Wed 22 Aug 2012 14:07

Intel desktop CPU price-performance: Hyperthreading not helping?
Typically, CPU prices follow performance. Faster CPUs command higher prices; slower CPUs are available for less. Recent Intel desktop CPUs continue to show this general pattern, but there appears to be more to the story than usual.

At first glance, everything seems to be what you would expect. Using current pricing in US$ at time of writing from newegg.com, we get:

Processor	PassMark	Price	PassMark/$	Price-Performance vs G640
Pentium G640	2893	$79	36.6	100%
i3-2120	4222	$125	33.8	92.2%
i5-3570	7684	$215	35.7	97.6%
i7-3770	10359	$310	33.4	91.3%

The PassMark (http://www.cpubenchmark.net/) to dollar ratio is pretty consistent across all these processors, roughly 35 ± 2.

But what happens if we look at a more real-life benchmark? Consider SPEC CPU 2006 Integer (CINT2006) Baseline. For each CPU, I used the CINT2006 Baseline results from the most recently reported Intel reference system, as reported on spec.org. In the case of the G640, no Intel reference system was reported, so I used the results for a Fujitsu Primergy TX140 S1p.

Processor	CINT2006 Base	Price	CINT/$	Price-Performance vs G640
Pentium G640	34.4	$79	0.44	100%
i3-2120	36.9	$125	0.30	67.8%
i5-3570	48.5	$215	0.23	51.8%
i7-3770	50.5	$310	0.16	37.4%

When looking at CINT2006 Baseline, we see the price-performance ratio drop off dramatically as the processor price increases. We would expect this from the i3 to the i5, since SPEC cpu int is a single job benchmark and the i3 to the i5 represents a transition from two to four cores, but it's curious to see the dropoff in the price-performance ratio between the G640 and the i3 (both dual-core CPUs), and the i5 and the i7 (both quad-core CPUs). What might be going on?

A look at hyperthreading may provide some answers. Intel hyperthreading is a feature of some Intel CPUs that allow each physical core to represent itself to the OS as two different "cores". If those two "cores" simultaneously run code that happens to use different parts of the physical core, they can proceed in parallel. If not, one of the "cores" will block. The i3 and i7 CPUs offer hyperthreading, the Pentium G and i5 do not. It turns out that the PassMark benchmark sees significant speedups when hyperthreading is turned on. SPEC CINT2006, and many ordinary applications, do not.

What about SPEC CINT2006 Rate Baseline, then? The SPEC CPU Rate benchmarks measure throughput, not just single-job performance, so maybe hyperthreading helps more here? Let's see:

Processor	CINT2006 Rate Base	Price	Rate Base/$	Price-Performance vs G640
Pentium G640	61.7	$79	0.78	100%
i3-2120	78.8	$125	0.63	80.7%
i5-3570	146	$215	0.68	87.0%
i7-3770	177	$310	0.57	73.1%

If we look at the transition from two to four cores (by comparing the i3 to the i5), we now see that the price-performance of the i5 is better than the i3: this is no surprise, since we are now measuring throughput, and from the i3 to the i5, we go from two to four cores. But there still is a dropoff in price-performance between the Pentium G and the i3, and again between the i5 and the i7. It's not as extreme as before, but it is still significant. This suggests that hyperthreading may help with throughput, but not as much as the increase in price would suggest.

What does this mean, then? It suggests the increase in price from a non-hyperthreaded to a hyperthreaded Intel desktop processor may reflect more an increase in PassMark performance than an increase in real performance. Hyperthreading may have a positive effect, it seems, but typically not as much as PassMark suggests. At present, for best real-world price-performance in Intel desktop CPUs, I would consider models without hyperthreading.

/it permanent link