Official AMD Ryzen Benchmarks, Reviews, Prices, and Discussion

[mikk]

This last bit may explain why Intel PR sent out a last-minute “call us before you write” email to most of the press,
^^ from semiaccurate?
My question what Intel mean it by writing Call us before you write.
link: http://semiaccurate.com/2017/02/22/amds-ryzen-7-1800x-beats-intels-i7-6900k-half-price/

It's possibly a lie because it's from Charlie, he hates Intel. What you need is a confirmation from a reliable source if Intel really send out messages to the press.

 

[JDG1980]

Someone should ask Ian Cutress on Twitter if Anandtech got a "call us before you write" email, and if so, what it said. (I don't have a Twitter account)

 

[lolfail9001]

XFR has to be enabled in windows. That came out in the Donanimhaber leak.

1. Do not remember Windows mentioned even once in any of transcripts.
2. XFR by design does not need enabling in OS. If they talk about enabling something, they are not talking about XFR.
3. These folks sound a whole lot like Turkish wccftech. So apply a mountain of salt if neccessary.

It's possibly a lie because it's from Charlie, he hates Intel. What you need is a confirmation from a reliable source if Intel really send out messages to the press.

They[Intel] probably did. Remember AMD informing press of price cuts and all that jazz before 1050's launch? Same story here, most likely.

 

[poohbear]

Nope, all are unlocked. There is speculation you're paying for a binned part however (that will overclock better).

all are unlocked?? I just don't see why anyone would pay the premium for the others then? The 300mhz difference must be ridiculously easy to achieve in an overclock. Speculation about binned CPUs doesn't make any sense based on how CPUs are created in the factory.

 

[lolfail9001]

The 300mhz difference must be ridiculously easy to achieve in an overclock.

Hence the speculation they are binned. For one, it looks like the "safe" frequencies are 3.8/4/4.2 for 1700/1700X/1800X.

 

[Arachnotronic]

It's possibly a lie because it's from Charlie, he hates Intel. What you need is a confirmation from a reliable source if Intel really send out messages to the press.

It's true, confirmed it with multiple sources.

 

[Kenmitch]

all are unlocked?? I just don't see why anyone would pay the premium for the others then? The 300mhz difference must be ridiculously easy to achieve in an overclock. Speculation about binned CPUs doesn't make any sense based on how CPUs are created in the factory.

Not sure what you mean by how they are created in the factory. Testing for functionality I'd imagine is a part of the process. Doing some automated power characteristic type of testing could easily bin the chips into different flavors rather quickly I'd imagine.

 

[.vodka]

all are unlocked?? I just don't see why anyone would pay the premium for the others then? The 300mhz difference must be ridiculously easy to achieve in an overclock. Speculation about binned CPUs doesn't make any sense based on how CPUs are created in the factory.

I heard in some youtube video from one of the guys at the event that there are things we still don't know about the X models, and that as it's known now, X doesn't mean XFR.

In similar fashion in one of those videos it was said that the X models will overclock better than the non X models, suggesting some binning is going on. I mean, there's a reason why the 1700 is so underclocked relative to the X models, it's probably that those dies don't meet the voltage/frequency specs for the 1700x or 1800x, it's plausible.

Remember that AMD is creating 8, 6, 4 core models out from a single 8 core die, so you can expect lower quality silicon as you drop down the core count.


We'll see about this last remaining mystery in the reviews...

 

[inf64]

It's true, confirmed it with multiple sources.

But what it actually means? Guidelines for benchmark selection (which would be very very wrong) or guidelines for how to setup intel systems optimally , ie. TB3.0 (which would be perfectly OK).

 

[inf64]

I heard in some youtube video from one of the guys at the event that there are things we still don't know about the X models, and that as it's known now, X doesn't mean XFR.

In similar fashion in one of those videos it was said that the X models will overclock better than the non X models, suggesting some binning is going on. I mean, there's a reason why the 1700 is so underclocked relative to the X models, it's probably that those dies don't meet the voltage/frequency specs for the 1700x or 1800x, it's plausible.

Remember that AMD is creating 8, 6, 4 core models out from a single 8 core die, so you can expect lower quality silicon as you drop down the core count.


We'll see about this last remaining mystery in the reviews...

What I will speculate below is my own opinion and I have no knowledge of any NDA stuff.

What if X in the name denotes better half core Turbo boost states? For example X models will boost to "ST" turbo clocks if only up to 4 cores(8 threads) are loaded? That would be a great distinction from non-X models and will help Ryzen perform admirably even in low threaded workloads, rivaling mainstream i5/i7s in those scenarios.

 

[lolfail9001]

But what it actually means? Guidelines for benchmark selection (which would be very very wrong) or guidelines for how to setup intel systems optimally , ie. TB3.0 (which would be perfectly OK).

Certainly latter, but the former is interesting one.

 

[Arachnotronic]

But what it actually means? Guidelines for benchmark selection (which would be very very wrong) or guidelines for how to setup intel systems optimally , ie. TB3.0 (which would be perfectly OK).

No it was just the canned statement that PCper published.

 

[inf64]

No it was just the canned statement that PCper published.

Hmm I missed that, was it published on their website or via twitter?

 

[bjt2]

I heard in some youtube video from one of the guys at the event that there are things we still don't know about the X models, and that as it's known now, X doesn't mean XFR.

In similar fashion in one of those videos it was said that the X models will overclock better than the non X models, suggesting some binning is going on. I mean, there's a reason why the 1700 is so underclocked relative to the X models, it's probably that those dies don't meet the voltage/frequency specs for the 1700x or 1800x, it's plausible.

Remember that AMD is creating 8, 6, 4 core models out from a single 8 core die, so you can expect lower quality silicon as you drop down the core count.


We'll see about this last remaining mystery in the reviews...

The 1700 is a 65W part.
What does it mean?
To have such a low TDP, even if it's binned at low frequency, it has to be a low leakage part. Namely mean leakage of each transistor is lower. At same process geometry, a low leakage transitor have higher threshold and/or lower transconductance. This is good for low clock, low voltage operation. But for higher clocks it requires more Vcore and hence they will draw more power than higher leakage parts at high frequency. Not only they will overclock less, but also at same frequency, greater than a certain threshold, they will draw more power.
So the lower binned parts probabily scale worse with voltage...

 

[bjt2]

What I will speculate below is my own opinion and I have no knowledge of any NDA stuff.

What if X in the name denotes better half core Turbo boost states? For example X models will boost to "ST" turbo clocks if only up to 4 cores(8 threads) are loaded? That would be a great distinction from non-X models and will help Ryzen perform admirably even in low threaded workloads, rivaling mainstream i5/i7s in those scenarios.

As per XFR description (and Bristol ridge shadow p-states), the frequency increment depends loosely by the load. Loosely in the sense that it depends from the power drawn, that of course is lower with a load that uses less cores. But instead of cores/threads used, we must consider the physical units exercised, that in turn translates to higher or lower power drawn.

 

[looncraz]

all are unlocked?? I just don't see why anyone would pay the premium for the others then? The 300mhz difference must be ridiculously easy to achieve in an overclock. Speculation about binned CPUs doesn't make any sense based on how CPUs are created in the factory.

Every process has its average defect density distribution. There are areas on the average wafer where chips will tend to have specific characteristics that are better/worse than the average on the wafer.

In this wafer map, ANY die with a color has a different mode of defect based on that color:

 

[Arachnotronic]

Hmm I missed that, was it published on their website or via twitter?

We take any competition seriously but as we’ve learned, consumers usually take a ‘wait and see’ approach on performance claims for untested products. 7th Gen Intel® Core™ delivers the best experiences, and with 8th Gen Intel Core and new technologies like Intel® Optane™ memory coming soon, Intel will not stop raising the bar.

https://www.pcper.com/news/Processo...r-Starts-Today-Specs-and-Performance-Revealed

8th gen core for desktops isn't coming until 1Q 2018, what are they talking about?

 

[inf64]

https://www.pcper.com/news/Processo...r-Starts-Today-Specs-and-Performance-Revealed

8th gen core for desktops isn't coming until 1Q 2018, what are they talking about?

Maybe Skylake-X? That is the only thing that will launch soon but I don't see how that will help them much. They will probably get ~15%-20 higher performance than 6900K with the higher clocked 16T Skylake X but the glaring problem of 2x cheaper 1800X (and maybe 1900X by that time?) that performs ~10% slower on average in consumer workloads , remains.

 

[Tuna-Fish]

maybe Zen+ will be quad channel?

It won't. AMD has committed to shipping future CPUs that are compatible with current AM4 boards. Also, the reason that Intel E-series is quad channel is that they share the dies with the server CPUs, and on server, more channels are more useful than on the desktop. AMD's server socket will support 8-channel CPUs.

 

[Arachnotronic]

Maybe Skylake-X? That is the only thing that will launch soon but I don't see how that will help them much. They will probably get ~15%-20 higher performance than 6900K with the higher clocked 16T Skylake X but the glaring problem of 2x cheaper 1800X (and maybe 1900X by that time?) that performs ~10% slower on average in consumer workloads , remains.

Skylake-X should help them in three ways:

1. Higher IPC than Broadwell-E (probably bigger gap between BDW-E and SKX-X than there is between BDW-E and SKL-S/KBL-S here, due to changes in cache sizes/bandwidth).

2. Skylake has a better SMT implementation than Broadwell/Haswell do, so SMT yield should improve, boosting MT competitiveness for a given core/frequency count.

3. Skylake-X should clock a lot better than Broadwell-E due to both process improvements as well as a much better physical implementation (Broadwell-E's was quite poor).

 

[raghu78]

I think Skylake-X is most likely to launch at Gamescom in Aug.

http://www.guru3d.com/news-story/skylake-x-will-be-launched-at-gamescom-2017.html

AMD could get another stepping out in 6 months time and bump the clocks up another 200 Mhz as GF 14 LPP process matures and AMD/GF use their Summit Ridge production ramp learning. I doubt Intel is going to be able to clock higher than AMD at any core count for a base clock comparison because of a larger core becoming even more larger with AVX-512. Intel could turbo higher as we see Kabylake run at 4.5 Ghz. But it remains to be seen as there is a lot of difference between clocking a 4C/8T chip and a 8C/16T chip to very high clocks. As die size increases its not easier to hit the same clocks as smaller chips do. I think Intel's slightly higher IPC will allow them to maybe edge ahead of higher clocked Summit Ridge SKUs. But still I don't think Intel can ever get back to USD 1000 for 8C/16T. The 8C/16T Skylake-X can sell for USD 600- USD 700 and maybe the 10C/20T can come in at USD 1000 and the rumoured 12C/24 T SKU can sell for USD 1300-USD 1400. Still its going to be a tough sell for those higher priced and higher core count SKUs against the flagship Summit Ridge 8C/16T SR SKUs from a perf/$ point of view. I am waiting to see what kind of base/turbo clocks Intel is going to hit for Skylake-X at each core count. I am also curious to know if the Skylake-X chips can overclock beyond 4.5 Ghz given their much higher core counts and larger cores due to AVX-512.

 

[PhonakV30]

Here link from SA Forum

Grabbed this from Drunkenmaster on OCUK. Some pretty interesting tidbits.

Some food for thought which I posted in the memory thread.

First off we had some rumours about weak IMC due to supposedly low speed. This was followed by a rumour that actually Zen had insane efficiency way above Intel and as such lower speeds actually had same performance as higher on Intel.

http://ranker.sisoftware.net/show_r....1b7c4f9c9&l=en

This is a Zen getting 33.99GB/s out of 2133Mhz memory, which has a theoretical maximum bandwidth of 34.128GB/s... meaning, epic efficiency.

For comparison you have

http://ranker.sisoftware.net/show_r....5b3c0fdcd&l=en

Broadwell-e with 3200Mhz memory achieving 74.97GB/s with max theoretical of 102.4GB/s, so around 75% efficiency.

http://ranker.sisoftware.net/show_r....4b2c1fccc&l=en

Skylake with similar (lower latency though) memory achieving 26.52GB/s at 2133Mhz, with obviously max theoretical of 34.128GB/s. A little over 75% efficiency. Another way to look at the results, if true, is that Zen has 28% higher bandwidth at the same memory speed.

If that scales at the same rate then 2666Mhz on Zen would give the same bandwidth that Skylake(and thus Kaby) would at 3400Mhz.

Along with those rumours was a mention that Zen was currently locked at 1t, we also have Asus managing to get higher memory speed. I wonder if Asus unlocked some settings that aren't supposed to be unlocked yet and in doing so managed to lower timings to up memory speeds, but sacrificing efficiency.

Basically I'm just trying to say, heads up, new platform, don't assume 2666Mhz is even bad, OR that 3200Mhz(at potentially much lower timings) is automatically better than 2666Mhz.

http://www.corsair.com/en-eu/blog/20...ddr4_synthetic

Posted this before, but another thing to bear in mind, Haswell tanked in bandwidth efficiency after 2400mhz. We absolutely do not know that super fast memory speeds will bring ANY better performance yet. I mean Sisoft doesn't usually get faked afaik, and the result fits with some rumour 1-2 weeks back, it would also explain most issues of getting higher memory speeds.

How did he get 33.99GB/s ? 33.99/34.128G = 0.9959

 

[.vodka]

That is some pretty freaking insane efficiency if true!

Who is providing the memory controller for AMD here? Rambus? Infinity fabric at work....?

 

[zinfamous]

http://imgur.com/a/Xl4gN

MSI X370 motherboard manual slides

Zen overclocks to 11. niiiiiiiice

EDIT: later learned that Intel MSI boards have been doing the same thing. Great, now I lazed my way into an obvious and likely very old observation.

 

[looncraz]

That is some pretty freaking insane efficiency if true!

Who is providing the memory controller for AMD here? Rambus? Infinity fabric at work....?

AMD rolled the memory controller, but the PHY is supposedly from RAMBUS (physical interface circuitry).

It appears that the PHY, the IMC, and the infinity fabric are all up to snuff - it would take all three to work nearly perfectly for the results we are seeing.