Re: [gentoo-user] new machine : CPU : 22 nm vs 32 nm

[Michael Mol]

On Fri, Jul 27, 2012 at 8:30 PM, microcai <microcai@×××××××××××××.org> wrote:
> 2012/7/26 Michael Mol <mikemol@×××××.com>:
>> On Thu, Jul 26, 2012 at 10:18 AM, Евгений Пермяков <permeakra@×××××.com> wrote:
>>> On 07/26/2012 05:50 PM, Michael Mol wrote:
>>>>
>>>> On Thu, Jul 26, 2012 at 9:43 AM, Евгений Пермяков <permeakra@×××××.com>
>>>> wrote:
>>>>>
>>>>> On 07/26/2012 12:05 AM, Philip Webb wrote:
>>>>>>
>>>>>> I've listed what's available at the local store,
>>>>>> which I trust to stock reliable items, tho' I wouldn't ask their advice.
>>>>>>
>>>>>> All the AMD's are  32 nm , while the Intel recommended by one commenter
>>>>>> -- Core i5-3570 4-Core Socket LGA1155, 3.4 Ghz, 6MB L3 Cache, 22 nm --
>>>>>> is  22 nm : it costs  CAD 230  & they have  3  in stock,
>>>>>> which suggests demand, but not the most popular ( 9  in stock).
>>>>>>
>>>>>> Isn't  22 nm  going to be faster than  32 nm  ?
>>>>>>
>>>>>> In the same price range, AMD offers  Bulldozer X8 FX-8150 (125W)
>>>>>>    8-Core Socket AM3+, 3.6 GHz, 8Mb Cache, 32 nm  ( CAD 220 ,  2  in
>>>>>> stock).
>>>>>>
>>>>>> How do you compare cores vs nm ?
>>>>>> How far is cache size important ( 6 vs 8 MB )?
>>>>>>
>>>>>> When I built my current machine 2007, the CPU cost  CAD 213 ,
>>>>>> so both look as if they're in the right ballpark.
>>>>>>
>>>>> If you're building new, performance-oriented box, you should take latest
>>>>> intel with AVX because of AVX.  As I recall, recent gcc has support for
>>>>> avx,
>>>>> so some performance gain may be achieved.
>>>>> If you want home box, you may be interested in AMD A8 and similar chips,
>>>>> as
>>>>> they are reasonably fast and very chip
>>>>
>>>> AMD parts have had AVX since the Bulldozer core release in Q3 2011.
>>>
>>> Are they already available in reasonable numbers on market?
>>
>> http://www.cpubenchmark.net/cpu.php?cpu=AMD+FX-8120+Eight-Core
>>
>> At $150, fitting into existing Socket AM3+ boards, that looks like the
>> best part for my money right now.
>>
>>>>> In any case, I'd put most of my money in 2-4 big 3Tb HDD's for media and
>>>>> 8+
>>>>> Gb fast memory, as modern browsers eat memory like crazies and CPU is
>>>>> usually fast enough. Decoding HDTV mkv's should occur on gpu block in any
>>>>> case, so general performance for most uses is irrelevant, as it was fast
>>>>> enough four yesrs earlier. Simply check, that you can offload HDTV
>>>>> decoding
>>>>> to GPU in your config.
>>>>
>>>> Here, you're talking about either VDPAU or VAAAPI support. VDPAU is
>>>> only offered by nVidia cards, and even then you need to run the
>>>> proprietary driver. VAAPI is supported by Intel graphics and ATI's
>>>> proprietary driver.
>>>
>>> I do not see any problems with this. A blob in system is not best practice,
>>> of course, but it does not need any configuration and is not a performance
>>> bottle-neck, so there is no reason to care.
>>
>> I only bring it up because some people do care. I'm running fglrx at
>> home right now. When I run nVdia, I run the nVidia drivers. In part
>> because I like accelerated video decoding (which a Geforce 210 does
>> wonderfully), in part because the nv, nouveau and radeon drivers
>> historically worked very poorly for me in 2D performance when faced
>> with multiple 1080p displays. They're always getting better, of
>> course.
>>
>>>
>>> I personally would prefer AMD A8 if I can offload decoding to GPU unit there
>>> (not sure if I can, so won't change my box till next summer), but discrete
>>> video card will not be the most costly part in good non-gaming box, hard
>>> drives will, so again, what the matter?
>>
>> Computer usage breaks down into more than gaming and non-gaming. My
>> "non-gaming" boxes at home tend to have their CPU, RAM or NICs as
>> their most expensive components, because that's where I need them to
>> perform better.
>>
>
> CPU speed does not matter. what matters most is the I/O speed.
>
> As far as I can tell, AMD chip suffered with a lot of I/O. Their
> Hyper-transport seems not competitive with Intel's ring bus

(please don't top-post, especially if the thread's already been
primarily organized as bottom-post)

I hadn't read that, but remember that HyperTransport is intended for a
mesh architecture. In single-CPU systems, you'll only have one HT
link, the link between your CPU and your north bridge. In multi-CPU
systems, you'll have additional links between the CPUs. In systems
with many CPUs, you may even have a fully-connected mesh.

The I/O characteristics will greatly depend on the topology of your network.

That said, HyperTransport may just be getting old; when it came out,
it (and AMD's crossbar switch for memory management) beat the pants
off of Intel's SMP solution. Intel's solution ran at lower and lower
clock rates the more CPUs you added, and their first pass at multicore
gave each core its own port onto the memory bus, with predictably poor
results. Intel's had plenty of time to catch up, but with their
price-per-part, it's taken me a long time to pay much attention.

(It also doesn't help that Jon "Hannibal" Stokes stopped writing
detailed technical articles for Ars Technica; I sincerely miss him and
the precision and clarity of his writing on such arcane subjects.)

-- 
:wq