Intel Quad-Core Price and Performance Showdown

ThinSkin writes "The folks over at ExtremeTech have had enough time on their hands to benchmark Intel's entire quad-core lineup to determine which has the best performance for the dollar. While prices range from $183 to $1399, the real bargain is with Intel's latest Core i7 architecture which outpaced many other more expensive processors. For comparison's sake, Intel's fastest dual-core CPU was thrown into the mix and was, at times, not even competitive, which suggests that we're beginning to see more and more multi-threaded applications take advantage of four cores."

Print version

[n1ckml007]

http://www.extremetech.com/print_article2/0,1217,a%253D235027,00.asp [extremetech.com]

Print Link

[HaloZero]

Full print article should anyone not want to deal with the multipage click-through: http://www.extremetech.com/print_article2/0,1217,a%253D235027,00.asp [extremetech.com]

the computer is not just the cpu

[phr1]

The Nehalem/i7 uses DDR3 which is a lot more expensive per GB compared to DDR2 and not available in as high capacity. It has more bandwidth but its latency (which matters more) is about the same. The usual desktop mobo is limited to 2 dram modules per channel. DDR2 boards usually have 2 channels (4 sockets max) while DDR3 boards have 3 channels (6 sockets). But 4GB DDR2 modules are around $100 (link [newegg.com]) while DDR3 currently maxes out at 2GB. So you can populate a Phenom or Core 2 mobo with 16gb of ram for $400 but you can't put that much on a normal consumer i7 board for any amount of money. 2GB DDR2 parts are a lot cheaper still, you can put on 8gb (4x 2gb) at around $15/gb, $120 total. Right now a 2gb DDR3 part is $50-ish, 3x as expensive (link [newegg.com]). It helps that you can put 6 of them on a board (12gb total, $300) but you have to take the cost difference per GB into account with 2GB parts, and comparing with 4GB DDR2 parts there is $/GB parity but lower total capacity (4x4gb vs 6x2gb). And of course when 4gb ddr3 does come out, it will bring a welcome increase to 24gb total capacity, but it will be WAY expensive for quite a while (the 4gb ddr2 modules that are $100 now were $500+ for most of this year).

I just don't understand why there aren't more consumer boards with a lot more sockets, using FB-DIMM or registered DDR. You have to go to server boards for that ($$$).

Re:Whoopee

[Z00L00K]

Summary in short is that the Core i7 series is the way to go unless you just run office apps in which case the dual-core processors are sufficient.

The Q-series seems to be expensive and slow compared to the Core i7. And unless they can make a considerable price reduction on them it's no idea to select a Q-series processor.

For Canadian prices,

[Anonymous Coward]

For Canadian prices, add $40-140 (wtf, man)

Comment removed

[account_deleted]

Comment removed based on user account deletion

Re:Print version

[thePowerOfGrayskull]

Eh? No it's not... it seems that there's some referrer check going on, redirecting back to the multi-page article.

Re:the computer is not just the cpu

[rudeboy1]

Actually, I had to look that up recently. It's not 3GB, it's 4GB. Here comes the science:

"Microsoft Windows XP Professional, designed as a 32-bit OS, supports an address range of up to 4 GB for virtual memory addresses and up to 4 GB for physical memory addresses. Because the physical memory addresses are sub-divided to manage both the computer's PCI memory address range (also known as MMIO) and RAM, the amount of available RAM is always less than 4 GB.

The PCI memory addresses starting down from 4 GB are used for things like the BIOS, IO cards, networking, PCI hubs, bus bridges, PCI-Express, and video/graphics cards. The BIOS takes up about 512 KB starting from the very top address. Then each of the other items mentioned are allocated address ranges below the BIOS range. The largest block of addresses is allocated for today's high performance graphics cards which need addresses for at least the amount of memory on the graphics card. The net result is that a high performance x86-based computer may allocate 512 MB to more than 1 GB for the PCI memory address range before any RAM (physical user memory) addresses are allocated.

RAM starts from address 0. The BIOS allocates RAM from 0 up to the bottom of the PCI memory addresses mentioned above, typically limiting available RAM to between 3 GB and 3.4 GB."

I actually learned something last week, thought I'd pass it on...
*Cue the "The More You Know" logo*

Re:Flawed analysis, I just bought a dual-core E840

[Bujang Lapok]

You do NOT want to use 4GB ram for an i7. It has to be in multiples of 3GB since the i7 uses a triple channel ddr3. Some mobos have 4 slots (eg Intel DX58SO), however populating that last slot will sacrifice triple channel with single channel performance.

Re:Whoopee

[Hatta]

While we're bitching about the format, why the hell are they connecting the points on the line graph? Their X axis is meaningless, the order of those chips is arbitrary, so the slope of the line connecting the points is absolutely meaningless.

This is what bar charts are for.

Re:You don't need an article.

[pjbgravely]

I don't normally reply to AC's and this one maybe trolling but PC2-6400 (DDR2-800 SDRAM) runs at 200Mhz. Also running your FSB at 400Mhz is taking you back to P-IV levels.

Re:RAM Question

[lysergic.acid]

I understand the fun of fiddling around with a system and experimenting with something new.

no you don't, as demonstrated by the rest of your post.

not all processors/memory are made equal. some handle being overclocked better than others (certain models of high performance memory are designed by the manufacturer specifically to be overclocked). CPUs rated at different clock speeds within the same family are usually manufactured the same way, and often from the same wafer. manufacturers separate processors into different bins based on the standards each unit conforms to as determined through testing.

however, this bin assignment process is far from a perfect science. not every processor within the same bin will perform the same, and the actual performance differences between processors assigned to different bins are not always as significant as the vendor tries to project. in fact, processors may sometimes be relegated to a lower bin while conforming to the performance standards of a higher speed bin just because the manufacturer needs more units to sell at the lower price point.

well-informed overclockers understand these intricacies, and so do hardware manufacturers and system builders. that's why many modern motherboards have overclocking features/functionality integrated into their BIOS. by selecting an appropriate processor (taking into account the CPU family, stepping number, and manufacturing week) and method of overclocking (FSB/clock multiplier), there's no reason why a moderately intelligent user couldn't have a stable overclocked system. all it takes is some careful research and diligent testing.

it's not about pushing anything to "the edge." it's about pushing your system within safe limits to extract the optimal performance out of your hardware and saving money. obviously, if you're completely clueless about computers you shouldn't try to overclock anything. luckily, most overclockers know what they're doing and are able to improve the performance of their system without risking damage to their hardware.

I saved a lot. CPUs have to survive OEM builds.

[Behrooz]

There's absolutely no reason NOT to overclock Core2s. The small investment required to avoid cut-rate bare-bones parts gives you faster clock at stock/lower voltage with lower temperatures and much higher reliability.

For goodness sake, you're using a CPU that has to perform reliably in machines cobbled together by *OEM* suppliers who have to slash their costs to the bone and use the cheapest possible components in order to not go immediately out of business. There's enough headroom in the designed ratings to massacre rated performance levels. Have you ever looked inside a Dell?

Thus, my design philosophy is twofold: Excellent performance at a reasonable price with high reliability. For these reasons, I'm already going to spend more than a bare-bones system, so overclocking potential is a freebie!

In my view, the most critical components for reliability in a system are, in order:
Motherboard: Proven, tested, highly-rated enthusiast-quality mobo from a top-notch manufacturer. Abit IP35 Pro, +$80 over barebones.
Power Supply: Well-reviewed supply from a top-notch manufacturer, with at least 100w headroom over expected use. Antec TruePower Trio650, $40 over barebones.
RAM: Quality manufacturer, rated for where you expect to overclock to. - 4GB OCZ Platinum DDR2, say +$20, RAM is cheap.
Case: Antec 900. No reason to get anything else from a performance standpoint, they're on sale regularly for sub-$100.
Oh, and $15ish for a no-frills CPU fan that doesn't suck.

So, $200-ish over a barebones system for massively higher reliability. What did that get me, a YEAR ago?

Core2 E6750 OC'd 27% to 3.4GHz, 100% stable at stock voltage, running much cooler than the stock cooler did at 2.66GHz.

I'll take that any day.

Re:the computer is not just the cpu

[swillden]

32-bit OS can only support 4gb of RAM total, meaning installed RAM + video card RAM maxes out at 4gb. That's why you're not getting the full 4gb of motherboard RAM.

Wrong, twice.

First, video card RAM is not mapped into the kernel's address space and is irrelevant.

Second, 32-bit OSes can use Intel's Physical Address Extension (PAE) [wikipedia.org] mode to support up to 64 GiB of RAM. Linux does this. 32-bit Linux actually has three different memory management modes:

• HIGHMEM off: This is the traditional "split address" mode, which assigns part of the 4 GiB address space to the kernel (for kernel use and for mapping devices) and part of it to user programs. Linux uses a 3/1 split by default, with 3 GiB available to user processes and 1 GiB available to the kernel.
• HIGHMEM on: This sticks with 32-bit pointers but avoids permanently mapping chunks of the address space. The kernel gets preferential use of the lower 896 MiB of address space, but can also allocate from ZONE_HIGHMEM as needed, and user space memory can also be allocated from anywhere. Bottom line: with HIGHMEM on, the Linux kernel will use all of 4 GiB, but no more.
• HIGHMEM64G on: This enables PAE mode, which allows the kernel to use page tables to map different parts of up to 64 GiB total RAM into per-process 32-bit (4 GiB) address spaces. No single process can use more than 4 GiB in this mode, but each process can use up to 4 GiB, using all of the available RAM. This comes at a small performance hit in terms of CPU cycles.

Most desktop Linux distros today (inluding Ubuntu) ship kernels with HIGHMEM on, but not HIGHMEM64G, avoiding the PAE performance cost, but enabling the use of up to 4 GiB of RAM -- assuming the hardware supports it, which my laptop does not.

Re:RAM Question

[Detritus]

Like I tell people at work, if it doesn't have to produce the correct results, we can make it run as fast as you wish. Just because your system seems stable, doesn't mean that some obscure part of the chip isn't failing in a subtle manner. Intel has insanely expensive test jigs to ensure that their parts meet published specs at their marked speed. You have what?

For games, who cares. For real work, it's absolutely unacceptable.

Re:Whoopee

[Mike1024]

While we're bitching about the format, why the hell are they connecting the points on the line graph?

Or, given that we're comparing price and performance, a scatter plot.

I decided to replot some of the graphs properly. Here are the results [uwcs.co.uk].