[SemiAccurate] AMD to differentiate cores

[naukkis]

ARM started out for desktop in the archimedes line and later as a replacement for the 68000 found in the Amiga and Aplle PCs of the time it made it until around the year 2000 where it became obvious that the technology can't scale upwards enough to remain relevant for desktop use.
The same problems remain today,sure technology came a long way and ARM are much more capable than back then but X86 didn't sleep all this time either.
ARM doesn't scale up with power and x86 doesn't scale down with power that's just the way it is,it might change at some point in the future but it's still a long time until then.

Aarch64 is ground up ISA made for todays knowledge(2011) to make easy to design, power efficient and well performing cpu's. It isn't even compatible with older ARM-versions. AMD and Apple started to design aarch64 cpu's about as soon as it released.

For most important part of performance, IPC, measured as different ISA so not literally by count of instructions, Apple current aarch64 core is about twice as fast as best performing x86.

 

[whm1974]

For most important part of performance, IPC, measured as different ISA so not literally by count of instructions, Apple current aarch64 core is about twice as fast as best performing x86.

Again benchmarks please. And iOS is a stripped down version of MacOS w/applications tailored toward such hardware and uses.

 

[naukkis]

Again benchmarks please. And iOS is a stripped down version of MacOS w/applications tailored toward such hardware and uses.

There's plenty of benchmarks out there, some people just refuses to believe them. How about Tech Report Jim Keller review from K12 vs Zen?:

"Presumably, these sister x86 and ARM cores will perform about the same, but they evidently are not just two variants of the same microarchitecture adapted to different ISAs. Keller was very complimentary about the ARMv8 ISA in his talk, saying it has more registers and "a proper three-operand instruction set." He noted that ARMv8 doesn't require the same instruction decoding hardware as an x86 processor, leaving more room to concentrate on performance. Keller even outright said that "the way we built ARM is a little different from x86" because it "has a bigger engine." I take that to mean AMD's ARM-compatible microarchitecture is somewhat wider than its sister, x86-compatible core. We'll have to see how that difference translates into performance in the long run.

 

[whm1974]

There's plenty of benchmarks out there, some people just refuses to believe them. How about Tech Report Jim Keller review from K12 vs Zen?:

"Presumably, these sister x86 and ARM cores will perform about the same, but they evidently are not just two variants of the same microarchitecture adapted to different ISAs. Keller was very complimentary about the ARMv8 ISA in his talk, saying it has more registers and "a proper three-operand instruction set." He noted that ARMv8 doesn't require the same instruction decoding hardware as an x86 processor, leaving more room to concentrate on performance. Keller even outright said that "the way we built ARM is a little different from x86" because it "has a bigger engine." I take that to mean AMD's ARM-compatible microarchitecture is somewhat wider than its sister, x86-compatible core. We'll have to see how that difference translates into performance in the long run.

And yet K12 was never released and for which applications and use cases?

 

[DrMrLordX]

And yet K12 was never released and for which applications and use cases?

AMD didn't want to blaze that trail. They sort-of tried with Seattle. It didn't get much uptake so there you go.

Some server rooms could easily switch to ARM without anyone noticing/caring as long as the underlying software support was there. Again, who wants to blaze that trail? The answer, right now, appears to be Huawei. They have the backing they need to fill in software gaps. Uptake will be slow in markets hostile to Huawei; that's what you get when you sell yourself out to the CLA.

Will AMD someday regret not developing K12 further? Hard to say. @naukkis brought up some timely remarks. Keller likely won't get to do anything with ARM over at Intel, either, unless they make drastic changes to their CPU designs. For the time being, AMD is better-served locking up as much of the x86 market as they can. They still have K12 in the lineup, and they can update it if they get the cash to expand their design teams and update that for 7nm or 5nm nodes in the future. Let Huawei and Cavium/Marvell do the heavy lifting, examine the software infrastructure that supports their ISA tweaks (if any), and then realign a modernized version of K12 to support the same software. AMD could hit markets interested in ARM but wary of Huawei and Marvell. And that is assuming that Marvell makes it much further with their ThunderX line. They might not.

 

[whm1974]

AMD didn't want to blaze that trail. They sort-of tried with Seattle. It didn't get much uptake so there you go.

Some server rooms could easily switch to ARM without anyone noticing/caring as long as the underlying software support was there. Again, who wants to blaze that trail? The answer, right now, appears to be Huawei. They have the backing they need to fill in software gaps. Uptake will be slow in markets hostile to Huawei; that's what you get when you sell yourself out to the CLA.

Will AMD someday regret not developing K12 further? Hard to say. @naukkis brought up some timely remarks. Keller likely won't get to do anything with ARM over at Intel, either, unless they make drastic changes to their CPU designs. For the time being, AMD is better-served locking up as much of the x86 market as they can. They still have K12 in the lineup, and they can update it if they get the cash to expand their design teams and update that for 7nm or 5nm nodes in the future. Let Huawei and Cavium/Marvell do the heavy lifting, examine the software infrastructure that supports their ISA tweaks (if any), and then realign a modernized version of K12 to support the same software. AMD could hit markets interested in ARM but wary of Huawei and Marvell. And that is assuming that Marvell makes it much further with their ThunderX line. They might not.

That is the thing, users will need proper software support and the applications they need available before they consider buying into a different Platform. On top of that, the new Platform has to offer enough advantages to be worthwhile for users to switch to.

 

[moinmoin]

AMD did make one cpu design to support both ARM and X86 instruction sets. Why would they abandon that path in later versions, as I see offering both high performance x86 and ARM versions of their high performance core is brilliant idea. Mass producing die at top-notch production lines is what needs much funding, for cpu design itself AMD has them already done. And there's more use cases for high performance ARM-core, now AMD semi-custom chips are limited to x86 but how about doing semi-custom mobile SOC for some phone maker which itself just can't do it or want more performance that in-shelf ARM cores would have to compete against Apple, Samsung etc.....

Note that AMD never abandoned their effort on ARM. Both K12 and Zen were worked on by the same team. ARM is integral part of the AMD Secure Processor compatible with ARM TrustZone, as well as of the Scalable Control Fabric part of IF, both of which are part of every Zen chip. AMD just didn't enter the market of selling ARM chips. If somebody were interested to pay for that I'm sure AMD would be open to creating ARM chips as part of its semi custom business. (Feel free to take the fact that this didn't happen so far as an indication that so far nobody sees a sufficient market for such to make it worth happening.)

 

[whm1974]

Note that AMD never abandoned their effort on ARM. Both K12 and Zen were worked on by the same team. ARM is integral part of the AMD Secure Processor compatible with ARM TrustZone, as well as the Scalable Control Fabric part of IF, both of which are part of every Zen chip. AMD just didn't enter the market of selling ARM chips. If somebody were interested to pay for that I'm sure AMD would be open to created ARM chips as part of its semi custom business. (Feel free to take the fact that this didn't happen so far as an indication that so far nobody sees a sufficient market for such to make it worth happening.)

And what market there is, it is already crowded by many other players. Right now there are only two other companies AMD has to complete with, Intel and Nvidia.

 

[moinmoin]

And what market there is, it is already crowded by many other players. Right now there are only two other companies AMD has to complete with, Intel and Nvidia.

Indeed. And semi custom is an service neither of the other two competitors is offering (though Intel is partly opening up in the high end datacenter business at least).

As I wrote if there is actual demand for ARM chips done by AMD there should be a customer paying for such through AMD's semi custom business. The fact that hasn't happened so far points to there being not sufficient demand.

 

[whm1974]

Indeed. And semi custom is an service neither of the other two competitors is offering (though Intel is partly opening up in the high end datacenter business at least).

As I wrote if there is actual demand for ARM chips done by AMD there should be a customer paying for such through AMD's semi custom business. The fact that hasn't happened so far points to there being not sufficient demand.

If I need a custom ARM based Processor/SOC, why would I go to AMD instead of well established and more experienced vendors?

 

[moinmoin]

If I need a custom ARM based Processor/SOC, why would I go to AMD instead of well established and more experienced vendors?

To combine it with AMD's superior GPU tech for example?

The fact that AMD doesn't make mobile SoCs at all is a disadvantage in that regard of course. But there the competition is diverse enough that AMD wouldn't add much anyway I'd say. (This would have been different hadn't AMD sold its handset division to Qualcomm back in 2009, see Imageon/Adreno.)

 

[DrMrLordX]

That is the thing, users will need proper software support and the applications they need available before they consider buying into a different Platform. On top of that, the new Platform has to offer enough advantages to be worthwhile for users to switch to.

Correct. Right now there is some FOSS support for ARM64 in the server room. For large organizations that can produce their own in-house software solutions, that might be all that they need. And, as I stated, I think Huawei will produce custom software solutions where possible. That alone won't be enough for shops looking to use existing software/hypervisors.

As long as Huawei continues to push their multicore ARM server products, I expect them to make significant contributions to existing Linux support for ARM64. That will probably set the standard that other ARM server vendors will have to follow in the future.

If I need a custom ARM based Processor/SOC, why would I go to AMD instead of well established and more experienced vendors?

Depends on what you need. Right now, you wouldn't. It's hard to say exactly what K12 was or wasn't, but I suspect it was similar enough to Zen that it would be aimed first and foremost at server/workstation loads. AMD has demonstrated that their latest iteration of Zen - Zen2 - is suitable for a console, so theoretically they could update K12 and sell it as a CPU worthy of (for example) a console or similar. Then you sidestep the issue of software compatibility since a custom job is going to run in-house software anyway. But how many years would it take them to iterate upon K12 and bring it up to date? How much money would that take?

All that aside, I see no evidence that AMD could update and scale down K12 to make it competitive with something like A12x or A13 on a per-watt basis. AMD is free to prove me wrong.

 

[IntelUser2000]

I'd find the bigger SIMD units for server curious. Intel tries very hard to sell those for server customers, but I have not seen much interest outside HPC.

The wider SIMD units have worked for them for many years. I think its only recently with AVX-512 its being questioned. I think even they realize this with segregating the lineup into two with client having 1 512-bit FPU and server having 2. Server has HPC to use AVX-512 on, what's there on client? AVX-512 allows parallelizing instructions that were previously not doable so it does go beyond just FP performance.

The reason for wider SIMD in CPUs are twofold and its related.

1. Not all HPC scientific code is amenable to GPU or benefits a lot from it. You have few extremes. Very parallel, very vectorizable, and memory bandwidth bound. And you have code that's exact opposite of the three. Most codes are complex combinations of them in varying degrees. If you have a customer that can benefit from faster vector units they'll comply.
2. Intel has the server market cornered. Nvidia is trying to get more with their GPUs. Since it takes a lot of development effort to switch, faster vector units in CPUs stave off efforts to switch to GPUs. Let's say if Intel just stayed with SSE2. Then the GPU market might be bigger. They had AVX2 in Haswell, its natural to expect they'll continue to enhance it. You don't think they'd have stayed still do you?

The technologies used and markets are constantly in flux because requirements change over time and the responses change accordingly. You have near-future server CPUs that'll try to address memory bandwidth portion with on-package HBMx memory, and even Intel entering the GPU market.

Who knows how it'll be in 10 years? One thing I can say for sure is it'll continue to be dynamic and never be black-and-white. Analysts that have views that tend towards the extremes are always wrong for this reason.

The question for AMD is whether they'll go the Skylake-SP route and make a variant with more beefy vector units or make a small core like Goldmont Plus. Maybe they'll even do both!

 

[SarahKerrigan]

Again benchmarks please. And iOS is a stripped down version of MacOS w/applications tailored toward such hardware and uses.

Anandtech has run SPEC on multiple generations of Apple cores. 2x is optimistic, but those cores are exceptionally fast - look at here vs here for instance. And the OS is largely irrelevant to SPEC numbers, except for scheduling differences - the timed section of SPEC runs doesn't make system calls.

"Only x86 is magically able to hit high single-thread performance targets" is nonsense. Any ISA, unless fundamentally not fit for fast implementation (like some MCU ISAs, or perhaps VAX), can do it. Microarchitecture matters far more.

 

[NTMBK]

AMD did make one cpu design to support both ARM and X86 instruction sets. Why would they abandon that path in later versions, as I see offering both high performance x86 and ARM versions of their high performance core is brilliant idea. Mass producing die at top-notch production lines is what needs much funding, for cpu design itself AMD has them already done. And there's more use cases for high performance ARM-core, now AMD semi-custom chips are limited to x86 but how about doing semi-custom mobile SOC for some phone maker which itself just can't do it or want more performance that in-shelf ARM cores would have to compete against Apple, Samsung etc.....

AMD don't have a modem, and they don't have GPU IP that efficiently scales down to smartphone power levels. (They sold that division to Qualcomm years ago.) They don't have a hope of competing for smartphone contracts, there's just no way they could beat Qualcomm on their own turf.

And yes, mass production of chips is expensive, but the real killer is the cost to produce chip #1. Doing the design, the physical layout, and generating masks for a new 7nm chip costs literally hundreds of millions of dollars. (I've seen estimates at around $300m.) Why spend such a huge chunk of money to make an ARM server CPU when you already have an AMD64 one? You might make a few more sales, but enough to justify that price? And enough to justify undermining the AMD64 market that has served them so well?

 

[IntelUser2000]

there's just no way they could beat Qualcomm on their own turf.

Not only that, AMD will have to be substantially better than Qualcomm. The underlying reasons for business lies in human behavior, not necessarily technology. Apple didn't use Intel modems because they were better, but because basically they didn't like Qualcomm.

Intel failed after how many billions, acquisitions and years in mobile development? I don't think this is just due to ISA differences. You'd want to go into Apple or Samsung devices. Apple said no, so its Samsung. Samsung has their own chips, why would they use Intel's? The rest are fighting for scraps. 80% of the Smartphone revenue is taken up by Apple, and 80% of the leftover 20% is taken up by Samsung. Scraps.

Why spend such a huge chunk of money to make an ARM server CPU when you already have an AMD64 one?

Yep. 15% or so advantage quoted by Jim Keller is nowhere near enough to switch to a different ISA.

 

[scannall]

ARM started out for desktop in the archimedes line and later as a replacement for the 68000 found in the Amiga and Aplle PCs of the time it made it until around the year 2000 where it became obvious that the technology can't scale upwards enough to remain relevant for desktop use.
The same problems remain today,sure technology came a long way and ARM are much more capable than back then but X86 didn't sleep all this time either.
ARM doesn't scale up with power and x86 doesn't scale down with power that's just the way it is,it might change at some point in the future but it's still a long time until then.

I'm going to respectfully disagree with you on this one. ARM V8.x isn't the same as ARM from 20-25 years ago. They don't really share anything at all. Going back to Apple, simply because it's a good example. By going wide, and out of order like any other high performance CPU whether x86 or Power (IBM) gets a lot of throughput. Clock for clock better than Intel and x86.

For example:
Xeon 4655 at 3Ghz vs Apple A12X at 2.5 Ghz

And this performance comes at a fraction of the power draw of the Xeon. Clearly there is something to ARM. Like any high performance CPU it's a matter of throwing transistors at it, until it performs the way you want, on a die you can afford to make.

 

[NostaSeronx]

The best case for ARM is SVE support. AMD got some architects from the Broadcom's Vulcan-SVE microarchitecture. So, SVE wouldn't be a completely blind implementation for AMD.

-from the whitepaper-

Vector Length Agnostic (VLA) programming:
Unlike traditional SIMD architectures, which defne a fixed size for their vector registers, SVE only specifes a maximum size. This freedom of choice is done to enable diferent ARM architectural licensees to develop their own implementation, targeting specifc workloads and technologies which could benefit from a particular vector length.

A key goal of SVE is to allow the same program image to be run on any implementation of the architecture (yes, which might implement different vector lengths!), so it includes instructions which permit vector code to adapt automatically to the current vector length at runtime.

-hpc compiler team-

What’s the vector length?
- There is no preferred vector length
1. Vector Length (VL) is hardware choice, from 128 to 2048 bits, in increments of 128
- Does not need to be a power-of-2
- Vector Length Agnostic programming adjusts dynamically to the available VL
- No need to recompile, or to rewrite hand-coded SVE assembler or C intrinsics
- Has extensive implications for loop optimizations

One code, across multiple architectures. Is better than SSE -> AVX -> AVX512 -> AVX1024 -> AVX2048 -> etc. This leads to a question of why not have a hybrid x86 & ARM core instead. Run AMD64(+up to AVX2) as legacy and ARM+SVE as modern or a hybrid mode; AMD64(GPR) + SVE(VGPR).

 

[naukkis]

All that aside, I see no evidence that AMD could update and scale down K12 to make it competitive with something like A12x or A13 on a per-watt basis. AMD is free to prove me wrong.

If released 2016 as planned K12 would have been greatly competitive against Apple A10. To compete against A12 or A13 K12 need to be scaled way up. A12 is huge, though from Keller quotes AMD might have planned making K12 wider than Zen so after all K12 might end being pretty close to A12 design.

Now threre's A76 from ARM available which basically is about equal to best x86 designs, if someone want to compete against that with custom core it needs to be way better.

 

[naukkis]

This leads to a question of why not have a hybrid x86 & ARM core instead. Run AMD64(+up to AVX2) as legacy and ARM+SVE as modern or a hybrid mode; AMD64(GPR) + SVE(VGPR).

Because design cycle time for Aarch64 is way lower than for x86, and for aarch64 cpu front end can be much simpler. No need for predecoders to get instruction lengths, no need for micro-op cache and so on which space can be used to make core itself faster.

 

[naukkis]

Yep. 15% or so advantage quoted by Jim Keller is nowhere near enough to switch to a different ISA.

15 more performance. Someone thinks that ARM can't scale up. Here we have AMD which was making two comparable designs and ARM version was to be faster - and presumably would have also been more energy efficient at the same time. And that with first take on new architecture......

ISA has newer meant anything for best performing cpu, if AMD can make much better performing ARM cpu than intel can make x86 they will find markets for it. If changing ISA gets them few tens of percent free performance it would be nuts to not use it against such a big competitor like Intel which management are usually always trying to use suboptimal ISAs.....

 

[NTMBK]

If released 2016 as planned K12 would have been greatly competitive against Apple A10. To compete against A12 or A13 K12 need to be scaled way up. A12 is huge, though from Keller quotes AMD might have planned making K12 wider than Zen so after all K12 might end being pretty close to A12 design.

Now threre's A76 from ARM available which basically is about equal to best x86 designs, if someone want to compete against that with custom core it needs to be way better.

Except AMD still couldn't have produced a competitive mobile SoC, because they're missing all the rest of the mobile IP you need to compete with Qualcomm like a mobile-suitable GPU and a modem. And who would they sell to? Apple has their own SoCs with custom CPU. Samsung has their own SoCs with custom CPU. Huawei has their own SoCs. After that you're left with scraps of the mobile market.

 

[naukkis]

Except AMD still couldn't have produced a competitive mobile SoC, because they're missing all the rest of the mobile IP you need to compete with Qualcomm like a mobile-suitable GPU and a modem. And who would they sell to? Apple has their own SoCs with custom CPU. Samsung has their own SoCs with custom CPU. Huawei has their own SoCs. After that you're left with scraps of the mobile market.

They don't have to have all needed ip in their own portfolio, they don't have all inhouse design even for currrent x86 designs.

And potential customers? How about Qualcomm and Huawei? Their fastest Soc is build to as fast as what available from ARM, nothing prevents them licensing big cores to their SOC from AMD if they have something better available.

 

[NTMBK]

They don't have to have all needed ip in their own portfolio, they don't have all inhouse design even for currrent x86 designs.

And potential customers? How about Qualcomm and Huawei? Their fastest Soc is build to as fast as what available from ARM, nothing prevents them licensing big cores to their SOC from AMD if they have something better available.

Sure, they could license the Mali GPU, but who on earth would they license a modem from?

And as far as I'm aware, the ARM architectural license does not allow you to sell your custom CPU IP to other companies. We've not seen a single company doing this, and that's probably for a reason. Why would ARM grant you a license to compete with one of their core businesses? Even if it were possible to talk ARM into it, I'm sure it would require some hefty per-device licensing fees to make up for the Cortex A76 licenses that ARM weren't selling.

Maybe you could do some convoluted reverse licensing semicustom product, where Huawei could license their modem IP to AMD exclusively for use in devices for Huawei, AMD could license the Mali GPU IP from ARM (on top of paying ARM for the architectural license in the first place), and then AMD would "produce" the final chip (in a TSMC fab) and sell it back to Huawei... Maybe? It's a stretch though, and doesn't sound like a super profitable business for AMD.

 

[DrMrLordX]

I think that if AMD ever goes the ARM route, it will be for servers and not for mobile/tablets. Or even PCs.