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

[OrangeKhrush]

AMD's minimum accpted board for a 1700X is a B350, the test was run on a A320 which is why Turbo doesn't work, it is 100% certain that that score is at 3.4Ghz, that is very impressive.

 

[The Stilt]

AMD's minimum accpted board for a 1700X is a B350, the test was run on a A320 which is why Turbo doesn't work, it is 100% certain that that score is at 3.4Ghz, that is very impressive.

Thats not true.

 

[OrangeKhrush]

The real question is if the processor is really running at 3.4 GHz or if it's running at a boost clock that the benchmarking program can't properly defect. For all we know, it might have some exotic liquid cooling system that lets it hit 5 GHz.

I'd still wait for official benchmarks before placing a pre-order.

As above, the R7's are not designed for A320 which is why XFR and turbo is off.

 

[OrangeKhrush]

Thats not true.

The source writer even stated that the board could not detect turbo. The slides showed SR7 suited only to X370 and B350 boards as 320 chipset doens't support clocking. Either way turbo was not detected in the benches.

 

[The Stilt]

The source writer even stated that the board could not detect turbo. The slides showed SR7 suited only to X370 and B350 boards as 320 chipset doens't support clocking. Either way turbo was not detected in the benches.

Boost is completely invisible on Ryzen. The only way to detect it is to continuously monitor the clocks (e.g HPET, TSC).
Obviously the board doesn't need to support overclocking to run the chip at stock. Same rules apply as before: The chip is allowed to run at it's rated speeds, unless the some protection kicks in (i.e VRM OTP, CPU temperature etc). In that regard Ryzen is no different to previous generations.

 

[OrangeKhrush]

Boost is completely invisible on Ryzen. The only way to detect it is to continuously monitor the clocks (e.g HPET, TSC).
Obviously the board doesn't need to support overclocking to run the chip at stock. Same rules apply as before: The chip is allowed to run at it's rated speeds, unless the some protection kicks in (i.e VRM OTP, CPU temperature etc). In that regard Ryzen is no different to previous generations.

Not to sure I buy into invisable turbo when CPC had reflected turbos

 

[The Stilt]

Not to sure I buy into invisable turbo when CPC had reflected turbos

Care to elaborate?

 

[lolfail9001]

Care to elaborate?

CPC actually mentioned that their sample had 3.15 base clock, 3.3 all core boost and 3.4 single core, i believe. Though the first and last can be deduced from ES name and the second can be just tracked manually.

 

[The Stilt]

CPC actually mentioned that their sample had 3.15 base clock, 3.3 all core boost and 3.4 single core, i believe. Though the first and last can be deduced from ES name and the second can be just tracked manually.

Or, they alternatively could have used software like CPU-Z to see the clocks the CPU was running at.

 

[bjt2]

It's possible, yes. IMO.

But Vt 'binning' is never done. Those chips consumers see as 'high leakage', they are way too hot on current to validate on most MBs.

If you mean intentional Vt variation among transistors, that's a different topic altogether and not what is referred to as 'high/low leakage chip'.

I am non native english, so maybe i didn't explain well what i meant.

When you perform binning, you usually choose the fast bin in which the CPU can run under the rated TDP of that bin. If you pick chips that go into a 65W bin, they will have mean leakage lower that chips that go into 95W bin simply because high leakage chips have much lower probability of drawing lower power.
I was not suggesting of making binning on leakage.

This sometimes is done by users on the GPU, after reading the ASIC quality (or the name that it has) into the GPU ROM, but usually after buying more than one sample and reselling the others...

I don't mock you. I mocked your repetitive assertions because they were adrift of reality.

You were claiming 4-4.5GHz for launch under 95W, and theorizing more is possible due to the Neon FPU on 14nm test vehicle. Then you extrapolated that to EXC max clock and added the gains on top. Plus the >BDe level IPC and better than BDe power.

I have those posts saved

For the clocks, i stated "up to" and added that maybe not in the first batches. For the IPC, the resources are plenty and from haswell to skylake there are 8-10% of difference or less. So it's easy to be right or wrong, depending on the actual IPC value. If i am wrong, probabily i will not be off by more than 10%...

Clocks after some tuning is when I said that it is surely possible but certainly not for launch under 95W. You always hit a knee-jerk reaction on this, until now.

Because all my calculations tends to confirm clocks around 4GHz (or just below) and i keep reading of unsupported claims of 2.8-3.2GHz, just because INTEL 8c16t is 3.2GHz base... And I am nearer to the truth that those unsupported claims...

It's been shown that Samsung's process is inferior on every metric to Intels original nontuned 14nm in 2014. Just imagine Intels process right now in 2017, but you still kept promoting without data how AMD will easily beat those IPC x Clocks@Power. I kept trying to explain to you it's a seesaw, but then gave up after you would ignore answering pertinent questions, providing data or discussing actual historical trends and data.

I am making calculations on differences between 28nm bulk and 14nm FF. I don't care of intel, differently of others that base their prediction to te mantra "AMD can't beat intel".

Moreover there is someone on this forum that said that GF has 50% more conductance. I don't know if it's true, but I know that INTEL's is denser. But this is not an advantage if the power needed is superior, because of power density: a small chip can't dissipate much power...

That linked is about transistor level not chip level.

These trannys are used in different areas depending on if it is in the critical speed path or power critical, they pose different benefits. HVT allows lower power but slower delay. LVT allows the converse. I don't disagree with any of that. If you search back I explained these same points months ago on here.

I couldn't find the graph that i once saw (some day ago) with a true and clear correlation between Vt and leakage. This was the best i could find. I know that i'ts not the same. The different Vt caused by process variability is reflected in the leakage...

But I will say, I dislike anyone trying to save face backtracking\euphemisming on lengthy claims after they realize how incorrectly exaggerated they were for months, now that reality is hitting home.

It's intellectual dishonesty and I expect you to not take this route.

It's better to live and learn, but take a measured approach for the future processor releases.

Sent from HTC 10
(Opinions are own)

I will not negate my statements. I was stating 4GHz base at 95W since the start. 2 years ago the only condition was IF zen had the same FO4 than XV. Since this seems confirmed, also by Keller, i am fully convinced of my predictions...

 

[OrangeKhrush]

Or, they alternatively could have used software like CPU-Z to see the clocks the CPU was running at.

This is my CPU with invisible turbo, you enable it by turning turbo off in BIOS, no application can detect it.

http://imgur.com/0FJ6Q20

(cannot upload) - this is my i5 4460 with a very visible turbo now.

I am just going to call a spade a spade and say that turbo was legitimately disabled, due to instability with firmware. Passmark has good enough intrusion software to read turbo states. There is no such thing as undetectable turbo frequency.

 

[itsmydamnation]

Look at the physics score. That is what is going to condemn it to gaming mediocrity. The memory score is also a grave concern, if it is cache bandwidth that is pulling it down.

no its not, systematically speaking from a diabolical point of view your fundamental facilities are not sufficiently sophisticated to bombast your philosophies....

first point, my ivb 3770@4.3
CPU Mark 10673
Integer Math 19504
Floating Point Math 8336
Prime Numbers 31.5
Extended Instructions (SSE) 219.1
Compression 14806
Encryption 1944
Physics 597
Sorting 8577
CPU Single Threaded 2358

So physic's is a multi-threaded task, so my ivb that has 256bit AVX units but no AVX2 or FMA scores 597 in physics which is in the 87% percentile of this measured benchmark. So 87% of computers are gaming mediocrity are they? Sounds like lazy devs to me /s. Normalizing that to 3.8ghz is 527. Now my memory score is way better then that Zen sample(2762 95th percentile) which if this is a real world representative benchmark is very important to this kink of benchmark.

Now on to point number two, Memory and your complete inability to actual read number and then the second and most important point, COMPREHEND NUMBERS. So we have three types of memory numbers , latency , cache throughput and memory throughput. lets look at memory throughput first, compare to my DDR3 1866 (passmark isn't getting the clock or the timings correct CBF looking at that they really are im around 2000mhz DDR).

memory read uncached 14915 (zen) vs 15799 (my ivb)
memory write 7917 (zen) vs 10959 (my ivb)

So it looks like Zen is a little worse here on reads but it is using crap memory, writing are quite a lot lower so that looks like a problem but it probably isn't. the reason is simple, A+B=A|B, this is what x86/x64 operations do, two reads, one write, write bandwdith is >50% of read so all is good. Now on to cache throughput

memory cache read 28006 (zen) vs 28313 (my ivb)
memory threaded 34001 (zen) vs 25498 (my ivb)

So cached read (its a single core test) is actually very very good for Zen, my ivb runs L1,L2,L3 all at 4.3ghz, Zen is L1,L2 @ 3.4/3.8 and L3 is unknown. Memory threaded is a read test with a threaded launched per hardware thread, its ability to saturate DDR4 2400 is obviously not an issue. Now onto latency,

Database operations 78 Zen, 106 my ivb
memory latency 76ns vs 23ns

So first off my memory does not have an access time of 23ns, something is broken in this benchmark, ADIA64 is around 50ns which is far more realistic, if i run the passmark advanced memory latency test i get, Random Range 23ns ( cant be that random, i guess the prefetcher is guessing it) and single random as 72ns.

Now the DB test is heavy compute and memory latency, we know from the compute results compute isn't the problem. So lets look at latency numbers:
my DDR3 has CAS of ~10 and clock of ~ 1000mhz , giving access latency of around 9.5ns
Zen DDR4-2400 has CAS of 17 and clock of 1200mhz, giving access latency of around 14ns.

So my memory latency is 47% better, guess what 78 * 1.47 = 114

CRYSIS AVERTED!@!#!@#!@#342434

Now onto the third point, AVX , AVX 2 and FMA usage in games.

Thax to intel, its not going to be prevalent any time soon because there are to many chips that have AVX and AVX2 fused off. Then add the number of people still on SB and IVB that dont have AVX2 or FMA . Then add that the consoles only have 2x128b AVX units a core and we can see Dev's are going to target data structures that align to 128bit operations for a good number of years yet. My bet is ~5 years before they (256bit ops) are the common target, thax Intel


So Now that i have explained your own numbers to you, What was your point again?

I am now going to go cut my memory speed in 1/2 and rerun the test to check the impact both the physics test and memory tests, Aren't you going to look very silly if the physics test is memory latency bound

 

[OrangeKhrush]

Every Ryzen benchmark is being gimped in some way, the AOTS running far more samples and half terrain off yet came withing 7% of a i7 6950X's score now this benchmark where Turbo was definitely disabled on a meh level motherboard with the loosest timings possible.

 

[Glo.]

Every Ryzen benchmark is being gimped in some way, the AOTS running far more samples and half terrain off yet came withing 7% of a i7 6950X's score now this benchmark where Turbo was definitely disabled on a meh level motherboard with the loosest timings possible.

Lets not spread FUD, and lets wait for reviews, shall we?

Downplaying numbers is not fine. Hyping the CPU before its released is also not fine. We see, what we see. End of it.

 

[itsmydamnation]

no its not, systematically speaking from a diabolical point of view your fundamental facilities are not sufficiently sophisticated to bombast your philosophies....

first point, my ivb 3770@4.3
CPU Mark 10673
Integer Math 19504
Floating Point Math 8336
Prime Numbers 31.5
Extended Instructions (SSE) 219.1
Compression 14806
Encryption 1944
Physics 597
Sorting 8577
CPU Single Threaded 2358

So physic's is a multi-threaded task, so my ivb that has 256bit AVX units but no AVX2 or FMA scores 597 in physics which is in the 87% percentile of this measured benchmark. So 87% of computers are gaming mediocrity are they? Sounds like lazy devs to me /s. Normalizing that to 3.8ghz is 527. Now my memory score is way better then that Zen sample(2762 95th percentile) which if this is a real world representative benchmark is very important to this kink of benchmark.

Now on to point number two, Memory and your complete inability to actual read number and then the second and most important point, COMPREHEND NUMBERS. So we have three types of memory numbers , latency , cache throughput and memory throughput. lets look at memory throughput first, compare to my DDR3 1866 (passmark isn't getting the clock or the timings correct CBF looking at that they really are im around 2000mhz DDR).

memory read uncached 14915 (zen) vs 15799 (my ivb)
memory write 7917 (zen) vs 10959 (my ivb)

So it looks like Zen is a little worse here on reads but it is using crap memory, writing are quite a lot lower so that looks like a problem but it probably isn't. the reason is simple, A+B=A|B, this is what x86/x64 operations do, two reads, one write, write bandwdith is >50% of read so all is good. Now on to cache throughput

memory cache read 28006 (zen) vs 28313 (my ivb)
memory threaded 34001 (zen) vs 25498 (my ivb)

So cached read (its a single core test) is actually very very good for Zen, my ivb runs L1,L2,L3 all at 4.3ghz, Zen is L1,L2 @ 3.4/3.8 and L3 is unknown. Memory threaded is a read test with a threaded launched per hardware thread, its ability to saturate DDR4 2400 is obviously not an issue. Now onto latency,

Database operations 78 Zen, 106 my ivb
memory latency 76ns vs 23ns

So first off my memory does not have an access time of 23ns, something is broken in this benchmark, ADIA64 is around 50ns which is far more realistic, if i run the passmark advanced memory latency test i get, Random Range 23ns ( cant be that random, i guess the prefetcher is guessing it) and single random as 72ns.

Now the DB test is heavy compute and memory latency, we know from the compute results compute isn't the problem. So lets look at latency numbers:
my DDR3 has CAS of ~10 and clock of ~ 1000mhz , giving access latency of around 9.5ns
Zen DDR4-2400 has CAS of 17 and clock of 1200mhz, giving access latency of around 14ns.

So my memory latency is 47% better, guess what 78 * 1.47 = 114

CRYSIS AVERTED!@!#!@#!@#342434

Now onto the third point, AVX , AVX 2 and FMA usage in games.

Thax to intel, its not going to be prevalent any time soon because there are to many chips that have AVX and AVX fused off. The add the number of people still on SB and IVB that dont have AVX2 or FMA . Then add that the consoles only have 2x128b AVX units a core and we can see Dev's are going to target data structures that align to 128bit operations for a good number of years yet. My bet is ~5 years before they are the common target, thax Intel


So Now that i have explained your own numbers to you, What was your point again?

I am now going to go cut my memory speed in 1/2 and rerun the test to check the impact both the physics test and memory tests, Aren't you going to look very silly if the physics test is memory latency bound

Man i am so right it just hurts

my IVB @4.3 DDR3 1066 10-11-10-30 T2

CPU Mark This Computer 9230
Integer Math This Computer 19408
Floating Point Math This Computer 8121
Prime Numbers This Computer 19.8
Extended Instructions (SSE) This Computer 225.8
Compression This Computer 14193
Encryption This Computer 2024
Physics This Computer 359.7
Sorting This Computer 8723
CPU Single Threaded This Computer 2370

Memory Mark This Computer 2059
Database Operations This Computer 82.6
Memory Read Cached This Computer 27786
Memory Read Uncached This Computer 11720
Memory Write This Computer 6455
Available RAM This Computer 13141
Memory Latency This Computer 33.6
Memory Threaded This Computer 14229


Be gone you trolling, shill!

 

[OrangeKhrush]

Lets not spread FUD, and lets wait for reviews, shall we?

Downplaying numbers is not fine. Hyping the CPU before its released is also not fine. We see, what we see. End of it.

It is not hyping, it happened, benches of a stock CPU released against overclocked an highly clocked CPU's to paint it in bad light, the information was not conveyed meaning it was made in intention to sabotage. I found a bunch of Benches on what I believed was stock clocked parts and the 6950X scored about 77.7 to Ryzen's 70.5, the 6800 was about 67.7 and 5930K around 63 all those CPU's were benched with 12 million samples instead of 16 and half terrain detail on while Ryzen was off. I believe in being fair.

Passmark is not the best benchmark but is shows Ryzen's IPC levels somewhere between Haswell and Broadwell, while dipping to IVY levels in AVX/FMA3 instructions, since the latter is irrelevent for high performance gaming PC's and power users the result shows that Ryzen's baseclock IPC is as strong as Haswell, that was never an overhype by any stretch of the imagination.

 

[Glo.]

Passmark is not the best benchmark but is shows Ryzen's IPC levels somewhere between Haswell and Broadwell, while dipping to IVY levels in AVX/FMA3 instructions, since the latter is irrelevent for high performance gaming PC's and power users the result shows that Ryzen's baseclock IPC is as strong as Haswell, that was never an overhype by any stretch of the imagination.

That is correct conclusion, based on what we have seen so far. For the rest of it - lets wait for reviews. It will not be too long from now.

 

[.vodka]

Man i am so right it just hurts

my IVB @4.3 DDR3 1066 10-11-10-30 T2

CPU Mark This Computer 9230
Integer Math This Computer 19408
Floating Point Math This Computer 8121
Prime Numbers This Computer 19.8
Extended Instructions (SSE) This Computer 225.8
Compression This Computer 14193
Encryption This Computer 2024
Physics This Computer 359.7
Sorting This Computer 8723
CPU Single Threaded This Computer 2370

Memory Mark This Computer 2059
Database Operations This Computer 82.6
Memory Read Cached This Computer 27786
Memory Read Uncached This Computer 11720
Memory Write This Computer 6455
Available RAM This Computer 13141
Memory Latency This Computer 33.6
Memory Threaded This Computer 14229


Be gone you trolling, shill!

Man those tests did take quite the dive with the slower memory!

Now I wonder what could this Ryzen sample do with faster memory up to whatever it supports with decent timings.

 

[OrangeKhrush]

Lol got banned from Toms for posting the turbo off/on state and how it appears on Passmark as N/A or only shows base clocks. Aparently if you use the "you" term you are calling out. I guess they rather promote wrong information or assumptions over someone that actually wasted 10 minutes running the test. Good riddence.

 

[bjt2]

Boost is completely invisible on Ryzen. The only way to detect it is to continuously monitor the clocks (e.g HPET, TSC).
Obviously the board doesn't need to support overclocking to run the chip at stock. Same rules apply as before: The chip is allowed to run at it's rated speeds, unless the some protection kicks in (i.e VRM OTP, CPU temperature etc). In that regard Ryzen is no different to previous generations.

I don't know for Ryzen, but last AMD BIOS and kernel developers guide (yes: i have fun reading this stuff even if i am not a bios or kernel developer) AFAIK, contained instructions on how to program the various P-states, including turbo P-states, that can be modified, at least in linux, even at run time. If you don't program correctly the P-states, it won't work. Probabily also the energy saving features will be broken.
If that AMD document specify how to program these p-states, i suppose that there are not default valuses. And this is understandable: if there is a very low end motherboard with weak VRMs, the bios is responsible of programming not too high p-states...

 

[OrangeKhrush]

I don't know for Ryzen, but last AMD BIOS and kernel developers guide (yes: i have fun reading this stuff even if i am not a bios or kernel developer) AFAIK, containde instructions on how to program the various P-states, including turbo P-states, that can be modified, at least in linux, even at run time. If you don't program correctly the P-states, it won't work. Probabily also the energy saving features will be broken.
If that AMD document specify how to program these p-states, i suppose that there are not default valuses. And this is understandable: if there is a very low end motherboard with weak VRMs, the bios is responsible of programming not too high p-states...

Or turbo was turned off, which is likely.

 

[bjt2]

Or turbo was turned off, which is likely.

Yes. If i remember well, the boost p-states must be enabled explictly by the BIOS. This is understandable also because of VRM limitations specific of that MB (think of a microitx MB with support of up to 65W...)

 

[itsmydamnation]

Man i am so right it just hurts

my IVB @4.3 DDR3 1066 10-11-10-30 T2

CPU Mark This Computer 9230
Integer Math This Computer 19408
Floating Point Math This Computer 8121
Prime Numbers This Computer 19.8
Extended Instructions (SSE) This Computer 225.8
Compression This Computer 14193
Encryption This Computer 2024
Physics This Computer 359.7
Sorting This Computer 8723
CPU Single Threaded This Computer 2370

Memory Mark This Computer 2059
Database Operations This Computer 82.6
Memory Read Cached This Computer 27786
Memory Read Uncached This Computer 11720
Memory Write This Computer 6455
Available RAM This Computer 13141
Memory Latency This Computer 33.6
Memory Threaded This Computer 14229


Be gone you trolling, shill!

Just to be sure i tried to run my memory @ 2000mhz 20-22-20, i cant get post, or anything with such high numbers.
So i did 1066@ 7-7-7, which 100% confirms a strong memory latency to performance of the two "poor" performing sub tests:

Ivb@4.3                       2000mhz 10-11-10    1066mhz 10-11-10    1066 mhz 7-7-7
Database Operations           106                   82.6                  89
Physics                       597                   359.7               435
ns access time time aprox   10ns               20ns               13ns

edit: typo in the data

 

[CentroX]

posted?

 

[The Stilt]

I don't know for Ryzen, but last AMD BIOS and kernel developers guide (yes: i have fun reading this stuff even if i am not a bios or kernel developer) AFAIK, contained instructions on how to program the various P-states, including turbo P-states, that can be modified, at least in linux, even at run time. If you don't program correctly the P-states, it won't work. Probabily also the energy saving features will be broken.
If that AMD document specify how to program these p-states, i suppose that there are not default valuses. And this is understandable: if there is a very low end motherboard with weak VRMs, the bios is responsible of programming not too high p-states...

Yes, that worked fully until Excavator.
Excavator was the first design to implement the Shadow PStates. To make it unnecessary to fully redesign the whole power management, the implementation is much more simple than in Zen.
In Zen they are truly invisible and only the power management processor can access and control them.