Logo   Banner   TopRight
TopUnder
Transparent
Case in Point: The Best CPU Under $300
Transparent
Date: Oct 08, 2009
Section:Processors
Author: Loyd Case
Transparent
Case in Point: The Best CPU for Under $300

I’ve recently built up two midrange Core i7 based systems, which I discuss on my blog at Improbable Insights. One is based on Bloomfield, more specifically, the popular (among performance enthusiasts, anyway) Core i7 920. The other is the new Lynnfield-based Core i7 860. What’s interesting is the similarity in pricing between the two systems. I’ll talk about the individual system builds, but I also want to explore why you might build one type of system over the other.

These systems are similar, but don’t have identical components (motherboards obviously differ.) But they are good case studies into what to think about when building a system that has some legs.

Both CPUs are priced identically on Intel’s price list -- $284 – but differ in other respects. Both have advantages and disadvantages. Let’s dig into the details of each system, then do a little postmortem analysis.

Core i7 920 and Core i7 860 Features
The Core i7 920 is the entry level of the Bloomfield CPUs. It has become something of a darling among the overclocking crowd, particularly since the “D” stepping came out a few months back. The 920 requires an LGA 1366 socket and has a triple channel DDR3 memory controller. The core i7 860 uses the newer LGA 1156 socket, and has a dual channel DDR3 memory controller. Here’s a table of their base specs:

 

Feature

Core i7 920

Core i7 860

Clock Frequency (Default)

2.66GHz

2.80GHz

                Number of QPI / DMI Paths

2

1

                Turbo Boost “bins” (1 core active)

2

5

                PCI Express Controller

Chipset

On Die

                DDR Memory Channels

3

2

                Shared L3 Cache Size

8MB

8MB

                Price (Intel Official Wholesale)

$284

$284

               Price (Average of top 3 sites)

$280

$290

 
Of course, you have to factor in platform costs as well. Comparing similar motherboards and memory averages out to around a $40-60 cost disparity, with the Core i7 920 systems on the more expensive side. X58 motherboards generally cost more, for similar feature sets, and you need triple channel memory kits, rather than dual channel. On the other hand, triple channel, 6GB DDR3 kits are cheaper on a per-gigabyte basis, than similar dual channel kits.
 

Transparent
System Configurations

I built two systems, one using an Asus P7P55D Pro P55 board and the other built around an eVGA X58 SLI Micro X58 board. The Asus board is typical of most Asus products – it’s solid, reasonably priced (you can find it for around $170), and it's moderately overclockable. Mostly, it’s just stable and works. The eVGA board looks quite cool, but has a somewhat sparse, old-school BIOS. It’s overclockable, to be sure, but you have to work at it. The graphics card in both test systems was an eVGA GeForce GTX 260 Core 216 SSC.


Asus P7P55D Pro Motherboard

Interestingly, the Core i7 860 costs slightly more if you’re actually shelling out real dollars, as opposed to Intel price list dollars. This might be simply the effect of a new product hitting the channel, or possibly a glut of Core i7 920s on the street. The difference of $10 is pretty minor though.

On the performance side, there are several parameters that can affect performance: base clock frequency, turbo boost availability, QPI channels and the number of memory channels. The last factor – the number of memory channels – can affect the overall configuration as well. With the 920, you either get 3GB or 6GB with affordable memory kits (I don’t consider $1,700 for Kingston’s 12GB kit as “affordable.) Even 12GB is relatively affordable in motherboards supporting six DDR3 sockets. With the Core i7 860, you can load up on 2GB, 4GB or drop in four modules at 8GB.


eVGA X58 SLI Micro Motherboard

QPI or DMI bandwidth can be an issue, but only in some applications. Perhaps more important from a pure performance play is the number of Turbo Boost “bins” available. Suffice it to say that a single execution core (three cores idle) on the 860 will run at a considerably higher Turbo Boost frequency than a single execution core operating in Turbo mode on the 920.

Transparent
Future Growth & Performance Considerations

Secondary Considerations: Future Growth
When I build a system, I like to believe that I can eventually drop in a new, faster CPU as the manufacturer releases updated, higher clocked versions of that processor. The bifurcation of Intel’s desktop CPUs into two different sockets has really muddied the waters. Given what Intel has publicly revealed of its CPU roadmap, here’s the growth path for Bloomfield and Lynnfield:

  • Bloomfield (LGA 1366): Will accept 32nm Gulftown 6-core CPUs. There may be clock speed bumps to the existing Core i7 900s. If so, those haven’t yet been revealed.
  • Lynnfield (LGA 1156): Will accept 32nm Clarkdale. The problem here is that Clarkdale will be dual core, four threads. Even at higher clock speeds, this seems like a step down. And P55 boards won’t be able to take advantage of Clarkdale integrated graphics (not that integrated graphics are of much interest for HotHardware-reading desktop PC users.)
  • There is currently no four core successor to Lynnfield on Intel’s public roadmap.

Now, Intel’s next generation architecture – the true successor to Nehalem – is code named Sandy Bridge. What’s known is that some variant of Sandy Bridge will offer four cores and be LGA 1156. Whether they’ll work in existing motherboards is an open question. Intel’s past suggests otherwise. When Core 2 shipped, relatively few LGA 775 motherboards capable of running Pentium 4 or Pentium D CPUs could run Core 2 CPUs, even though the socket was the same.

Performance Considerations
So it’s tough to figure out what to get, if you’re concerned about future growth. So we’ll fall back on performance: which CPU actually performs better? Core i7 860 has a 133MHz clock speed advantage, but only supports two memory channels. On the other hand, i7 860’s Turbo Boost capability is more sophisticated than the 900 series, with five bins for single thread apps, versus the 920s default two bins. Similarly, two execution threads can get up to four bins of boost, while Bloomfield only manages two.

That means that if you’re running single or dual threaded tasks on a system that’s otherwise relatively idle, the single core under stress on Lynnfield will spool up to a higher clock speed than Bloomfield.

Also, overclocking is less relevant. What overclocking will get you is better performance on three and four threaded tasks – and that’s still relatively uncommon in most day-to-day use. If you use a good CPU cooler, and have good airflow through the case, you’re already guaranteed by Intel to get 4 – 5 bumps in clock speed during dual and single threaded execution. No voltage tweaks, no experimenting with BCLK, none of that stuff needed.

One other factor that comes into play is memory bandwidth. By default, Lynnfield’s DDR3 controller handles 1333MHz DDR3. Bloomfield’s DDR3 controller is rated at 1066MHz. That makes up for the triple channel bandwidth advantage somewhat, though theoretically, Bloomfield has an edge in raw bandwidth.

Note that the actual disparity in clock frequency is only a little over 5% -- 5.26% to be more précis. In the past, a clock speed jump of about 5% might show some minor performance bumps, but nothing to write home about. But what’s the practical effect of Lynnfield versus Bloomfield? Let’s look at a few benchmarks.

Transparent
PCMark 2005

PCMark 2005
I’m using this older synthetic test because it actually breaks out some tests into single, dual and four thread tests within the CPU benchmark. Let’s look at the overall CPU score, then a single-threaded, dual-threaded and four-threaded benchmark.

In theory, even a heavily CPU oriented benchmark won’t see perfect scaling unless it all fits in the L2 cache. These tests fit nicely in the 8MB L3 cache of the Nehalem architecture, but are probably a big large for the 256KB per core L2 cache. This is a pretty old benchmark, so take the results with a lump of salt.


Hmm. There’s only a 4% difference in the overall CPU score.That’s what you’d expect, maybe a little worse. But if we look at individualtests, we see a nearly 6% difference in the File Decompression single threadedtest – better than you’d expect from the clock rate difference. Curiously,there’s almost no difference in the dual-threaded test, yet an expected 4%difference in the four-thread test.

Remember, this is a very old benchmark by today's standards. Cache hierarchies and sizes have changed, but I used this test because the more current PCMark Vantage doesn't break down tasks this way.

Transparent
CPU Intensive Task: 3D Rendering

CPU Intensive Task: 3D Rendering
Next up, we’ll look at two 3D application benchmarks. The freeware renderer POV-RAY (2.70, beta 32), which also allows you to run the test with all CPUs and threads, or just as a single thread. Cinebench 10 breaks out single versus multithreaded scores.

POVRAY sees a 5.2% gain in the All CPU benchmark, but 6% in the single threaded test. That’s slightly better than perfect scaling (if clock speed were the only criteria.) From our past experience, POVRAY sees no real benefit from higher memory bandwidth.

Cinebench, on the other hand, has a real time component as well as a software rendering piece. What’s interesting is that the single CPU score is lower than expected, while the multi-CPU test is higher than expected. 

Transparent
Games

Games
Let’s look at a pair of games. Far Cry 2 is known to be multithreaded, whereas STALKER: Clear Sky is a single threaded game. We use the Far Cry 2 Action benchmark scene, which is a little more CPU intensive than the flythrough scenes. We also dial down the level of detail considerably on one test. Both Clear Sky tests are at fairly low detail levels, but the slower result has DX9 full dynamic lighting and other settings at default, while the faster result has everything dialed way down. We’re just trying to isolate the CPU here.

With gaming, things get really interesting. Performance in the multithreaded Far Cry 2 is almost a wash. In fact, Lynnfields runs over 5-1/2% slower in the very low res, DX9 test. It’s a dead heat in the igher resolution, DX10 test.

On the other hand, the single-threaded Clear Sky game sees a massive performance increase with Lynnfield, which we can chalk up partly to memory bandwidth improvements, but probably mostly to Turbo Boost.
 

Transparent
Bottom Line

Bottom Line
If you aren’t really concerned with having very large memory capacities, or upgrading to the six core Gulftown CPU next year, then the Core i7 860 is a much better deal than the Core i7 920. It costs less, and in most cases, we see expected performance gains. In a few cases, we see some big gains in single threaded performance. When you factor in the overall system cost – Socket 1156 motherboards, on average, cost less – you can build a more cost effective system with Lynnfield.

The big unknown is the long term upgrade factor. Will Sandy Bridge just plug into current LGA 1156 boards? That’s unknown currently. But if you’re building a system today, then Lynnfield systems are cheaper than Core i7 900 series systems, and as fast or faster. What’s not to like?
 



Content Property of HotHardware.com