Kanter to clarify.

The "two designs" in his comment was, by context, more likely referring to AMD's two brand new cores in the sentence right next to the one you highlighted. They were known to be built "a bit differently", so he was adding some colors to this with what he expects the difference that he knew of the two cores will contribute to the final performance (probably single thread performance at a certain clock, as usual in most of the comparative analysis of microarchitectures).

 

[pTmdfx]

It's certainly an interesting question. I wonder how much of the R&D spending is common to the two? Things like the platform design will be common across both, obviously, and the uncore/on die fabric design is likely to be the same too. Will the memory controller be the same? A fast DDR4 controller is a fast DDR4 controller, will it need to be different for the two? Will research into stacked memory pay off for both designs? (I would suspect so, in some way.) Can research into better cache hierarchies be reused between both- in terms of fundamental design concepts, if not in precise implementation? I'm looking forward to finding out.

A lot of design features within the microarchitecture shall be common either. I would say in K12's case it would always cumulatively spend more than the new x86 core, because AMD had nothing in its R&D pipeline for bringing up an ARM product (let alone a core), and they have to fund it and build it up from nothing.

 

[Hans Gruber]

When I first joined this forum in 2006 I was using and AMD Barton 2500+. Then an AMD 64 Athlon and then a AMD 64 Athlon x2. After that I went Intel with the Q6600 and never looked back. I have an Ivy Bridge 3570K @4.5ghz today.

My question is where does the thermal load come from for AMD? Since their architecture is inferior to Intel's why not build on the same architecture and simply build CPU's that can clock up to 7Ghz? Speed wars like the good old days of Intel Vs. AMD.

Off the subject, Intel has been bottlenecked with Ivy Bridge, Haswell and Devil's Canyon with the same 22nm thermal wall which makes the overclocking of the CPU's around the 4.5-4.7ghz range. Sandy Bridge hit 5ghz but mostly with crazy voltages.

Also, why doesn't both AMD and Intel integrate their graphics on a chip to run parallel to the PCI Express GPU's? They have software that attempts to do this but it should be in house from AMD and Intel.

Both XBOX ONE and PS4 run on AMD CPU's and GPU cores. It's like steroids, PC gamers need a boost. They could easily have the integrated graphics do the little things while the GPU's do the big stuff. I have seen simulations with a 40% increase in graphics performance.

I agree with the OP, AMD's stuff sucks. Since they are struggling to innovate, simply create a CPU that can crank up the party to 7-9Ghz. If you can't beat them with core structure you can beat them with pure speed.

 

[witeken]

You can't simply create a 9GHz CPU. The amount of heat it will produce will be insane. Only few people are waiting for that.

 

[Erenhardt]

It seems AMD CPUs are fine performance wise:
http://forums.anandtech.com/showthread.php?t=2394740
The problem is in the software that doesn't utilize their architecture well. We just see a shift to 4 and more cores in what is widely (wrongly?) used as general performance benchmark - Games. AMD is too far in front with their designs. Software can't keep up.
The same with HSA. They have igpu co-compute with cpu, Hans. Problems is the same as it was with FX CPUs. The software just isn't there yet. Before it catches up AMD will be releasing their next big thing, which again, will lack software support. Rinse and repeat.

 

[frozentundra123456]

Wow, havent heard this excuse for a few years now.

 

[meloz]

The problem is in the software that doesn't utilize their architecture well.

If anyone should be complaining about software not utilizing their arch properly, it is Intel. Poor Haswell, everyone runs software compiled for AMD64 on it, when they could get upto 20% or more performance boost in certain loads if software was compiled for Haswell (AVX2). And yet, somehow, Haswell thrashes AMD across the spectrum. AMD have absolutely nothing to complain about.

We just see a shift to 4 and more cores in what is widely (wrongly?) used as general performance benchmark - Games. AMD is too far in front with their designs. Software can't keep up.

And yet, somehow, Intel does well on same benchmarks with its four cores / eight threads CPUs. Let's stop blaming software. If anyone has not kept up in this industry it is AMD.

The same with HSA. They have igpu co-compute with cpu, Hans. Problems is the same as it was with FX CPUs. The software just isn't there yet. Before it catches up AMD will be releasing their next big thing, which again, will lack software support. Rinse and repeat.

HSA is another distraction they created for themselves, guess they need something to talk about since their Bulldozer and now Steamroller is an abject failure. In reality any software that needs iGPU can access it with openCL, and that is the prefered way to use iGPU because with openCL the compiled binary works across Intel and AMD CPUs.

AMD are unnecessarily creating more burden for themselves by trying to promote HSA, a solution for a problem that does not truly exist. So far the biggest benefit of HSA has been that it has given AMD something to crow about, and fool the gullible media that AMD has something special in its hand.

 

[mrmt]

It's certainly an interesting question. I wonder how much of the R&D spending is common to the two?

The R&D spending is part of the equation, the compromises that AMD is having to make to reuse the IP is the interesting part. I'm betting that AMD is sacrificing a lot the x86 core while being a lot more open-handed to the ARM core. This, and not a x86 tax should be what is making the x86 inferior to the ARM core and overall a mediocre product.

Bulldozer is dead, long live Bulldozer!

 

[monstercameron]

We all know that there are a few well threaded apps where AMD does better than Intel in. AMD's design has its strengths. HSA also has its strengths, people are interested in gpgpu and HSA style processing but the tools and methods are still early days. Remember that HSA. Isn't just a AMD thing, many independent companies have joined the HSA foundation.

 

[mrmt]

We all know that there are a few well threaded apps where AMD does better than Intel in. AMD's design has its strengths. HSA also has its strengths, people are interested in gpgpu and HSA style processing but the tools and methods are still early days. Remember that HSA. Isn't just a AMD thing, many independent companies have joined the HSA foundation.

There's not a single case where AMD does better than Intel on perf/area and perf/watt. The fact that you can get more perf/$ with AMD has more to do with AMD manufacturing bad products and having to sell them in a lower market bracket than it should than with the product strengths.

As for HSA, wake me up when someone other than delivers HSA-enabled hardware.