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

[inf64]

I have done my own "estimation" on Summit Ridge 8C/16T performance a while ago, I just waited for more information from AMD so I could see if I was near the expected numbers based on released data regarding the uarchitecture choices/features AMD opted for. Now when they released one benchmark and some new information on Zen, I am confident I can post this up now:

I used AT's article on Excavator and applied the ~40% ST uplift on Carrizo's number while adjusting for clocks I expect Zen will launch at (3.2Ghz base and 3.7Ghz ST Turbo). I also applied a 1.2x SMT boost to well threaded benchmarks and adjusted for scaling 12% penalty that Carrizo has when running 2 threads on a module( 0.88x scaling is an average, it varies from benchmark to benchmark). The rest of the numbers used in the sheet are from AT's bench page which is accessible from AT homepage.

This is just a rough estimate of course, Zen might end up at lower clocks and/or lower performance per clock than what I used in the table above. On the other hand, Zen might end up clocking even better and performing better than 40% . We will see on Computex 2017 I guess

Time to bump up my guesstimate now in the light of new information we have just gotten .
Almost feels like this Zen in my table above is "sandbagging" version if the one that was talked about in the previous posts has Haswell/Broadwell-like IPC .

 

[Abwx]

http://www.digitimes.com/news/a20161026PD206.html

AMD's Zen-based processors are expected to feature a price/performance ratio better than expected

Motherboard players are also aggressively preparing products based on the high-end X370, mainstream B350 and entry-level A320 chipsets. Motherboard players are expected to start shipping X370-based motherboards at the end of November

 

[jpiniero]

AMD's Zen-based processors are expected to feature a price/performance ratio better than expected

Considering what AMD has said, that would be quite a strategy shift from "done being the budget brand". Perhaps the 8 core around $300 then? That's what the 1090T went for at launch.

 

[Abwx]

Rather BDW-E level of perf at 50% of the price, something like 490$ for a 8C/16T and 220$ for the 4C/8T, that s in line with their current pricings since the latter should replace the most expensive FX parts.

 

[Ajay]

Rather BDW-E level of perf at 50% of the price, something like 490$ for a 8C/16T and 220$ for the 4C/8T, that s in line with their current pricings since the latter should replace the most expensive FX parts.

I sure hope AMD can get $500 (or more) for the 8 core Zen. They cannot afford to play to the low end and remain viable.
That said, they price will be very performance dependent relative to Intel.

 

[moonbogg]

Hopefully Zen is an amazing deal and everyone buys them. That way I can get an 8 core Skylake for less than a grand. I'd buy Zen, but I wouldn't be happy with broadwell ipc. As long as I keep my CPU's these days, I can't afford to settle for less ipc.

 

[witeken]

Interesting how people keep comparing Zen IPC to Broadwell.

Lol, we'll see .

 

[scannall]

Interesting how people keep comparing Zen IPC to Broadwell.

Lol, we'll see .

My guess is around Ivy Bridge, which is what they have been saying forever. I doubt they are sandbagging it. That wouldn't make much sense. But, IB IPC, with 8 cores and 16 threads for a decent price should sell pretty well.

 

[witeken]

We'll see in February (removed due to user feedback)

 

[cytg111]

We'll see in February.

Fixed it. The rest you could have done without.

 

[witeken]

Fixed it. The rest you could have done without.

My post is neutral. As far as I know, crushing expectations of people is the opposite of shattering them .

 

[Doom2pro]

Interesting how people keep comparing Zen IPC to Broadwell.

Lol, we'll see .

I'd expect the first iteration of GloFo's 14LPP to be close to Broadwell IPC, and I expect future Zen+ to be well above it, perhaps well above what Intel could offer considering their precarious foundry/node situation compared to what the rest of the industry is doing/planning to offer.

 

[witeken]

I'd expect the first iteration of GloFo's 14LPP to be close to Broadwell IPC, and I expect future Zen+ to be well above it,

Wait, isn't Kaby Lake IPC not just 5% higher than BDW according to AnandTech? So well above BDW would mean well above KBL.

 

[Arachnotronic]

40% faster per clock than XV for Zen puts Broadwell at around 19% faster clock for clock relative to a Zen.. The delta grows to over 30% for SKL vs hypothetical Zen in GB4. Unless AMD dramatically beat its 40% target, Broadwell IPC on average is not realistic to expect from Zen.

 

[Arachnotronic]

Wait, isn't Kaby Lake IPC not just 5% higher than BDW according to AnandTech? So well above BDW would mean well above KBL.

AnandTech's testing was frankly very poor and put Skylake in a bad light with respect to Haswell and Broadwell.

 

[jpiniero]

40% faster per clock than XV for Zen puts Broadwell at around 19% faster clock for clock relative to a Zen.. The delta grows to over 30% for SKL vs hypothetical Zen in GB4. Unless AMD dramatically beat its 40% target, Broadwell IPC on average is not realistic to expect from Zen.

I should point out that Bristol Ridge gets killed in the HTML 5 DOM score presumably due to not having L3 cache. I'm beginning to think the infamous 40% is specifically on floating point and not on integer though.

http://browser.primatelabs.com/v4/cpu/compare/884997?baseline=637037

AnandTech's testing was frankly very poor and put Skylake in a bad light with respect to Haswell and Broadwell.

Not really; Skylake does get a nice performance boost in some tests when using faster DDR4 but using more than 2133 would be considered overclocking.

 

[Arachnotronic]

I should point out that Bristol Ridge gets killed in the HTML 5 DOM score presumably due to not having L3 cache. I'm beginning to think the infamous 40% is specifically on floating point and not on integer though.

http://browser.primatelabs.com/v4/cpu/compare/884997?baseline=637037



Not really; Skylake does get a nice performance boost in some tests when using faster DDR4 but using more than 2133 would be considered overclocking.

A9X doesn't have an l3 cache and it handles the HTML DOM workload just fine -- solidly faster than the A9.

 

[Ajay]

Not really; Skylake does get a nice performance boost in some tests when using faster DDR4 but using more than 2133 would be considered overclocking.

That the deal with SKL - really needs fast DDR4 for Benchmarks, IRL, I expect the difference as so significant - unless one spends all their time doing video editing and such.

 

[bjt2]

A9X doesn't have an l3 cache and it handles the HTML DOM workload just fine -- solidly faster than the A9.

A9 has 2 or 3MB (I don't remember) of fast(er than L3?) L2 for 2 cores... I think it's enough to counterbalance the lack of L3...
BR has only 1MB of L2 for 2 cores and we know how slow it is due to sharing...

 

[FIVR]

Hopefully Zen is an amazing deal and everyone buys them. That way I can get an 8 core Skylake for less than a grand. I'd buy Zen, but I wouldn't be happy with broadwell ipc. As long as I keep my CPU's these days, I can't afford to settle for less ipc.

If you are waiting for Intel to drop prices after AMD releases Zen you will be disappointed. Even when intel was losing nearly every benchmark and was being panned for its 32 stage pipeline and terribly expensive RAMBUS ram they still charged 2x what AMD did for a comparable sku. This back when the P4C Northwood and Prescott were out, and intel was paying dell hand over fist to please please please not buy athlons even if they perform better.


Intel will continue to charge exorbitant pries while conducting its usual anti-competivie attempts to squeeze OEMs and retailers. Intel does not participate in "price wars" and never has. It's bad for margins, which is bad for share price, which in turn is bad for intel execs. Loss of sales and such just requires them to do lower-level layoffs and who gives a crap about 2000 H1Bs sent back to india or wherever? Not intel, that's who.

 

[TheELF]

40% faster per clock than XV for Zen

It's not 40% faster,it's 40% more IPC,that's throughput not speed,it will only be 40% faster if you actually find a software that will be able to use all 10 instructions the ZEN core has available per cycle.
Which will be pretty difficult since there aren't many CPUs out there (if there are any) with 10 instructions per core,I guess that's why they went with blender instead of some "traditional" benchmark.

 

[Abwx]

It's not 40% faster,it's 40% more IPC,that's throughput not speed,it will only be 40% faster if you actually find a software that will be able to use all 10 instructions the ZEN core has available per cycle.
.

You are wrong on this one, at HC AMD s Clark explicitely said that the 40% is the gain on a single thread, so it s definitly not throughput, this latter is still unknown safe for Blender where it s comparable to a BDW-E.

 

[Arachnotronic]

It's not 40% faster,it's 40% more IPC,that's throughput not speed,it will only be 40% faster if you actually find a software that will be able to use all 10 instructions the ZEN core has available per cycle.
Which will be pretty difficult since there aren't many CPUs out there (if there are any) with 10 instructions per core,I guess that's why they went with blender instead of some "traditional" benchmark.

The Blender benchmark seems like a best case for Zen. The Blender build that AMD used did not use FMA or AVX2, so the Zen core had the advantage of having 2 FMUL and 2 FADD units while Broadwell has the equivalent of half of that in such code.

Stilt talked about unusually high SMT yield and I bet that SMT helped the Zen core fully utilize those execution units quite nicely.

Anyway if AMD's staged demo puts it in a statistical dead heat with Broadwell, expect it to fall far below in real world applications. I stand by my prediction of single thread perf/MHz being around 19% in favor of Broadwell.

 

[Abwx]

The Blender benchmark seems like a best case for Zen. The Blender build that AMD used did not use FMA or AVX2, so the Zen core had the advantage of having 2 FMUL and 2 FADD units while Broadwell has the equivalent of half of that in such code.

There s no advantage in a dual FPM/FPADD if the individual threads code is highly serialised, wich is apparently the case for Blender..

Stilt talked about unusually high SMT yield and I bet that SMT helped the Zen core fully utilize those execution units quite nicely.

And if you took attention rather than re interpreting according to wishfull thouhgts you would knew that Intel get 50-60% SMT gain in Blender, so i m not sure that it s AMD that benefited much of multithreading...

Besides Clark stipulated that in ST the thread would benefit from all ressources, because it s obvious that in SMT the first thread lose throughput but that the sum is superior to a single thread throughput, that s not documented but Intel s HT doesnt work by providing 30% higher throughput over say 100% throughput of a single thread, reality is that the second thread eat in the first thread throughput and that with HT the repartition is rather 90 + 40 and not 100 + 30.

Anyway if AMD's staged demo puts it in a statistical dead heat with Broadwell, expect it to fall far below in real world applications. I stand by my prediction of single thread perf/MHz being around 19% in favor of Broadwell.

If your "estimation" was right it would mean that AMD s SMT bring 85-90% gain in Blender, and generaly 19% more gain than Intel s HT, as you can see previsions based on wishfull thoughts end being fairy tales.

 

[TheELF]

You are wrong on this one, at HC AMD s Clark explicitely said that the 40% is the gain on a single thread, so it s definitly not throughput, this latter is still unknown safe for Blender where it s comparable to a BDW-E.

40% is the gain on a single thread if the single thread can use all 10 instructions,it's the same thing I said.
This is the throughput of one single thread running on one single core.