New Zen microarchitecture details

[itsmydamnation]

The deficiency that AMD has regarding memory seems to carry out to their GPUs, with either memory utilization being low or uarch being inefficient in case of Fury parts.

That as a statement doesn't even make any sense. If you understood how GCN works from a threading model you would understand fury's issue. Fiji in all likelihood is a horrible stopgap created by the poor performance of 20nm and AMD not backporting 20nm design to 28nm ( there is amost nothing in fiji that doesn't exsist in another 28nm part). Memory bandwidth and utilization isn't an issue, if anything memory size appears to be far more of an issue.

i dont see "only" 3000mhz DDR4 as an issue especially if you can take fast ram run it at 3000mhz and reduce CAS. Low ILP miss cares about latency in NS and nothing else. i would be surprised if there a very many real world workloads that dual channel 3000mhz memory will be the bottleneck to a 8 core chip with 16mb of L3 and "only" 256/128 bit load store a cycle.

 

[Yuriman]

^ Frequency is part of latency.

Not that I disagree with you, just sayin'.

 

[itsmydamnation]

^ Frequency is part of latency.

Not that I disagree with you, just sayin'.

I know, but high frequency memory has longer (in cycles) access/refresh/etc times. Now if the memory controller cant handle a faster clock rate but your memory can handle a reduction in latencies to go with its reduction in clock you can come out at the same total latency (in NS), just with reduced throughput(depending on the read/write pattern).

 

[AtenRa]

nm

 

[coercitiv]

I was just pointing out that in case of AMD, even with AM4 the max speed is DDR4-2400, overclocked.

How do we know that?

 

[IntelUser2000]

How do we know that?

That's obviously a mistake, sorry. DDR4-2933 is OC, and DDR4-2400 is stock.

 

[ShintaiDK]

That is for overclocked ram per motherboard. The header should have given that away. Nice that the memory can be overclocked like that and all, but the max supported un-overclocked speed is 2133 per Intel's Ark Info.

All these combinations are certified and warranted.

The 2133 is for any module, any combination.

 

[zentan]

Wrong.
http://www.intel.com/content/dam/ww...emory-for-intel-core-processors-datasheet.pdf

Yeah,some people think that memory OC support is some mobo oem magic,it's the chip designer(intel in this case) who creates/supports the memory divider.Of course some boards would have better mem OC support that others like OC formula but that doesn't mean it was some asrock magic not supported by Intel at the first place.
Officially it's 2133MHz for SKL,but for OC it's been 4000MHz(+) and that's been validated by intel and present in their XMP doc long ago.

 

[coercitiv]

All these combinations are certified and warranted.

Is DDR4 3000 on Skylake an overclock or not?

 

[ShintaiDK]

Is DDR4 3000 on Skylake an overclock or not?

Depends on the combination.

 

[ShintaiDK]

Yeah,some people think that memory OC support is some mobo oem magic,it's the chip designer(intel in this case) who creates/supports the memory divider.Of course some boards would have better mem OC support that others like OC formula but that doesn't mean it was some asrock magic not supported by Intel at the first place.
Officially it's 2133MHz for SKL,but for OC it's been 4000MHz(+) and that's been validated by intel and present in their XMP doc long ago.

It was also impossible to validate Skylake for any module above 2133 since it didn't exist at the time as a production standard.

 

[coercitiv]

Depends on the combination.

If you need to have a combination validated as an Extreme Memory Profile, then it's an overclock.

 

[Dresdenboy]

I think it's just another angle to pick on something. After over 40 years of computer repairs and upgrades, I've discovered that if you're sober and pay attention then things don't get bent.

+1

I'm also wondering a bit about this discussion. Maybe what you say about handling electronic parts worth >$100 didn't become common knowledge among enthusiasts yet.

OTOH I know people who were born with or got a good amount of clumsiness, due to many reasons like ADD, or pregnancy.

 

[The Stilt]

I would imagine the highest memory frequency officially supported on Zen's memory controller is 2666MHz. Hopefully on Raven Ridge it can reach significantly higher (similar to Skylake) since otherwise the it will be the same old story again :'(

 

[coercitiv]

I would imagine the highest memory frequency officially supported on Zen's memory controller is 2666MHz. Hopefully on Raven Ridge it can reach significantly higher (similar to Skylake) since otherwise the it will be the same old story again :'(

Is 2666 a Jedec standard yet? From their docs I concluded 2400 is, but 2666 & 3200 seem to be in TBD status.

 

[ShintaiDK]

Zen CPUs would need luck to get 2666. The APUs however is likely to get it due to their timeframe.

 

[PPB]

The article clearly states 2400 non OC/2933 OC for Summit Ridge and we can only suspect that those speeds at least will carry on to Zen, if not improve. Why I'm still reading BS such as that the IMC wont even do 2400mhz? :\

 

[ShintaiDK]

Summit Ridge is Zen

Bristol Ridge is renamed Carrizo.

 

[PPB]

More to my point. It will at least behave like HW-E regarding RAM speeds if 2400/2933 is true.

 

[dark zero]

And here i was expecting ddr4>3000mhz

It won't happen in a very long time....

 

[monstercameron]

It won't happen in a very long time....

only sith deal in absolutes :sneaky:

 

[AtenRa]

http://wccftech.com/amd-am4-%C2%B5opga-socket-1331/

 

[Soulkeeper]

I'd build a Bristol Ridge system if there was confirmation that I could later plop in a Summit Ridge upgrade. That'd be great.

 

[William Gaatjes]

De maximum memory speed is not only determined by just the memory controller of a given cpu.

It is of course the ram but also the motherboard.
I assume AMD delivers ( or hires an external company for) a reference design for the motherboard with optimized copper trace layout to reduce noise, ringing and undershoot/ overshoot. That design must also be tested and qualified to reach a given maximum memory speed. Of course, with the memory parameters a lot can be done. Drive strength, delay, slew rate, impedance matching. But only so much for optimum results.

 

[ShintaiDK]

I'd build a Bristol Ridge system if there was confirmation that I could later plop in a Summit Ridge upgrade. That'd be great.

So you are willing to buy a system for future upgrade without knowing how a future CPU will perform? D: