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

[TheELF]

We can guess looking at the gain that intel had in Fmax with 32nm BULK->22nmFF... Finfet is WAY better than BULK in all respects. Less capacitance, more transcounductance, less leakage. Why on earth the FMax can't go higher?

How much % energy do you save by all this, and whatever other savings zen has, compared to how much % more ALU FPU etc each zen core has? HT in intel also costs more power why shouldn't AMD's SMT ,especially since people think it will be giving better improvements then intel's, it should also draw more power?!
Unless AMD managed to get equal power gains to the amount of "more tech" in each core zen will be running slower(clocks) at the same power draw,max clocks might still be at ~4,8-5ghz but zen might be drawing way too much power at that point.

 

[mikk]

https://mobile.twitter.com/momomo_us/status/805050776087203841

2、关于Zen，据可靠消息，4C8T版本性能跟4790K差不多

Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost


Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

 

[AtenRa]

https://mobile.twitter.com/momomo_us/status/805050776087203841



Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost


Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

What if this performance is at the rumored $150 price point ??

 

[bjt2]

How much % energy do you save by all this, and whatever other savings zen has, compared to how much % more ALU FPU etc each zen core has? HT in intel also costs more power why shouldn't AMD's SMT ,especially since people think it will be giving better improvements then intel's, it should also draw more power?!
Unless AMD managed to get equal power gains to the amount of "more tech" in each core zen will be running slower(clocks) at the same power draw,max clocks might still be at ~4,8-5ghz but zen might be drawing way too much power at that point.

You are forgotting that zen is made on a process 2 node ahead than XV. An XV core has about half alu and half fp (has half of a full FPU), but same AGU and decoders, so it is more than half the resourcses of a Zen core. It is correct that SMT increase consumption, but mainly for more resource usage.
Zen is made on a process with 1/6 of leakage, more drive strength and less parasitic capacitance.

Even if we suppose a zen core with same transistors as an XV module, we have power draw gain projected of up to 70% (and this is almost respected on polaris: +20% clock and almost 50% less power draw).

Let's say it's 50% and we have that a Zen core, at same speed, draw the same as a half XV module, namely an XV "core"... But wait! This is just the AMD statement, that i quoted!

 

[mikk]

What if this performance is at the rumored $150 price point ??

From a technical standpoint the price doesn't matter. Not to mention that your rumored price point is wrong.

 

[Abwx]

https://mobile.twitter.com/momomo_us/status/805050776087203841

Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost

Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

It s unlikely that this Zen is at 4GHz like the 4790K, if it s at say 3.4GHz then it means that it has better throughput clock/clock than HW, so i dont understand how you came to your conclusion,.

I guess that it s of the same accuracy as your personnal translation of Heise.de article when stating that they said that it had slightly lower perf in SPEC than HW while the article state that it keep up with HW..

 

[DrMrLordX]

If AMD can produce a 4c/8t CPU on par with the 4790k, that puts them way ahead of the game. Right now the closest thing they have to that is XV which can't reach the clockspeeds necessary to compete on a thread-by-thread basis AND they can't launch it with more than the capacity for four threads.

And XV only shines when fully-loaded.

There are still instances of where a 4790k can beat 4m/8t Vishera CPUs in raw throughput.

Calling that kind of performance "extremely poor" is absurd.

According to Anandtech Bench:

Cinebench R10 MT (legacy SSE2 FPU throughput benchmark)

4790k stock (4.0 GHz): 33538
FX-9590 stock (4.7 GHz): 26635
Athlon x4 845 stock (3.5 GHz): 14426

Hypothetical 4m/8t XV @ 3.5 GHz: 28852

If 4c/8t Zen is hitting ~33500 in R10 in the 3.5 GHz range then it would be %16 faster than XV in total throughput for this benchmark, which wouldn't be too bad.

 

[mikk]

Calling that kind of performance "extremely poor" is absurd.

It is extremely poor if this is based on a Multithread test versus a 4 core CPU if you like it or not. An easy MT advantage over any 4C i7 is something we should expect.

 

[Dresdenboy]

https://mobile.twitter.com/momomo_us/status/805050776087203841



Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost


Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

Nice! But since when is the 4790K an 8C chip?

Edit:
Full translation:

1, on the RX490, perfect domestic media YY, mainly because of RTG now has a technology summit
2, on the Zen, according to reliable sources, 4C8T version performance with 4790K almost
3, Vega determine next year Q1 officially released (I mean the market, the paper surface release to discipline), performance is not a surprise, more power is improved
4, RX470 / 480 price collapse is not good for AMD
5, able to do so within USD10 buy AMD stock it, or buy it than buy U card also the actual point

http://tieba.baidu.com/p/4873908142

 

[Abwx]

Cinebench R10 MT (legacy SSE2 FPU throughput benchmark)

4790k stock (4.0 GHz): 33538
FX-9590 stock (4.7 GHz): 26635
Athlon x4 845 stock (3.5 GHz): 14426

.

Dunno if CB R10 is SSE2, isnt it rather X87 based..?..

Because in CB 11.5, wich is SSE2 for sure, the 4790K has barely 2.5% advantage over the 9590, and the X4 845 is 21% faster clock/clock than a 4C Vishera, this to point that CB R10 is completely obsolete..

 

[deasd]

https://mobile.twitter.com/momomo_us/status/805050776087203841



Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost


Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

LOL, translating disaster. Or perhaps you misinterpreted this, let me translate:

4C8T version of Zen has similar performance to 4790k.

DISCLAIMER: I don't pretty trust those chinese leakers due to their bad track record, I just doing translating work and have no responsibility about this info.

Dunno if CB R10 is SSE2, isnt it rather X87 based..?..

Because in CB 11.5, wich is SSE2 for sure, the 4790K has barely 2.5% advantage over the 9590, and the X4 845 is 21% faster clock/clock than a 4C Vishera, this to point that CB R10 is completely obsolete..

R10 might still use SSE but I suspect it use pretty simple datapath(<128bit) that is rarely used by majority of apps right now, it is now look much like superpi.......

 

[cytg111]

This is the graph of the famous ARM NEON test chip. The X axis are the frequency in GHz and the Y axis is the power x100mW.

Oki, interesting. What libraries is Zen built with? High dense? - That would be the limited graphs right?

 

[bjt2]

Oki, interesting. What libraries is Zen built with? High dense? - That would be the limited graphs right?

This chart was of 1-2 years ago and lacks also sLVT (that should be even better). Besides this, time pass and probabily the performance is improved...

9T is HDL and 12T is HPL. Now we know that the HPL library would never be used. A 10.5T intermediate is the maximum that will be allowed. Since even the HDL (9T) library performance is interesting, event if Zen will use LVT 9T transistors, we can expect quite interesting clocks... The consumption is interestingly linear up to 4GHz and over... If they use 10.5T and sLVTs in critical paths, we can expect fmax over 4GHz, thus in the same ballpark or slightly better than actual XV chips, that have Fmax of 4.3GHz. I don't se why not...

 

[AtenRa]

4C8T version of Zen has similar performance to 4790k.

From a technical standpoint the price doesn't matter. Not to mention that your rumored price point is wrong.

Nice! But since when is the 4790K an 8C chip?

Yea really, if this 4C8T ZEN can much Haswell 4C8T Core i7 4970K (4GHz) its a huge step. And with the ability to upgrade later to a 8C16T on the same platform its getting really nice. Lets wait and see the performance/price and then we can talk more things.

I have a feeling this thing is aiming the Core i5s KabyLake

 

[inf64]

Even without deasd's translation work I managed to understand that the Chinese leaker likely meant that Zen 4C/8T version roughly matches 4790K.

If this is true then it is huge achievement as it means several things:
- most likely scenario : 4C/8T Zen has similar IPC to Haswell and clocks to similar frequency as 4790K which in turn means there is no clocking problem with this generation of Zen cores, only barrier is TDP.
- less likely scenario : if 4C/8T Zen has similar performance to 4790K without clocking to those levels than IPC is >Haswell which is also very good prospect for AMD
- least likely scenario : if 4C/8T Zen can actually outclock 4790K (by more than 5-10%) this would imply that IPC is @ IvyBridge levels and clocks are at 4.4+Ghz territory. This wouldn't be bad per se but Zen clocking this high is the most improbable case out of the 3.

 

[jpiniero]

Not buying it, but I suppose the other option is that the SMT boost is closer to what AMD got with CMT.

 

[RussianSensation]

https://mobile.twitter.com/momomo_us/status/805050776087203841

Google translated: on Zen, according to reliable sources, 4C8T version of the performance with the 4790K almost

Not sure if this is a good result if true, given that it is 4C vs 8C. Depends on the test, Multithread performance on par with 4790k would be extremely poor for example.

Huh? It says 4C/8T Zen ~ 4C/8T 4790K.

Sounds almost too good to be true because it would mean Zen would come within 10-15% of Skylake IPC and it would also mean 8C/16T Zen would ~ 5960X. If this were to happen, AMD would sell out every single 4C/6C/8C unit they make if they price these between $150-500.

No matter how much $ NV has, they cannot manufacture a CPU that even comes close to 30% as fast as Intel's best CPU. No matter how much $ Intel has, they cannot even manufacture a GPU that comes close to 20% as fast as AMD's fastest GPU. Given how AMD is such a tiny firm in comparison and is fighting both NV and Intel on 2 fronts, while having a fraction of the resources of Intel OR NV, if it manages to release an 8C/16T CPU that comes close to the 5960X and can also then sell it for $499, that would be an unbelievable achievement!

Intel's cheapest 8C/16T CPU, the i7-6900K, costs $1089 USD on Newegg.

Even if Zen's 8C/16T CPU delivered just 75% of the total performance of the $1089 6900K for $499, it would already be a great value. If Zen comes in with 85-90% of the performance of the 6900K at $499, it would be an absolutely insane amount of value.

If 4C/8T Zen costs $150 and has Haswell IPC and can overclock to 4.4-4.6Ghz, it would make all i3 and i5s Intel makes instantly irrelevant. The era of i3s and i5s is coming to an end quickly.

i7 2600K is beating i5-6600K is an eye-opener of what's to come. PS5/XB2 aren't far away either and we should expect these consoles in 2019-2020. For anyone buying a 2017 CPU, a 'budget' 4C/8T Zen should be a great value for games.

i7 6700 is 41% faster than i5 6600 in Watch Dogs 2

i7 HT = fully utilized

For 4K gaming, the load shifts almost entirely to the GPU. At which point Zen's extra cores will be a nice free bonus for productivity/future-proofing over Intel's i3s and i5s and 4C/8T i7s.

From a technical standpoint the price doesn't matter. Not to mention that your rumored price point is wrong.

Ok, and when the consumer goes out to buy the product and they see 4C/8T, 6C/12T, 8C/16T Zen CPUs (real cores, with real Hyper-Threading like tech) selling $200-500 compared to Intel's anemic i3s and i5s and 4C/8T i7s, Zen will find a huge customer base.

To this day, there are tens of millions of PC gamers using 1st gen Nehalem/Lynnfield, 2nd and 3rd gen Sandy and Ivy i3/i5/i7s and they have not upgraded to Skylake despite Skylake's IPC. A lot of these users are simply fed up paying $350 for 4 core HT i7s and they will easily take a 15-20% lower IPC in exchange for 2X the cores.

The issue with Bulldozer/Vishera was that it never had 8 physical cores with 16 threads and it had IPC below that of Nehalem. In 1 shot AMD is going to give these users both an increase in IPC they wanted all this time and double the cores for each segment they own (i3 vs. 4C/8T Zen, i5 vs. 6C/12T Zen, i7 4C/8T vs. 8C/16T Zen). Most games now are CPU thread and GPU limited. The IPC advantage of i3s and i5s no longer helps them in games overcome the lack of cores/threads. We can see it already with i7 2600K / 3770K / 4770K vs. i3 6100/i5 6600K.

Spoiler: i5-6600K loses in almost every single benchmark at Sweclockers to the Ivy Bridge IPC[B] i7 3770K.[/B]

Productivity wise, nothing much needs to be said about Intel's i3 and i5 CPUs. They are absolute dogs, slower than a 6 year old (!) i7 2600K.

Spoiler: Skylake's IPC advantage over Sandy/Ivy i7 2600K/3770K does NOT help in multi-threaded and/or many productivity apps

4C/8T Zen just has to be as fast as an i7 3770K, use < 100W, and cost $150-230 and have the ability to overclock to 4.5-4.6Ghz and every modern desktop Skylake i5-6400-6600K or 7600K CPU is going to be extremely hard if not impossible to recommend for 2017 builds.

 

[itsmydamnation]

Not buying it, but I suppose the other option is that the SMT boost is closer to what AMD got with CMT.

The problem with this thread is so many people (particularly on the negative) dont even look at the mirco architecture and think before going off to war. AMD SMT isn't going give CMT style 2nd thread "performance" unless your workload gets bad cache hit rates and thats the same for intel SMT, the reason is simple. CON core has 2/1 load/store a cycle per core and so does Zen and intel, assuming the stack engine does its own simple address gen L/S is the obvious bottle neck given an aprox avg of 50% of x86 ops have a memory component.

To me it's always been the same people pushing negative agenda without being able to back anything up. look at all my posts none of them have any of the marking numbers in them, they are far to simplistic and people pick and choose how to interpret them, its pointless.

look at the architecture.
Everything in terms of the presented architecture puts Zen at a broadwell/skylake level.
much improved L1D (write-back) increased associativity
not detailed by AMD but will be a much lower L2 latency
L3 to have 5 times the bandwidth (we will have to wait and see about latency ) but hat put L3 performance in the same ball park (http://i.imgur.com/IAkGdxV.png, https://www.aida64.com/sites/default/files/shot3_cachemem_skylake.png)
Fetch looks pretty simlar, decoubled prediction larger TLB's then BD comparable but different to intel
Decode looks to be the same but now with additional uop cache so again around skylake/broadwell
issue/dispatch is bigger comparable to skylake
PRF files are comparable to broadwell (slightly behind skylake)
I would call Execution the same as >haswell (broadwell, skylake they are all the same) all the unit counting is silly potentially AMD can get one more FP ADD/MUL a cycle
load store system is hardest to gauge as no one especially intel talks about it. no idea about improvements to memory disambiguation*
finally SMT, almost every structure is shared, according to Mr kanter retire on skylake isn't shared and it isn't on Zen either so it looks very comparable.

one big unknown is memory controller performance, given that they are licensing rumbus controller hopefully its not terrible .....lol

Now you can take the position that Intel will squeeze out a little bit more performance from each unit but that still not going to radically change things. So again it come back to why can't Zen perform per clock at a broadwell/skylake level (outside 256bit ops).

*i went looking to see if amd had said anything about memory disambiguation for Zen, all i found is my own post from march saying the exact same thing as i am now.....lol
https://forums.anandtech.com/threads/new-zen-microarchitecture-details.2465645/page-12#post-38097201 good to see the discussion has moved on......

 

[cytg111]

Sounds almost too good to be true because it would mean Zen would come within 10-15% of Skylake IPC and it would also mean 8C/16T Zen would ~ 5960X. If this were to happen, AMD would sell out every single 4C/6C/8C unit they make if they price these between $150-500.

It doesnt smell right, sounds like an all out price war with Intel.. Intel with the deep pockets and AMD not so much? Even if their production costs were low enough to sustain that price I fully expect them to be in the ballpark of equal Intel offerings, siphoning as much $$ out of the deal as possible (or maybe with the WSA their hands is tied and they have no other play than to flood the market with somewhat cheap 8 cores ..).

 

[sirmo]

No matter how much $ NV has, they cannot manufacture a CPU that even comes close to 30% as fast as Intel's best CPU. No matter how much $ Intel has, they cannot even manufacture a GPU that comes close to 20% as fast as AMD's fastest GPU. Given how AMD is such a tiny firm in comparison and is fighting both NV and Intel on 2 fronts, while having a fraction of the resources of Intel OR NV, if it manages to release an 8C/16T CPU that comes close to the 5960X and can also then sell it for $499, that would be an unbelievable achievement!

+100

I'm hyped for the eternal underdog.

 

[bjt2]

The problem with this thread is so many people (particularly on the negative) dont even look at the mirco architecture and think before going off to war. AMD SMT isn't going give CMT style 2nd thread "performance" unless your workload gets bad cache hit rates and thats the same for intel SMT, the reason is simple. CON core has 2/1 load/store a cycle per core and so does Zen and intel, assuming the stack engine does its own simple address gen L/S is the obvious bottle neck given an aprox avg of 50% of x86 ops have a memory component.

To me it's always been the same people pushing negative agenda without being able to back anything up. look at all my posts none of them have any of the marking numbers in them, they are far to simplistic and people pick and choose how to interpret them, its pointless.

look at the architecture.
Everything in terms of the presented architecture puts Zen at a broadwell/skylake level.
much improved L1D (write-back) increased associativity
not detailed by AMD but will be a much lower L2 latency
L3 to have 5 times the bandwidth (we will have to wait and see about latency ) but hat put L3 performance in the same ball park (http://i.imgur.com/IAkGdxV.png, https://www.aida64.com/sites/default/files/shot3_cachemem_skylake.png)
Fetch looks pretty simlar, decoubled prediction larger TLB's then BD comparable but different to intel
Decode looks to be the same but now with additional uop cache so again around skylake/broadwell
issue/dispatch is bigger comparable to skylake
PRF files are comparable to broadwell (slightly behind skylake)
I would call Execution the same as >haswell (broadwell, skylake they are all the same) all the unit counting is silly potentially AMD can get one more FP ADD/MUL a cycle
load store system is hardest to gauge as no one especially intel talks about it. no idea about improvements to memory disambiguation*
finally SMT, almost every structure is shared, according to Mr kanter retire on skylake isn't shared and it isn't on Zen either so it looks very comparable.

one big unknown is memory controller performance, given that they are licensing rumbus controller hopefully its not terrible .....lol

Now you can take the position that Intel will squeeze out a little bit more performance from each unit but that still not going to radically change things. So again it come back to why can't Zen perform per clock at a broadwell/skylake level (outside 256bit ops).

*i went looking to see if amd had said anything about memory disambiguation for Zen, all i found is my own post from march saying the exact same thing as i am now.....lol
https://forums.anandtech.com/threads/new-zen-microarchitecture-details.2465645/page-12#post-38097201 good to see the discussion has moved on......

Just one point. INTEL archs have 4 shared int/fp ports. Zen has 4 int + 4 fp ports. This can be an advantage in smt code, but also in some complicated FP workload, because all branch, memory, compare and ancillary instructions are executed on the int ports: on Zen they don't impair FP performance, on INTEL they fight with FP instructions for the ports... And this is worse in MT tasks, expecially if they are FP. I think that this is the reason of slightly better performance in blender: it's an heavy FP task, but it has also int instructions for control flow and other tasks. On INTEL these 2 threads can execute max 4 uop/cycle of mixed int and FP. On Zen they can execute max 8 uop/cycle. I said in the past that most workloads have IPC around 1. Only heavy FP, like spec FP have higher IPC. SpecFP have around 2.5... So 2 threads require 5 instructions/cycle. Even if all instructions are of one uop, we are talking of 5uop/cycle. Probabily most of memory instructions are load/op or rmw ops, so we have probabily 5uop/cycle plus memory instructions. On INTEL 4 shared ports are not enough. On Zen we have 8. On such a workload (MT, heavy FP, low/medium memory load), Zen could perform better than INTEL...

We can forget CMT performance though, because BD for 2 threads has 2 L1 cache with 2L/1S ports, while zen has 1 such cache (albeit bigger). Since the L2 and L3 bandwidth is bigger, Zen could have CMT performance only on memory hungry tasks (streaming), where the work load does not fit into the Bulldozer L1 caches. Moreover L1D is double than INTEL's and also L2, so in such workloads it can even do better than INTEL's. I am talking of workloads with memory footprint of about 512KB, just around Zen L2 cache and above INTEL's L2 cache...

 

[bjt2]

+100

I'm hyped for the eternal underdog.

I saw liquid cooling and nitrogen... Zen extreme OC on stage?

 

[DrMrLordX]

It is extremely poor if this is based on a Multithread test versus a 4 core CPU if you like it or not. An easy MT advantage over any 4C i7 is something we should expect.

Are you serious? We're talking 4c/8t Zen vs 4c/8t Haswell here. Not 8c/16t Zen!

Yea really, if this 4C8T ZEN can much Haswell 4C8T Core i7 4970K (4GHz) its a huge step.

Thank you. And yes it is an "if" right now . . . we don't know the exact performance of Summit Ridge at this time.

R10 is, in my opinion, still very interesting as a benchmark. Look at the performance differences between Vishera, Excavator, and Haswell . . . Intel has an enormous advantage per thread vs. Vishera, but not so much of one compared to Excavator. Overall Con cores do better in R10 relative to the Intel competition than any other version of Cinebench.

If you prefer something newer, here's R15 MT (wish bench had 11.5 numbers too):

FX-9590: 728
i7-4790k: 948
Athlon x4 845: 313.96

Hypothetical XV 4m/8t chip @ 4.0 GHz: 717.6

If 4c/8t Zen manages a score of 948 in R15 at a clockspeed of 3.5 GHz, then that gives Zen a 32% throughput lead over XV in this benchmark. I don't know if I want to be that optimistic, but if that happened . . . yowza.

 

[KTE]

We can guess looking at the gain that intel had in Fmax with 32nm BULK->22nmFF...

Intel's process is typically FAR better than whatever AMD uses for the respective pMOS/nMOS values...

Did Intel just drastically change their architecture, hint much wider and more powerful?

Do you think IPC increases + SMT are free?

That happened from Core to Nehalem. Remember that?

Do you realize same chip, one running default and one tuned to extract higher IPC on a specific code = latter burns more power?

Have you ever designed a logic circuit at silicon level (W and L of transistors channel etc)? I did. At college (i am Italian, 39 years, so before the current reform that cutted a lot in course programs), electronic II course. It's all matter of transconductance, leakage and parasitic capacitance. If all is better, why on earth can't be clocked higher?

So you've designed nanotech ICs and you say there is no scientific chance possible that can make Zen clock lower or equal to Excavator at 95W.

Strong statements.

So if Zen turns out to clock lower than Exc at 95W, you seriously have no clue about processor design or nanotechnology.

I've already explained to you your compares and deductions are logically invalid but you just keep repeating the same words. You are missing most of the crucial factors to create crazy hype around Zen.

Your 4.0G/4.3Hz 8/16 belief is... we'll see by how much you're off

Sent from HTC 10

 

[itsmydamnation]

i wouldn't really use excavator as a guide it only had intentions of notebook/ small form factor OEM boxes and also probably developed with the least expense possible.

What is interesting is to look at 4 module bulldozer and pilediver first release 95watt TDP parts.

8100 base 2.8ghz all core turbo 3.1ghz max turbo 3.7ghz
8300 base 3.3 turbo 1 3.6 max turbo 4.2ghz

so the question is what are the Zen OPN numbers base and all core turbo? base and max turbo? all core turbo and max turbo?

if its the first then 95watt Zen has aprox the same clocks as piledriver 95watt parts. Would 125watt reach 8350 clocks?