[GTALUG] NUC NUC NUC
Lennart Sorensen
lsorense at csclub.uwaterloo.ca
Tue May 21 11:52:19 EDT 2019
On Fri, May 17, 2019 at 07:57:32PM -0400, D. Hugh Redelmeier via talk wrote:
> For the CPU's Lennart was talking about, the modern CPU is a little
> faster, but no enough to notice
>
> <http://hwbench.com/cpus/intel-core-i7-8550u-vs-intel-core-i7-3820qm>
>
> (The i7-8550u is in the NUC; the i7-3820qm is my guess at what
> Lennart's ThinkPad W530 has.)
You guess correctly. The desktop machine runs an i7-3960X.
It certainly would not be a big leap in performance, and if you touch
anything with graphics, the quadro K2000M will beat the intel graphics
easily. :)
> Desktops are a bit different. But I would not expect even a factor of
> two difference in comparable processors of these two generations.
>
> Which processor models were/are in your (Giles') two systems?
>
> - DDR4 (8th gen) is faster than DDR3 (required by 3rd gen). But most
> programs don't seem to be affected much by this. Does your program
> bust the (rather large) L2 cache?
As far as I know, DDR4 has more bandwidth but also higher latency,
so different work loads are affected differently.
> - 3rd gen i5's often have half the cores of 8th gen i5's. This may
> make a big difference. But Python programs often don't exploit
> multiple cores.
Well 3rd gen is usually 4 core (occationally 2) and 8th gen is 6 core.
> - a few instructions have been added, but I don't imagine that they
> affect your python program. AVX2 for floating point, for example.
>
> My best guess is that the program sped up due to (NVMe?) SSD vs HDD.
Yeah that could easily drop 20m to 5m or more.
> Perhap AMD flaws haven't been discovered yet. But the current score
> shows AMD ahead.
>
> My take on rowhammer is that it is a bug in the DRAM chips and should
> be fixed under warranty. They just don't meet specs. Exploiting
> rowhammer (as opposed to just making a machine go wrong) requires
> knowing the memory mapping, and that can be facilitated by some
> processor bugs.
--
Len Sorensen
More information about the talk
mailing list