[coreboot] Why does the Opteron 6328/6220 have two NUMA nodes?

[Nico Huber]

On 03.02.2018 20:38, Taiidan at gmx.com wrote:
> On 02/03/2018 01:59 PM, Nico Huber wrote:
> 
>> Hi Taiidan,
>>
>> On 03.02.2018 18:20, Taiidan at gmx.com wrote:
>>> Despite that probable memory slowdown I have noticed a significant 
>>> gaming performance increase with my higher clocked 8 core vs my 
>>> slower clocked 16 core but I am curious as to if it would be worth it 
>>> to use a C32 board for gaming instead.
>>
>> I would test that indeed. Games are often designed/optimized for very
>> simple dual/quad core cases (number increasing, though). There also
>> exist six-core C32 K10 processors, and, IIRC, they had a dedicated FPU
>> per core. 
> Oooh I would to know more :0
> Where did you hear about that?

Um, it's nothing special. I just assume K10 has an FPU per core, because
the shared FPU model seemed rather special (on x86) when they introduced
it with Bulldozer. So, a reference that Bulldozer cores share the FPU is
easy to find [1][2]. It's harder to find a reference that K10 doesn't
have that.

Though, the FPUs in later processors are much more advanced of course.
Hard to predict if that plus the higher clock, or dedicated FPUs win the
race. However, with 8 threads, I'd expect a newer CPU with 8 cores / 4
FPUs to be faster than an older one with 6 cores / 6 FPUs.

Nico

[1] https://en.wikipedia.org/wiki/Bulldozer_(microarchitecture)
[2] 
https://en.wikipedia.org/wiki/File:AMD_Bulldozer_block_diagram_(CPU_core_bloack).PNG