AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

[AtenRa]

If it's for a laptop then a lower tdp, more powerful igpu with better driver support. If it's for a desktop then price and not being tied to Intel's platform limitations.

For Laptops even Carrizo is better than Sandy/Ivy.

 

[bjt2]

AMD's better SMT? Do you really think AMD's first crack at it, after mocking Intel for using the technique back in 2011/2012, will be superior to a 4th gen implementation in Broadwell?

Every FP code requires also some INT code, for jump, loop, load/store. AMD has 4 INT PLUS 4 FP ports. INTEL has 4 SHARED ports for INT and FP.

While for ST code these can be enough (i don't expect AMD ST be superior to INTEL's for this fact alone), for MT code this can be not enough...

Here comes AMD's advantage.

Anyway, given blender results, either AMD's ST is superior or AMD's SMT... Your choice... No way otherwise...

 

[cdimauro]

My hypothesis is that ST is inferior and that AMD's better SMT makes up for the differences plus that 2% seen in blender test...

You cannot pick Blender as THE benchmark to measure how SMT works on both platforms. You need more, heterogeneous tests to be performed.

BTW, I want to stress one thing: the Blender test has only "measured" (with quotes because nobody knows exactly the platforms' settings, and nobody actually can reproduce the results) the IPC.

So, it's not a measure of possible performances, since there are missing two important things: frequencies AND Turbo (which was disabled, AFAIK). Don't underestimate the influence of Turbo on the real performances.

 

[AtenRa]

You cannot pick Blender as THE benchmark to measure how SMT works on both platforms. You need more, heterogeneous tests to be performed.

BTW, I want to stress one thing: the Blender test has only "measured" (with quotes because nobody knows exactly the platforms' settings, and nobody actually can reproduce the results) the IPC.

So, it's not a measure of possible performances, since there are missing two important things: frequencies AND Turbo (which was disabled, AFAIK). Don't underestimate the influence of Turbo on the real performances.

It only measured the CPU Throughput capability at fixed 3GHz, not the IPC.

 

[bjt2]

You cannot pick Blender as THE benchmark to measure how SMT works on both platforms. You need more, heterogeneous tests to be performed.

BTW, I want to stress one thing: the Blender test has only "measured" (with quotes because nobody knows exactly the platforms' settings, and nobody actually can reproduce the results) the IPC.

So, it's not a measure of possible performances, since there are missing two important things: frequencies AND Turbo (which was disabled, AFAIK). Don't underestimate the influence of Turbo on the real performances.

Yes, but we both agreed we were talking about IPC. And given AVFS and many power saving features, i think that Zen can even turbo more than INTEL...
Regarding the sole benchmark... THis is what we have... Yes, we have also AotS that shows similar outcome, but we'll see...

 

[bjt2]

It only measured the CPU Throughput capability at fixed 3GHz, not the IPC.

Having fixed the frequency, the IPC is only a division ahead...

 

[cdimauro]

Every FP code requires also some INT code, for jump, loop, load/store. AMD has 4 INT PLUS 4 FP ports. INTEL has 4 SHARED ports for INT and FP.

While for ST code these can be enough (i don't expect AMD ST be superior to INTEL's for this fact alone), for MT code this can be not enough...

You forgot the maximum of 6 uops/cycle that the Micro-op cache can deliver to the schedulers, which limits the usage of the 10 ports/units.

Here comes AMD's advantage.

2% in ONE test?

Anyway, given blender results, either AMD's ST is superior or AMD's SMT... Your choice... No way otherwise...

As I said before, it's only ONE test. And it tested only the IPC: not real-world performances.

Let's wait for more tests, with final frequencies AND Turbo activated.

 

[Arzachel]

Every FP code requires also some INT code, for jump, loop, load/store. AMD has 4 INT PLUS 4 FP ports. INTEL has 4 SHARED ports for INT and FP.

While for ST code these can be enough (i don't expect AMD ST be superior to INTEL's for this fact alone), for MT code this can be not enough...

Here comes AMD's advantage.

Anyway, given blender results, either AMD's ST is superior or AMD's SMT... Your choice... No way otherwise...

You forgot to qualify that with "in this specific benchmark". It is promising but far from conclusive evidence.

 

[AtenRa]

Having fixed the frequency, the IPC is only a division ahead...

Only if they would run the benchmark with a Single Core Single Thread. The test was performed with all 8Cores 16Threads, that measured the entire CPU Throughput (CPU MT Scaling, CPU IPC, CPU SMT etc etc).

 

[cdimauro]

It only measured the CPU Throughput capability at fixed 3GHz, not the IPC.

With fixed frequency and turbo disabled, your target is the IPC.

Yes, but we both agreed we were talking about IPC. And given AVFS and many power saving features, i think that Zen can even turbo more than INTEL...

Well, Intel has also a very long tradition of power-saving techniques and Turbo technologies. So, let me doubt a bit of what you stated.

Regarding the sole benchmark... THis is what we have... Yes, we have also AotS that shows similar outcome, but we'll see...

Ehm... no. AotS didn't show a similar outcome, and BTW it didn't measured the IPC.

Only if they would run the benchmark with a Single Core Single Thread. The test was performed with all 8Cores 16Threads, that measured the entire CPU Throughput (CPU MT Scaling, CPU IPC, CPU SMT etc etc).

Nevertheless, you're still measuring the IPC. Only one more division required...

 

[lolfail9001]

Nevertheless, you're still measuring the IPC. Only one more division required...

Division by what? You don't know MT scaling, you don't know SMT scaling and without those, you can't divide.

 

[cdimauro]

Division by what? You don't know MT scaling, you don't know SMT scaling and without those, you can't divide.

take a look at the definition of IPC

 

[lolfail9001]

take a look at the definition of IPC

In computer architecture, instructions per cycle (IPC) is one aspect of a processor's performance: the average number of instructions executed for each clock cycle.

Sure, now you may ignore MT/SMT scaling but you don't know amount of instructions executed. What now?

 

[coercitiv]

take a look at the definition of IPC

You were asked a different question: since you consider the test was indicative of IPC as the appropriate values are one division away, please develop on this further and explain how division is to be applied. Otherwise having people "read up" on what IPC means is hardly helpful for the purpose of this conversation.

 

[Glo.]

People who are hoping for a quad core Zen with Ivy IPC, what exactly has stopped you from buying an IVY BRIDGE CPU???

Nothing. I do have Ivy Bridge CPU.

But Ivy Bridge level of performance clock-for-clock will mean miracles for Zen APUs. And that is the most important part here. I would not buy Quad core Zen CPU.

Quad core APU - that is completely different story.

 

[bjt2]

You forgot the maximum of 6 uops/cycle that the Micro-op cache can deliver to the schedulers, which limits the usage of the 10 ports/units.

2% in ONE test?

As I said before, it's only ONE test. And it tested only the IPC: not real-world performances.

Let's wait for more tests, with final frequencies AND Turbo activated.

The 10 uops cycle probabily is avoid bottlenecks and keep scheduler queues almost empty. Even if also INTEL has a 6uop cycle from the cache, the shared uop scheduler can be a bottleneck and so potentially Zen can have greater IPC than intel's in corner cases (and we hope also in more common cases)... Anyway I am still convinced that Zen will be clocked higher than SKL...

Ehm... no. AotS didn't show a similar outcome, and BTW it didn't measured the IPC.

I remember that Zen was compared to similarly clocked INTEL CPUs in Aots and scored similar... If not, my fault...

 

[The Stilt]

Since we haven't had a proper derail for few days now: More information about nVidia GP107 (GTX 1050 / Ti) has been started to pop out, and it appears that the rumors about GP107 being manufactured on Samsung 14nm LPP were accurate. If it turns out to be true, it is pretty much ends the speculation about how the different processes compare between each other. If the GP107 is infact manufactured on 14nm LPP, it pretty much proves (together with the other available evidence) that Samsung process is inferior to TSMC 16nm FF+ in terms of Fmax. GP107 has 519MHz (or > 27%) lower maximum boost frequency than GP106 (GTX 1060), which puts the clocks to similar range as on AMD Polaris ASICs (< 1400MHz).

 

[deasd]

Also a geekbench result. Take with a grain of salt.

http://semiaccurate.com/forums/showpost.php?p=273864&postcount=3883

 

[leoneazzurro]

Or simply Nvidia decided to keep the frequencies low in order to keep a respectable difference with the GTX1050/1060 and/or maximizing the yield, being that a budget solution.

 

[Arachnotronic]

Also a geekbench result. Take with a grain of salt.

http://semiaccurate.com/forums/showpost.php?p=273864&postcount=3883

If this is legit, Zen IPC is slightly worse than Broadwell's.

 

[StrangerGuy]

If this is legit, Zen IPC is slightly worse than Broadwell's.

Last time I checked GB4 doesn't correctly report boost clocks.

 

[Glo.]

Since we haven't had a proper derail for few days now: More information about nVidia GP107 (GTX 1050 / Ti) has been started to pop out, and it appears that the rumors about GP107 being manufactured on Samsung 14nm LPP were accurate. If it turns out to be true, it is pretty much ends the speculation about how the different processes compare between each other. If the GP107 is infact manufactured on 14nm LPP, it pretty much proves (together with the other available evidence) that Samsung process is inferior to TSMC 16nm FF+ in terms of Fmax. GP107 has 519MHz (or > 27%) lower maximum boost frequency than GP106 (GTX 1060), which puts the clocks to similar range as on AMD Polaris ASICs (< 1400MHz).

You forgot that GTX 1050 Ti will be locked by no 6 pin connector. It can be bottlenecked by this factor, not the process.

 

[Arachnotronic]

Last time I checked GB4 doesn't correctly report boost clocks.

So this is apparently a GB3 result and it is not in the database. Probably fake.

 

[mikk]

Yes it's not there, I searched for all Oct 2 results, not available. Probably a fake from a random user.

 

[cytg111]

Or taken down by request?