Re: Advice on system purchase
On 11/2/2012 9:56 PM, Charles Kroeger wrote:
> On Wed, 31 Oct 2012 12:30:02 +0100
> Stan Hoeppner <email@example.com> wrote:
>> Now if they'd just smarten up
> I've pondered this sort of thing my whole adult life. I don't understand everything
> you're saying here but it sounds pretty straight forward for someone who does, like
> the 50 miles-to-the-gallon carburettor only that was just a myth, your description
> sounds actually plausible. I guess adding cores without adding anything else would
> be a way to get higher prices for the new and better, makes sense to me,
> that's pure Harvard Business School. We've come to the truth of it.
This was all written about at great length some 10 years ago when all
the CPU vendors started implementing Simultaneous Multi-threading (SMT),
called HyperThreading by Intel marketing wizards, and/or multiple CPUs
(cores) per silicon die. The die is what most people call a "CPU".
The reason the industry went the multi-core/SMT route was due to
transistor physics, and had nothing to do with consumer workloads or
demand. For any given process technology, whether 45/32/22 nanometer,
there is a natural frequency speed limit, be it 3GHz, 3.8GHz, 5GHz, etc.
Thus you can only make a single core go so fast by increasing clock speed.
The downside to this method is power draw, which Intel famously ran into
with its Netburst single core Pentiums and Xeons that ate over 100 watts
at 3.8GHz. The rate of power draw, and thus heat dissipation, increases
on a non-linear ever steepening curve as you increase frequency in a
given process technology. For example, a 3.1GHz dual core chip produced
on a 45nm process may have a TDP of 65 watts. If this die on this
process could be pushed to 4GHz, the TDP would jump to somewhere around
100 watts, a 53% increase in power, while only increasing frequency by 18%.
If we take the same die and add two more cores at 3.1GHz, our TDP is a
little over 100 watts, about the same as the super clocked dual core.
But, in this case we've increased effective clock rate by 100%, vs 18%,
because we've doubled the number of cores. From a transistor standpoint
this is a huge win for power vs instruction throughput. And for
multi-process or multi-threaded server workloads this is great.
However, for desktop end user workloads those extra two cores sit idle
almost all the time, so there's no benefit. This is why I said it's a
no-brainer for AMD to produce a dual core chip on 32nm and clock it as
high as they can while staying around 100 watts. That should be
somewhere around 4.6-5GHz. The fastest 45nm Regor die is currently
3.4GHz. That gives us a frequency increase of 35-47% for the same power
as a 4 core version of the chip. But in this case desktop applications
can actually make good use of this extra performance, whereas they don't
make use of extra cores.
Intel could be doing this already if they'd build an i3 22nm die sans
the GPU. IBM was shipping 5GHz Power6 CPUs built on the 65nm process
technology in 2008--4 years ago. 22nm transistors are capable of well
over 7GHz frequency. An out of order branch predicting x86 dual core
CPU based on Regor or Ivy Bridge would likely top out somewhere over
5GHz. AMD/Intel's latest quad core desktop offerings seem to top out