Discussion: AMD Ryzen

[cdrkf]

@gamerk, I've heard this argument before, but I've yet to see any tangible benefits.


If what you say is true, why are arm v8 cpus, with their dedicated ground up 64 bit architecture not vastly superior to x86-64? What great benefits do you envisage? most software only use the 64 bit extensions to increase memory capacity as 32 bit is sufficient for the majority of calculations....

 

[-Fran-]

The only tangible and real benefit of a tighter ISA is less bloated binaries. Also, less complex compilers.

In terms of actual performance, no matter the bloat the ISA has, the compiler is the one that actually decides how to create the binary and distribute the ASM instructions and all that.

Maybe, and just maybe, on time critical applications it does make a tangible difference, where a nano second might actually matter (think the stock market) when processing.

Cheers!

 

[jaymc]

Stuck with x86-64 for the next decade or two... jesus seems like a long time in the computer world... hell in any world.

Really that does seem a shame. I might not live to see 128bit CPU's in the end user market

 

[gamerk316]

@gamerk, I've heard this argument before, but I've yet to see any tangible benefits.


If what you say is true, why are arm v8 cpus, with their dedicated ground up 64 bit architecture not vastly superior to x86-64? What great benefits do you envisage? most software only use the 64 bit extensions to increase memory capacity as 32 bit is sufficient for the majority of calculations....

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

 

[juanrga]

@gamerk, I've heard this argument before, but I've yet to see any tangible benefits.

If what you say is true, why are arm v8 cpus, with their dedicated ground up 64 bit architecture not vastly superior to x86-64? What great benefits do you envisage? most software only use the 64 bit extensions to increase memory capacity as 32 bit is sufficient for the majority of calculations....

Gamerk was talking about Itanium, which used a scalable ISA named EPIC. This is radically different to ARM64, x64, Power, MIPS64, SPARC, RISC-V, or any other nonscalable 64bit ISA.

Regarding ARM64, Jim Keller gave a talk where he explained the advantages of the ARM64 ISA over the x64 ISA. He also mentioned that ARM64 was giving K12 an advantage over Zen. Some time ago I estimated that K12 would be about 20--30% faster than Zen.

 

[juanrga]

The only tangible and real benefit of a tighter ISA is less bloated binaries. Also, less complex compilers.

In terms of actual performance, no matter the bloat the ISA has, the compiler is the one that actually decides how to create the binary and distribute the ASM instructions and all that.

No. The less bloated ISA has a direct effect on the hardware and its performance. A simpler ISA without opcode bloat helps to reduce the decoder, which increases performance and reduces power consumption. A RISC ISA avoids the translation layer CISC-->uops and the need for a uop cache, saving again power and increasing performance. Native three-operand instructions can reduce logic pressure, by saving lots in the OOE window and buffers...

 

[juanrga]

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

ARM64 has been designed for high performance. And microarchitectures like K12 or Vulcan show it (both at Haswell level). It is also expected that X-Gene3 will be in this camp.

It is also worth to mention that the first exascale supercomputer announced will use ARM64. Fujitsu that currently builds supercomputers based in SPARC ISA has replaced SPARC by ARM64 for the future.

 

[gamerk316]

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

ARM64 has been designed for high performance. And microarchitectures like K12 or Vulcan show it (both at Haswell level). It is also expected that X-Gene3 will be in this camp.

It is also worth to mention that the first exascale supercomputer announced will use ARM64. Fujitsu that currently builds supercomputers based in SPARC ISA has replaced SPARC by ARM64 for the future.

Call me when someone actually produces a CPU that outperforms x86. Because so far: Hasn't happened. People WANT it to happen given x86 is never going to be power efficient (see: Atom), but what will happen is a trade-off: Power for Performance. This is why the embedded world I live in still uses PowerPC 7xx based CPUs; x86 just doesn't cut it.

My old boss has a statement: "Power, Performance, Price; pick two of three". Same logic applies to ARM64: It's never going to be faster then x86 without drawing a ton more power or costing a ton more. The architecture simply isn't designed for that type of workload.

 

[cdrkf]

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

ARM64 has been designed for high performance. And microarchitectures like K12 or Vulcan show it (both at Haswell level). It is also expected that X-Gene3 will be in this camp.

It is also worth to mention that the first exascale supercomputer announced will use ARM64. Fujitsu that currently builds supercomputers based in SPARC ISA has replaced SPARC by ARM64 for the future.

Call me when someone actually produces a CPU that outperforms x86. Because so far: Hasn't happened. People WANT it to happen given x86 is never going to be power efficient (see: Atom), but what will happen is a trade-off: Power for Performance. This is why the embedded world I live in still uses PowerPC 7xx based CPUs; x86 just doesn't cut it.

My old boss has a statement: "Power, Performance, Price; pick two of three". Same logic applies to ARM64: It's never going to be faster then x86 without drawing a ton more power or costing a ton more. The architecture simply isn't designed for that type of workload.

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

It will be fascinating to see if AMD do release K12- as that is rumored to be a very similar hardware design to Zen- then we should have as close to an apples to apples comparison of the difference the ISA makes as we're likely to get. Juan thinks 20- 30%, personally I think it will be smaller than that. I would expect the two to offer comparable performance though (honestly if K12 was *that* much better than Zen I think AMD would have been pushing it much harder as it would potentially put then in front of Intel given they are on pretty much equal manufacturing nodes- the fact K12 has dropped off the radar suggests it doesn't offer much in the way of tangible benefits over Zen so better to stick with x86-64 for compatibility).

 

[jdwii]

@gamerk, I've heard this argument before, but I've yet to see any tangible benefits.


If what you say is true, why are arm v8 cpus, with their dedicated ground up 64 bit architecture not vastly superior to x86-64? What great benefits do you envisage? most software only use the 64 bit extensions to increase memory capacity as 32 bit is sufficient for the majority of calculations....

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

I'd say Arm has a lot of room to grow where X86 has basically stopped. For example i could easily see Arm being in a next gen console now a days over X86 over it providing good enough performance at a TDP X86 can't touch. Really hoping Amd didn't cancel K12.

Edit but i'm kind of with Gamer when it comes to being skeptical, also wish Intel would just make it to i mean they would probably be the best at it they have the best fabrication for it

 

[-Fran-]

The only tangible and real benefit of a tighter ISA is less bloated binaries. Also, less complex compilers.

In terms of actual performance, no matter the bloat the ISA has, the compiler is the one that actually decides how to create the binary and distribute the ASM instructions and all that.

No. The less bloated ISA has a direct effect on the hardware and its performance. A simpler ISA without opcode bloat helps to reduce the decoder, which increases performance and reduces power consumption. A RISC ISA avoids the translation layer CISC-->uops and the need for a uop cache, saving again power and increasing performance. Native three-operand instructions can reduce logic pressure, by saving lots in the OOE window and buffers...

That is true and I forgot the bloat in the ISA does come with hardware bloat associated.

In my defense, there's no way your ISA can be successful and it stays "slim" for a long time.

Cheers!

 

[juanrga]

ARM64 doesn't see any advantages because ARM is to x86 what a Prius is to a Mustang. It's got almost none of the extra bits Intel has stuffed into x86 over the years to improve performance. ARM is focused almost exclusively on low power, at the expense of performance, and frankly will scale up a LOT worse then x86 will.

The better comparison is against other CPUs designed to be big; POWER and SPARC in particular. Even then, Intel manages to stay even due in large part to having better foundries, which hides a lot of the architectures inefficiencies.

ARM64 has been designed for high performance. And microarchitectures like K12 or Vulcan show it (both at Haswell level). It is also expected that X-Gene3 will be in this camp.

It is also worth to mention that the first exascale supercomputer announced will use ARM64. Fujitsu that currently builds supercomputers based in SPARC ISA has replaced SPARC by ARM64 for the future.

Call me when someone actually produces a CPU that outperforms x86. Because so far: Hasn't happened. People WANT it to happen given x86 is never going to be power efficient (see: Atom), but what will happen is a trade-off: Power for Performance. This is why the embedded world I live in still uses PowerPC 7xx based CPUs; x86 just doesn't cut it.

My old boss has a statement: "Power, Performance, Price; pick two of three". Same logic applies to ARM64: It's never going to be faster then x86 without drawing a ton more power or costing a ton more. The architecture simply isn't designed for that type of workload.

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

It will be fascinating to see if AMD do release K12- as that is rumored to be a very similar hardware design to Zen- then we should have as close to an apples to apples comparison of the difference the ISA makes as we're likely to get. Juan thinks 20- 30%, personally I think it will be smaller than that. I would expect the two to offer comparable performance though (honestly if K12 was *that* much better than Zen I think AMD would have been pushing it much harder as it would potentially put then in front of Intel given they are on pretty much equal manufacturing nodes- the fact K12 has dropped off the radar suggests it doesn't offer much in the way of tangible benefits over Zen so better to stick with x86-64 for compatibility).

My 20--30% estimation for the gap could be wrong. Once I asked about the gap to David Kanter and he said something as maybe 10%.

I think AMD's de-emphasization of the K12 in favor of Zen has two causes: (i) open software stack is not still there. Other server companies are developing custom software solutions, but AMD doesn't have the resources for that, and (ii) the ARM competition is stronger than AMD expected, specially after Qualcomm announced its server SoC.

Notice that in the x86 server camp AMD only has to fight Intel Skylake. Everyone going to purchase an x86 server in 2017 will have to choose between Skylake Xeons and Zen Opterons.

In the ARM server camp, AMD fights Cavium, APM, Broadcom, Qualcomm,... Everyone going to purchase an ARM server in 2017 will have freedom to choose between a broad spectrum of products. For instance here is the XGene3 I mentioned above

 

[juanrga]

My old boss has a statement: "Power, Performance, Price; pick two of three". Same logic applies to ARM64: It's never going to be faster then x86 without drawing a ton more power or costing a ton more. The architecture simply isn't designed for that type of workload.

That logic doesn't always apply. I have benchmarks of prototype HPC-class hardware outperforming Intel Xeons that cost more and consume more, and I have similar internal data from Paypal comparing their ARM servers to similar performance x86 servers. I cannot share this data, but I found this table online, which is not very good but can give an idea

 

[turkey3_scratch]

I cannot share this data, but I found this table online, which is not very good but can give an idea

Well, you have just shared the results, though, albeit indirectly. 😉 How do you know all this stuff anyway? Who do you work for?

 

[juanrga]

I cannot share this data, but I found this table online, which is not very good but can give an idea

Well, you have just shared the results, though, albeit indirectly. 😉 How do you know all this stuff anyway? Who do you work for?

I didn't share anything. That aren't the benchmarks I have. I only found online a table that shows one example of how one can have similar performance at lower costs and lower power.

 

[turkey3_scratch]

I cannot share this data, but I found this table online, which is not very good but can give an idea

Well, you have just shared the results, though, albeit indirectly. 😉 How do you know all this stuff anyway? Who do you work for?

I didn't share anything. That aren't the benchmarks I have. I only found online a table that shows one example of how one can have similar performance at lower costs and lower power.

My reference was to you sharing that the HPC-class hardware outperforms the Xeons.

 

[-Fran-]

There's a saying (in Spanish) that goes "tell about the miracle, not the saint behind it", trukey3_scratch 😛

In any case, for "simple" operations, I do remember RISC-type uArchs do fare better. For complex architectures (of software), with a lot of mixed workloads, CISC (and in turn, X86) performs way better. Now, that is from University times (~2003) information. That is independent of the ISA behind it according to what I remember. That also could be from data from that time, where the most impressive RISC type Arch was PowerPC and was actually faster in certain workloads with an ISA as "complex" as X86 of the time.

Cheers!

 

[gamerk316]

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

ARM won't scale up without adding all the little bits x86 already has, but without the 40+ years of tuning. ARM will easily win the power-efficiency battle, but not performance.

 

[juanrga]

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

ARM won't scale up without adding all the little bits x86 already has, but without the 40+ years of tuning. ARM will easily win the power-efficiency battle, but not performance.

Jim Keller gave a talk stating the advantages of ARM64 over x64 and why K12 will be a more powerful core than Zen. I mentioned a high-performance server-class core like Vulcan (which runs four threads per core). I also mentioned before that the first announced exascale supercomputer uses ARM64. I guess it is worth to give some more info about this in order to kill the myth that ARM is only for mobile or low power...

Piledriver core: 8 FLOP/core
Zen core: 16 FLOP/core
K12 core: 16 FLOP/core
Broadwell: 32 FLOP/core
KNL Phi: 64 FLOP/core
Fujitsu ARM: 64 FLOP/core

And Fujitsu is only using 1/4x the maximum 2048bit width defined on ARM SVE

It is time to kill the myth that ARM is not for performance. But I will not insist more, because this thread is about Zen.

 

[jimmysmitty]

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

ARM won't scale up without adding all the little bits x86 already has, but without the 40+ years of tuning. ARM will easily win the power-efficiency battle, but not performance.

Jim Keller gave a talk stating the advantages of ARM64 over x64 and why K12 will be a more powerful core than Zen. I mentioned a high-performance server-class core like Vulcan (which runs four threads per core). I also mentioned before that the first announced exascale supercomputer uses ARM64. I guess it is worth to give some more info about this in order to kill the myth that ARM is only for mobile or low power...

Piledriver core: 8 FLOP/core
Zen core: 16 FLOP/core
K12 core: 16 FLOP/core
Broadwell: 32 FLOP/core
KNL Phi: 64 FLOP/core
Fujitsu ARM: 64 FLOP/core

And Fujitsu is only using 1/4x the maximum 2048bit width defined on ARM SVE

It is time to kill the myth that ARM is not for performance. But I will not insist more, because this thread is about Zen.

ARM can be for performance. The thing is that their major money tends to lie in power efficiency.

Just an example but Intels Atom has been a superior performing chip vs ARM for quite a while. Their downfall and why Atom isn't leading the mobile market is the power efficiency advantage ARM has.

That said, when ARM is powerful its power efficiency tends to drop and it then becomes a factor of a lot of other parts rather than that.

On the topic, is Zen really only half the GFLOPs of Broadwell?

 

[juanrga]

On the topic, is Zen really only half the GFLOPs of Broadwell?

Yes,

BD/PD/SR/XV: 2 x 128bit per module --> 8 FLOP/core
Zen: 2 x 128bit per core --> 16 FLOP/core
Zen+: 2 x 128bit per core --> 16 FLOP/core*

SNB/IVB: 2 x 128bit per core --> 16 FLOP/core
HSW/BDW: 2 x 256bit per core --> 32 FLOP/core

This is the reason why I said that Blender only uses one half of the FPU units on Broadwell. Blender uses 128bit vectors. Broadwell supports 256bit vectors.

* My own estimation.

 

[cdrkf]

I think you are mixing up the *current ARM implementations* vs the *capability of the ISA*. Two very different things- I don't see how ARM64 cannot be made to be just as fast as a modern x86 processor if you build the two processors with the same execution resources and so on.

ARM won't scale up without adding all the little bits x86 already has, but without the 40+ years of tuning. ARM will easily win the power-efficiency battle, but not performance.

Jim Keller gave a talk stating the advantages of ARM64 over x64 and why K12 will be a more powerful core than Zen. I mentioned a high-performance server-class core like Vulcan (which runs four threads per core). I also mentioned before that the first announced exascale supercomputer uses ARM64. I guess it is worth to give some more info about this in order to kill the myth that ARM is only for mobile or low power...

Piledriver core: 8 FLOP/core
Zen core: 16 FLOP/core
K12 core: 16 FLOP/core
Broadwell: 32 FLOP/core
KNL Phi: 64 FLOP/core
Fujitsu ARM: 64 FLOP/core

And Fujitsu is only using 1/4x the maximum 2048bit width defined on ARM SVE

It is time to kill the myth that ARM is not for performance. But I will not insist more, because this thread is about Zen.

ARM can be for performance. The thing is that their major money tends to lie in power efficiency.

Just an example but Intels Atom has been a superior performing chip vs ARM for quite a while. Their downfall and why Atom isn't leading the mobile market is the power efficiency advantage ARM has.

That said, when ARM is powerful its power efficiency tends to drop and it then becomes a factor of a lot of other parts rather than that.

On the topic, is Zen really only half the GFLOPs of Broadwell?

Remember flops are measuring the floating point performance- and Intel has much wider 512bit fp units since haswell I believe. That said very little software makes use of that. In int math zen should be similar to Intel...

 

[juanrga]

Remember flops are measuring the floating point performance- and Intel has much wider 512bit fp units since haswell I believe. That said very little software makes use of that. In int math zen should be similar to Intel...

As stated above Haswell and Broadwell have 256bit units. Only the Xeon Phi line has 512bit units.

Very little Windows desktop software uses AVX2, but its use is more extended on servers, because it is not only for fp math. Precisely AVX2 introduced lots of new 256bit instructions for integer math. I guess this is one of the reasons why Zen has four ALUs: 64bit x 4 = 256bit.

 

[daerohn]

As many people have stated here AMD has falle short in performance race once again. I used to have a phenom II a while back, but my mainboard broke down. So I switched to FX 8320. Overall it is a good CPU, no problem till now however I am also using Intel chips and they are way better. I have been waiting for the Zen for at least 1 year now. And in the beginning I was waiting a very good CPU from AMD. Yet even there are few months left for the release, I could not find a throughout test on the chip, even AMD dodges the requests with some very uninforming slides. If they had a tremendous chip in their pocket, they would have been advertising it all around with big events, when taken in to account their financial situation this would not be unexpected. I even can not find a single web site with ES. Hell it took too long for AMD to come up with a good chip and it seems to be outdated even before it is released.

Still I wish and hope AMD to survive. I remember times Intel was the only option. We had to pay thousand of dollars for a simple computer. back than a middle range computer would cost around 3k $ while a high and was aroun 5k's. It would seem ridiculus now thanks to AMD and Cyrix. I remember paying around 3400 bucks for a Intel 486 DX 33 with 4MB's of ram, 200MB of HDD, 1MB of video card. No sound card, no CD-ROM 14" Monitor. I also remember creative kits with CD-ROM and Soundcard to cost around 500$. Long long ago... I hope AMD would act quick to develop Zen+ and beyond. As they are running out of time and money fast.

 

[jdwii]

Spoiler

As many people have stated here AMD has falle short in performance race once again. I used to have a phenom II a while back, but my mainboard broke down. So I switched to FX 8320. Overall it is a good CPU, no problem till now however I am also using Intel chips and they are way better. I have been waiting for the Zen for at least 1 year now. And in the beginning I was waiting a very good CPU from AMD. Yet even there are few months left for the release, I could not find a throughout test on the chip, even AMD dodges the requests with some very uninforming slides. If they had a tremendous chip in their pocket, they would have been advertising it all around with big events, when taken in to account their financial situation this would not be unexpected. I even can not find a single web site with ES. Hell it took too long for AMD to come up with a good chip and it seems to be outdated even before it is released.

Still I wish and hope AMD to survive. I remember times Intel was the only option. We had to pay thousand of dollars for a simple computer. back than a middle range computer would cost around 3k $ while a high and was aroun 5k's. It would seem ridiculus now thanks to AMD and Cyrix. I remember paying around 3400 bucks for a Intel 486 DX 33 with 4MB's of ram, 200MB of HDD, 1MB of video card. No sound card, no CD-ROM 14" Monitor. I also remember creative kits with CD-ROM and Soundcard to cost around 500$. Long long ago... I hope AMD would act quick to develop Zen+ and beyond. As they are running out of time and money fast.

I agree i actually just got done reading some old reviews on the K5, K6, K6-2, K6-3, and yes the K7-K8 that did some butt kicking to Intel(K7/K8). But its 2016 now a 8 core FX (sadly so sadly they ruined their FX name) is barely competitive to a I3 skylake CPU. Its only 250$ for a I5 that is easily overclockable to 4.0+ i really see no reason to wait for Zen unless one wants Amd but that is more emotional then logical. I guess i get it to a point. I used to be there myself.