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

[richierich1212]

Bulldozer@2GHz? OMG AMD is doomed!

No, Faildozer version 2.0.

Trolling is not allowed
Markfw
Anandtech Moderator

 

[CHADBOGA]

Don't Jinx us, please!!!!

I have invited him here, but he has not as of yet, accepted the invite.

 

[thedavexp]

Supreme commander replay, 80k map with 5000 unit cap and 7 AIs would be a good test for single core performance I have not ever had a PC that could cope with that at 1X simulation speed not even close. My i7 6700k clocked at 4.8ghz was the closest I got.

 

[DrMrLordX]

Y is he not here, he is usually everywhere.

What makes you think he isn't here already . . .

How about Feb 29th?

I kid, I kid...

In that case . . .

It's not one day.. it's 3 years. Boom Tiger Lake!

Woops you beat me to it.

He is a legend in his own mind.

At this point, he is a legend, period. Not for anything good mind you.

My question was more of a joke because I know people can't break NDAs.

IN SOVIET RUSSIA, NDA breaks YOU!!!!

My fear is that they're all fake and ryzen totally sucks on launch day

Reality is mealeable. He-who-can-not-be-named will warp spacetime, lobotomize Keller, and transform Summit Ridge into a steaming pile of garbage. Those of us who buy the CPU anyway will be rounded up for re-education. All will be well.

I have invited him here, but he has not as of yet, accepted the invite.

Anyway it's probably two weeks to NDA lift and launch so bring on the reviews. When do you guys think NewEgg will offer pre-orders? Or someone similar?

 

[Dresdenboy]

No, Faildozer version 2.0.

I knew what you actually meant.

 

[Dresdenboy]

Do we know what ram the Intel 8c was using.

Ryzen was using 2400. 17 17 17 I think.

We can normalize it using all these data and remove latancy effects.

At least for the 6800K used at Videocardz, it is 3200 16-18-18:
https://cdn.videocardz.com/1/2017/0...manceTest-9.0-Evaluation-Version-1000x866.png

IIRC Atkin sieve modulo operations are all by constants and so should be optimized away to faster operations by any decent compiler.

This needs to be checked. But it could be exactly that what's causing the mem latency dependency, as optimizing away the slow operations would shift the bottleneck somewhere else.

 

[RichUK]

2 weeks to go for Ryzen reviews to be released.

http://wccftech.com/amd-ryzen-reviews-live-february-28th-testers-receiving-samples/

 

[zinfamous]

i know we donot trust WCC, but they are claiming that 28th Feb is the release date? "With shipments expected to commence on the 28th"

i though 28th was the NDA lift date.

I'm visiting US between 18th Feb and 5th March. I just hope there will be enough reviews to show its worth making the purchase from there. Prices in UK are quite a bit inflated by tax. Then buy a motherboard at a later date once all the boards have been tested

I hope you're near a Microcenter or Fry's when you're visiting, because that would be your best chance, most likely. ...or I guess you could just place a pre-order on Amazon or Newegg and have it delivered to someone you know, or an Amazon delivery box near wherever you will be?

 

[zinfamous]

So all the leaks are all total BS.... The final samples haven't even gone out yet or not till later tonight!!

engineering samples have been out for some time to mobo manufacturers and others. the clock variance would be an issue, obviously, so you have to accept leaked benchmarks for what that chip was able to do, and compare that to leaked claims of what release clocks will be. benches can properly appear real or unreal because the chips themselves are different iterations of engineering samples, so you always have to take that with a grain of salt. There's a bit more reason to trust the majority of them than to distrust them, but certainly every reason to take it all with a grain of salt at the same time.

 

[lopri]

I still say that Oblivion's Imperial Market District (raining weather @ noon) is the best way to check single thread performance.

That is an odd choice. The game is way too old at this point and I do not think that scene is taxing enough. Besides which how do you control NPCs randomly teleporting in and out?

 

[tamz_msc]

That is an odd choice. The game is way too old at this point and I do not think that scene is taxing enough. Besides which how do you control NPCs randomly teleporting in and out?

A better choice would be Counter Strike: Global Offensive with ultra-low config and a custom match with a set number of bots.

 

[bjt2]

At least for the 6800K used at Videocardz, it is 3200 16-18-18:
https://cdn.videocardz.com/1/2017/0...manceTest-9.0-Evaluation-Version-1000x866.png


This needs to be checked. But it could be exactly that what's causing the mem latency dependency, as optimizing away the slow operations would shift the bottleneck somewhere else.

Someone that has a 6800K and RAMs that reach at least 3200 16-18-18 could do the tests both with 3200 16/18 ram and 2400 17-17 ram...

 

[CentroX]

I wonder what todays ryzen leaks will be

 

[Glo.]

Do we know what ram the Intel 8c was using.

Ryzen was using 2400. 17 17 17 I think.

We can normalize it using all these data and remove latancy effects.

The RAM used in Ryzen tests was ECC, on AM4 platform. All 2400 MHz RAM has CL17 timings. I have not spotted anything like this in "normal" RAM.

 

[CentroX]

RAM speed does not matter anymore. The performance difference between 2400 and 3200 is like 1 fps.

 

[tamz_msc]

RAM speed does not matter anymore. The performance difference between 2400 and 3200 is like 1 fps.

That is just false.
https://www.youtube.com/watch?v=Er_Fuz54U0Y

 

[Nothingness]

This needs to be checked. But it could be exactly that what's causing the mem latency dependency, as optimizing away the slow operations would shift the bottleneck somewhere else.

The algorithm is not very difficult to follow: https://en.wikipedia.org/wiki/Sieve_of_Atkin The only mod operations are by 60 and these can be replaced by some mul and other non-div ALU ops (look "division by constant" with Google).

If they are segmenting the sieve then some non-constant mod will appear but they should be negligible. And anyway if their limit indeed is 32 million, the sieve doesn't have to be segmented (unless they did heavy tuning of the algo to play with L1 cache). Next thing to note is that they say the test uses 4MB of memory, so definitely main memory should not have a big impact. OTOH cache latency might have some effect.

Anyway it's hard to say what exactly their benchmark does without reverse engineering it or try to run it with external profiling tools (such as Vtune).

 

[itsmydamnation]

The algorithm is not very difficult to follow: https://en.wikipedia.org/wiki/Sieve_of_Atkin The only mod operations are by 60 and these can be replaced by some mul and other non-div ALU ops (look "division by constant" with Google).

If they are segmenting the sieve then some non-constant mod will appear but they should be negligible. And anyway if their limit indeed is 32 million, the sieve doesn't have to be segmented (unless they did heavy tuning of the algo to play with L1 cache). Next thing to note is that they say the test uses 4MB of memory, so definitely main memory should not have a big impact. OTOH cache latency might have some effect.

Anyway it's hard to say what exactly their benchmark does without reverse engineering it or try to run it with external profiling tools (such as Vtune).

Just guessing (watching taskmgr) they run it on all threads in parallel so on an 8T processor its really a 32mb benchmark, that would make both cache latency and memory latency important.

 

[coercitiv]

Someone that has a 6800K and RAMs that reach at least 3200 16-18-18 could do the tests both with 3200 16/18 ram and 2400 17-17 ram...

The difference in Prime numbers will be in the realm of 10-15%, based on my own tests.

Core i5 6600k @ 4Ghz:
2133 CL 15 - Prime scores 34-36
3000 CL 15 - Prime scores 38-41

 

[Dresdenboy]

Just guessing (watching taskmgr) they run it on all threads in parallel so on an 8T processor its really a 32mb benchmark, that would make both cache latency and memory latency important.

Not to forget: 16T means 64GB then! And this would cause writes too. The question is: Is there some sharing of data? Of course, this only applies, if they don't work on these 4MB with all threads.

I'll make another plot with RAM B/W soon.

I hope at the end of each of the next days we don't have to sing:
No leaks today, my time is gone away,
the sample waits forlorn, the chip not even warm.

 

[lobz]

A better choice would be Counter Strike: Global Offensive with ultra-low config and a custom match with a set number of bots.

absolutely. my 300-500 fps drops massively from about the 8th bot, even though I have a 4790K @ 4.8 (and a gtx 1070 alleviates every concern about GPU bound problems).

 

[coercitiv]

absolutely. my 300-500 fps drops massively from about the 8th bot, even though I have a 4790K @ 4.8 (and a gtx 1070 alleviates every concern about GPU bound problems).

8 bots 8 threads...

 

[Dresdenboy]

The difference in Prime numbers will be in the realm of 10-15%, based on my own tests.

Core i5 6600k @ 4Ghz:
2133 CL 15 - Prime scores 34-36
3000 CL 15 - Prime scores 38-41

Being a 4T machine having 2 fast DIMMs the analysis might become skewed a bit. I'll include your numbers, too. Maybe cache thrashing causes a perf stall/drop (per core) on SMT machines.

 

[zinfamous]

I wonder what todays ryzen leaks will be

AMD has been working closely with the Russians, all along.

 

[Agent-47]

The algorithm is not very difficult to follow: https://en.wikipedia.org/wiki/Sieve_of_Atkin The only mod operations are by 60 and these can be replaced by some mul and other non-div ALU ops (look "division by constant" with Google).

If they are segmenting the sieve then some non-constant mod will appear but they should be negligible. And anyway if their limit indeed is 32 million, the sieve doesn't have to be segmented (unless they did heavy tuning of the algo to play with L1 cache). Next thing to note is that they say the test uses 4MB of memory, so definitely main memory should not have a big impact. OTOH cache latency might have some effect.

Anyway it's hard to say what exactly their benchmark does without reverse engineering it or try to run it with external profiling tools (such as Vtune).

the algorithm has a lot of If statements. one every few steps to see if the conditions are met. that's a lot of branching. remember the Flitz Chess benchmark? it had poor results too (8c vs 4c comparison).