Sandbo

引用:

原帖由 Puff 於 2017-3-8 23:39 發表
打算落搭 1700X
iMac 十世都唔出


過兩日落場睇吓有乜板同有冇貨先
唔記得之前 sandro 推果隻 modular 牛牛乜型號添

If you were referring to me, I always suggest Corsair RMx series (make sure you don't buy the old RM series).
I think they have the best performance given the price (not efficiency wise, but ripple performance).
650W: https://www.techpowerup.com/reviews/Corsair/RM650i/10.html
1000W: https://www.techpowerup.com/reviews/Corsair/RM1000x/9.html
*RMx and RMi use the same components and design, except for RMi can be connected to the PC for doing some monitoring.

But if you aim higher, now we have the best from Seasonic (and if I am right, they made their warranty 12 years long......while Corsair's still has 10)
https://www.techpowerup.com/reviews/Seasonic/Prime_750/6.html

[ 本帖最後由 Sandbo 於 2017-3-10 03:07 編輯 ]

Puff

引用:

原帖由 Sandbo 於 2017-3-10 02:58 發表

If you were referring to me, I always suggest Corsair RMx series (make sure you don't buy the old RM series).
I think they have the best performance given the price (not efficiency wise, but ripple p ...

Yep, sorry for misremembering your username.
Leaning towards OC 1700 with B350, perhaps 3.7-3.8 GHz all cores. Just that the idea of burning more power when browsing web pages kinda bugs me. Not sure if idle downclocking still works in OC mode.


[ 本帖最後由 Puff 於 2017-3-10 03:12 編輯 ]

Sandbo

引用:

原帖由 Puff 於 2017-3-10 03:02 發表

Yep, sorry for misremembering your username.
Leaning towards OC 1700 with B350, perhaps 3.7-3.8 GHz all cores. Just that the idea of burning more power when browsing web pages kinda bugs me. Not sure ...

A quick search gave me this:
https://www.bit-tech.net/hardwar ... zen-7-1700-review/6
On overclocking, the idle power does go up a bit, but it is just on par with 7700k's idle......far from being a worry.
While it seems every 1700 you can buy now overclock pretty well, there is no guarantee yours is going to be the same, just bare that in mind

*And I stand corrected - the max load power is higher than my 6700k overclocked to 4.4 GHz, but not by far.
With this it means you really would like a water cooler (some Corsair H80 or above) to proceed with clocking it at 4 GHz.

[ 本帖最後由 Sandbo 於 2017-3-10 08:22 編輯 ]

Sandbo

So it seems those available AM4 boards are already supporting ECC RAM to a certain extent:
https://www.reddit.com/r/Amd/com ... _and_other/def6vs2/

After some confirmation later I will probably upgrade my first gen Xeon E3-1235 to Ryzen 1700.

[ 本帖最後由 Sandbo 於 2017-3-10 11:32 編輯 ]

Puff

引用:

原帖由 Sandbo 於 2017-3-10 08:18 發表

A quick search gave me this:
https://www.bit-tech.net/hardware/2017/03/08/amd-ryzen-7-1700-review/6
On overclocking, the idle power does go up a bit, but it is just on par with 7700k's idle......far  ...

Probably not 4Ghz, since I heard hitting 3.9+ GHz needs a high-end board. I'd like somewhere around 3.7 and tune the voltage for perf-watt.
Saw some nice results in power consumption.


[ 本帖最後由 Puff 於 2017-3-10 16:08 編輯 ]

Puff

https://forum.beyond3d.com/posts/1970348/
DF: 32B bi-directional @ memclk

[ 本帖最後由 Puff 於 2017-3-10 20:06 編輯 ]

Sandbo

引用:

原帖由 Puff 於 2017-3-10 19:51 發表
https://forum.beyond3d.com/posts/1970348/
DF: 32B bi-directional @ memclk

Not knowing much about this, but there is a thread about the interconnect being the bottleneck for some cases:
https://www.reddit.com/r/Amd/com ... g_like_a_dual_4c8t/
It might not be a serious issue given the task is distributed smartly by the OS.

Puff

引用:

原帖由 Sandbo 於 2017-3-11 01:04 發表

Not knowing much about this, but there is a thread about the interconnect being the bottleneck for some cases:
https://www.reddit.com/r/Amd/com ... ving_like_a_dual_4c ...

It is not really a "design flaw" but apparently an intentional design choice (no system LLC). Yes, it is a bottleneck, but only for implementations that assume a giant shared LLC for all cores, and ping-pong the working set between cores on top of the assumption. Otherwise, it is no different from any UMA chips without a shared LLC running at core clock.

qcmadness

引用:

原帖由 Puff 於 2017-3-11 02:10 發表

It is not really a "design flaw" but apparently an intentional design choice (no system LLC). Yes, it is a bottleneck, but only for implementations that assume a giant shared LLC for all cores, and p ...

The OS needs to be NUMA-aware to fully support Ryzen

Puff

引用:

原帖由 qcmadness 於 2017-3-11 15:59 發表

The OS needs to be NUMA-aware to fully support Ryzen

It needs to be aware of the cache topology, not NUMA.

qcmadness

引用:

原帖由 Puff 於 2017-3-11 16:54 發表

It needs to be aware of the cache topology, not NUMA.

Both

Puff

引用:

原帖由 qcmadness 於 2017-3-11 17:13 發表

Both

The chip is UMA, and doesn't statically partition the physical memory address space betweeb CCXs. So essentially apps just need to be aware of the topology, e.g. porting console CPU scheduling code to the PC. XB1 and PS4 have essentially the same situation.

NUMA awareness isn't a must. Let alone the fact that the chip is not advertising itself as two NUMA nodes, and thus NUMA aware apps relying on OS API would not treat it as two.

[ 本帖最後由 Puff 於 2017-3-11 18:21 編輯 ]

Sandbo

引用:

原帖由 qcmadness 於 2017-3-11 15:59 發表

The OS needs to be NUMA-aware to fully support Ryzen

It might end up locking everything to a single CCX and uses only four core that case, this isn't ideal.

There is a new article from PCProspective :
https://www.pcper.com/reviews/Pr ... utm_medium=facebook

They tested and claimed the Windows 10 scheduler wasn't to blame for swapping task among core, but I don't think it's very conclusive; people consistently found Windows 7 to perform better with Ryzen, there might be something else behind the OS.

[ 本帖最後由 Sandbo 於 2017-3-12 09:26 編輯 ]

qcmadness

引用:

原帖由 Sandbo 於 2017-3-11 23:11 發表

It might end up locking everything to a single CCX and uses only four core that case, this isn't ideal.

There is a new article from PCProspective :
https://www.pcper.com/reviews/Processors/AMD-Ryzen ...

There are other claims that Windows 10 misreports the cache size for Ryzen CPUs

qcmadness

引用:

原帖由 Puff 於 2017-3-11 18:09 發表

The chip is UMA, and doesn't statically partition the physical memory address space betweeb CCXs. So essentially apps just need to be aware of the topology, e.g. porting console CPU scheduling code t ...

Cat. it as NUMA simplifies the problem.
But don't treat NUMA is an advertisement as NUMA reduces efficiency and possibly performance.