Discussion: AMD Ryzen

[cdrkf]

Is the IPC claimed for per core? Or it is total die? I think it is the later. So, 6 core Zen has 40% faster ipc than 6 core excavator.

They can also increase ipc just by giving each core a dedicated FPU.

No IPC as a measurement is specific to a single thread. The thing to keep in mind though- whilst 40% single thread IPC increase sounds unrealistic, you have to remember that it isn't enough for them to catch Intel from where they are (although it will get them fairly close hopefully).

They should get better scaling than that going from Piledriver FX -> zen with same 'core' count due to Zen having double the FP units (as you say) and double the front end (full cores not shared) *and* twice the threads thanks to smt (essentially Hyperthreading, although that's an Intel trademark so I'm sure it will get a different name from AMD).

From what we've heard, octacore Zen will likely be slower than latest Intel kit in single thread (but closer) and sit in-between Intels 6 and 8 core extreme edition parts for multi thread performance. That's a much better position than they are in currently.

 

[juanrga]

Is the IPC claimed for per core? Or it is total die? I think it is the later. So, 6 core Zen has 40% faster ipc than 6 core excavator.

They can also increase ipc just by giving each core a dedicated FPU.

It is per core. If you check the original slide there is a footnote

The footnote says "for Zen x86 CPU core compared to Excavator x86 CPU core". The IPC gain already includes that each core has its own FPU.

 

[juanrga]

I've been following Zen for quite some time, and although I don't expect it to totally dominate Intel (even with Zen+) I do expect Zen to bring AMD back into the high end CPU market. Like most others have speculated I fully expect Zen's IPC to be at the performance levels of Haswell, not Sandy Bridge. Heck my FX 8370 gets single core Cinebench R15 scores very close to Sandy Bridge, overclocked past 5Ghz. 40% IPC gains over Excavator will be roughly equal to Haswell performance.

With IPC roughly equal to Haswell but running 8 cores and 16 threads Zen should be able to make a nice "splash" into the CPU market. It should be able to give 4 core 8 thread Skylake processors a run for their money, but will of course be outperformed by Intel's top tier offerings. It will be interesting to see where the stock clock setting will be for Zen FX as well as where AMD will price Zen. Considering the uphill battle AMD will have in getting back into the high end CPU market it would be nice to see Zen priced between Intel i5 and 4 core 8 thread i7 Skylake processors. At that price point Zen would be highly affordable and desirable.

Zen microarchitecture is more close to Sandy than Haswell. We know that the X4 845 (Excavator with 3.8 GHz max turbo speed) does about 95 points in single-thread Cinebench R15. 40% higher IPC than Excavator core means 95 x 1.4 = 133. The Sandy Bridge i7-2600K also have a 3.8 GHz max turbo and does 135 points. so Zen IPC is at Sandy Bridge levels. But the above score is for a hypothetical Zen that hits 3.8Ghz. Zen is expected to have lower clocks, I guess 3.5GHz. Zen would have to hit around 5GHz to hit the single thread performance of a quad-core Skylake.

About pricing I expect octo-core Zen to be somewhat priced as Broadwell six-core.

 

[juanrga]

Just to place things in perspective

40% on top of 74.60 gives 104.44 which is below the Sandy Bridge Score. Skylake is a good 25% ahead and if we consider clocks then the total gap between Skylake and Zen can be something as big as 50%!!!!! We are back to Piledriver Haswell gap levels!!!!

 

[gamerk316]

The footnote says "for Zen x86 CPU core compared to Excavator x86 CPU core". The IPC gain already includes that each core has its own FPU.

Are we counting the SMT core as part of that core? Because that makes a HUGE difference in performance.

 

[ComputerSecurityGuy]

Since SMT adds logical cores, a core to core comparison usually would not be on logical cores.

 

[cdrkf]

The footnote says "for Zen x86 CPU core compared to Excavator x86 CPU core". The IPC gain already includes that each core has its own FPU.

Are we counting the SMT core as part of that core? Because that makes a HUGE difference in performance.

SMT would only count towards the multi thread performance- Zen should be a massive boost in Multi thread support over same core count Piledriver as it has double the front end, double the threads, double the FP units and more int units per core and doesn't have the CMT scaling penalty that FX '8' core parts do.

See further up for my calcs on what it should score in multi thread- essentially Zen should offer a decent step up from Intels quad core i7 in multi thread scenarios (thanks to better scaling and more cores). It will be behind in single thread as IPC is still behind and clocks will be lower.

What I think is encouraging is we are in fact back to Phenom II X6 vs first gen i7 (and NOT Piledriver vs Haswell like some suggest). The Phenom II X6 was actually a really good option at the time if you had software that could use the threads (I got one for rendering for example).

People looking for a high performance processor will have a choice- better multi thread performance (Zen), or better single thread performance (Skylake) for similar money. That is actually something worth consideration, rather than the current situation that the FX 8XXX parts excel at nothing at all, and can only offer cheaper prices vs much smaller intel Parts.

The power consumption also sounds fairly good- I think Zen looks like it could be quite a good option for servers in the high core count configurations- it might not be faster than Intel per core but sounds like the perf/w should be on the money and the overall throughput will be good thanks to the high thread count. The money is in servers anyway- and given AMD's current market share the only way is up (I mean a few % points swing to AMD could make all the difference).

 

[juanrga]

The footnote says "for Zen x86 CPU core compared to Excavator x86 CPU core". The IPC gain already includes that each core has its own FPU.

Are we counting the SMT core as part of that core? Because that makes a HUGE difference in performance.

That 40% is single thread on Zen vs single thread on Excavator. For multiple threads the IPC gap is about 70% because you have to count the SMT gain on Zen. For Piledriver you have to count the core gain Piledriver-->Excavator and the CMT module penalty on Piledriver

2 threads on Zen vs 2 threads on Excavator module is 1.4 * 1.2 = 1.7
2 threads on Zen vs 2 threads on Excavator module is 1.5 * 1.2 * (2/1.8) = 2

 

[ComputerSecurityGuy]

We tend to assume a 20% gain from SMT, and the same scaling. Of course, it could be more. IBM got very good gains out of SMT in POWER8.

 

[-Fran-]

But the Power8 is a behemoth of a CPU though. It has specific circuitry to improve the SMT logic even further from what you would imagine. Anandtech's wrote an excellent article about it not so long ago, so go and read it.

I guess we should start seeing more numbers for Zen in the coming weeks. I bet some engy samples were put to some benchmarks in public places. We just need to go and find them XD

Cheers!

 

[gamerk316]

If we get benchies, we can start to compute relative IPC/Performance gain across varying benchmarks. You all know I'll do the math if I get the numbers 😀

 

[juanrga]

We tend to assume a 20% gain from SMT, and the same scaling. Of course, it could be more. IBM got very good gains out of SMT in POWER8.

20% is what I have been assuming for Zen. IBM has much more experience on implementing SMT since the Power5. Also Power8 is a wider core than Zen, which means that Power8 core has lots of execution units ready to be used by other threads. Recall Power8 can run up to 8 threads simultaneously.

 

[gamerk316]

I would like to remind everyone IPC is not a constant number; it varies by the workload you are trying to accomplish. The BD architecture starved FP workloads, but Integer workloads had the execution resources to scale. Point being, Zen is going to see a much larger boost in FP performance then Integer performance. So depending on workload, you could see performance increases range from small (Integer) to large (Floating-Point) improvements.

 

[juanrga]

Indeed, the observed IPC depends on the workload, that is why I am using the 40% IPC gain only on CB scores.

 

[Oliver_21]

Is this going on

My advice:

Don't expect anything until zen comes out

 

[Johnpombrio]

How close are engineering sample chips clock speeds match the finished products? Years ago, there used to be a substantial difference but lately it seems to be pretty much the same between engineering samples and the shipping product (at least as far as Intel). Anyone know a web page that can show why newer engineering sample chips are deliberately down-clocked? I cannot find a single page that is not just someone's "common sense" opinion.
EDIT Never mind. The Skylake engineering samples had 2.2 GHz ,2.9 GHz boost while the Skylake 6700K chip runs at 4.0 GHz -4.2 GHz boost. The engineering sample clocks are NOT what the finished chips will be running at so take the Zen clock speeds as undefined until the real chips are launched (or leaked!)

 

[8350rocks]

How close are engineering sample chips clock speeds match the finished products? Years ago, there used to be a substantial difference but lately it seems to be pretty much the same between engineering samples and the shipping product (at least as far as Intel). Anyone know a web page that can show why newer engineering sample chips are deliberately down-clocked? I cannot find a single page that is not just someone's "common sense" opinion.
EDIT Never mind. The Skylake engineering samples had 2.2 GHz ,2.9 GHz boost while the Skylake 6700K chip runs at 4.0 GHz -4.2 GHz boost. The engineering sample clocks are NOT what the finished chips will be running at so take the Zen clock speeds as undefined until the real chips are launched (or leaked!)

It will be difficult to tell where they end up in terms of clocks...

 

[8350rocks]

Given history, yes. AMD has always been a "first try fail, second try recover" with their CPUs. I just hope it doesn't happen with Zen and they break out of that cycle.

Cheers!

Original Athlon was "first try blow the doors off intel across the board and leave them reeling for the next 5-8 years".

So, that was not always the case...

 

[-Fran-]

Original Athlon was "first try blow the doors off intel across the board and leave them reeling for the next 5-8 years".

So, that was not always the case...

The original Athlon core you mean? Uhm... I had to look for information, cause I only have memory from the T-Bird and K6-III era, haha. It was faster than the P3 (original) and then the T-Bird increased the lead a bit more. But previous incarnations (K6 line for example) wasn't better than the Intel equivalent at launch. K5 and K6 needed 1 or 2 revisions for parity. Specially the K6, since it kept on using the old socket 7 and moved too late to slot A (a bad idea to copy Intel's stupid slot 1 XD).

Cheers!

 

[Cazalan]

Slots were necessary at the time. That is where they had to put the caches.

Spoiler

 

[-Fran-]

But the Socket7 CPUs had the cache on the side (like RAM), which was the thing slots were trying to solve... That did not need solving XD

Well, maybe MoBo complexity was lessened and latency improved, but cooling for those CPUs was horrible. I still remember the need for "socket adapters" for some P3 models using Socket370! OC'ing those was... An interesting endeavor. That is why I did not like Slots for CPUs xP

Cheers!

 

[cdrkf]

But the Socket7 CPUs had the cache on the side (like RAM), which was the thing slots were trying to solve... That did not need solving XD

Well, maybe MoBo complexity was lessened and latency improved, but cooling for those CPUs was horrible. I still remember the need for "socket adapters" for some P3 models using Socket370! OC'ing those was... An interesting endeavor. That is why I did not like Slots for CPUs xP

Cheers!

The CPU 8350 was referring to was the Athlon 64- look at the reviews- brand new cpu, 64 bit capable, integrated IMC and faster than anything Intel had at a much lower clock speed and lower power consumption.

The original Athlon was also a great cpu. As for the Socket 7 based K6-2, that was a really good gaming cpu, as it gave people on existing Socket 7 setups a superb upgrade path. It allowed a very flexible upgrade approach- get a good socket 7 motherboard, keep your ram, keep your existing CPU... I remember keeping a P200 MMX and moving it to an up to date board (that gave a noticeable performance boost). Then later I was able to swap out the cpu. I initially tried a Cyrix 686 300, which was rubbish (slower than the P200 most of the time) so in the end went with a K6 -2 500. That machine lasted me quite a long time. The K6-2 wasn't the outright performance champ that gen (the Pentium II and III were faster most of the time) *but* it offered people with older Socket 7 systems from either camp an upgrade option- whereas going Intel meant an entire new platform, and back then desktop machines were overall somewhat more expensive than they are today.

 

[-Fran-]

But that is something else; the point I made is AMD has needed 2nd versions of CPUs through out their history to reach parity with Intel.

The only CPUs that did not need parity were Athlons up till Toledo, yes.

My point is very simple though: don't be surprised if AMD needs a second version of Zen to reach parity. Although, everything already points it won't be equal to neither Intel CPU on the market anyway 😛

Cheers!

EDIT: Typo.

 

[8350rocks]

Original Athlon was "first try blow the doors off intel across the board and leave them reeling for the next 5-8 years".

So, that was not always the case...

The original Athlon core you mean? Uhm... I had to look for information, cause I only have memory from the T-Bird and K6-III era, haha. It was faster than the P3 (original) and then the T-Bird increased the lead a bit more. But previous incarnations (K6 line for example) wasn't better than the Intel equivalent at launch. K5 and K6 needed 1 or 2 revisions for parity. Specially the K6, since it kept on using the old socket 7 and moved too late to slot A (a bad idea to copy Intel's stupid slot 1 XD).

Cheers!

K7/K8

Yeah...K62 and K63 were good options for older rigs to upgrade.

 

[Ecky]

I'm going to steal a quote from a poster on Anandtech:

https://forums.anandtech.com/threads/first-summit-ridge-zen-benchmarks.2482739/

First Summit Ridge (Zen) Benchmarks

Straight from Ashes of the Singularity's CPU benchmark:

#COMPARISON 1

- Summit Ridge 8C/16T ES - 2.8-3.2 GHz (2016/2017 Zen)
Average: 58.9 FPS
Normal batch: 65.8 FPS
Medium batch: 62.8 FPS
Heavy batch: 50.5 FPS

http://www.ashesofthesingularity.com/metaverse#/personas/50728c1c-7384-447b-99a9-1a7c330cce99/match-details/bfba4b4a-4b1e-4ab3-8f2f-2375321ea68b

- Core i7-980 6C/12T - 3.33 GHz (2010 Westmere)
Average: 58.6 FPS
Normal batch: 65.2 FPS
Medium batch: 59.7 FPS
Heavy batch: 52.3 FPS

http://www.ashesofthesingularity.com/metaverse#/personas/388c81f2-8199-4446-b54b-675f5a0aa833/match-details/45e74e9d-4b0c-4557-8225-a3b8664dd4a8

- Core i7-4770 4C/8T 3.4-3.9 GHz (2013 Haswell)
Average: 66.0 FPS
Normal batch: 74.5 FPS
Medium batch: 69.5 FPS
Heavy batch: 56.6 FPS

http://www.ashesofthesingularity.com/metaverse#/personas/2fb047a6-39a4-4039-8156-ee759c81bd89/match-details/4baad586-6271-42bd-84c5-6884cfb3341d

- Core i7-6700K 4C/8T - 4.0 GHz (2015 Skylake)
Average: 107.3 FPS
Normal batch: 125.7 FPS
Medium batch: 113.8 FPS
Heavy batch: 89.2 FPS

http://www.ashesofthesingularity.com/metaverse#/personas/045e4b2f-dd7b-4187-aafa-cde9dfe20eb6/match-details/6096dd00-7e11-4368-b588-97413922f5c7

The Haswell score is 100% at stock and based on the latest version of the benchmark included in the search engine (1.24.20823.0). The benchmark likes cores/thread and scales with more than 4C/8T:



3.0 GHz Haswell-E beating 4.0 GHz Skylake-S.


#COMPARISON 2

- Summit Ridge 8C/16T ES - 2.8-3.2 GHz (2016/2017 Zen)
Average: 31.5 FPS
Normal batch: 36.5 FPS
Medium batch: 33.8 FPS
Heavy batch: 26.2 FPS

- Core i5-4670K 4C/4T 3.4-3.8 GHz (2013 Haswell)
Average: 52.6 FPS
Normal batch: 56.9 FPS
Medium batch: 54.4 FPS
Heavy batch: 47.5 FPS

https://forums.anandtech.com/threads/first-summit-ridge-zen-benchmarks.2482739/page-4#post-38414245