Spring 2007 PC Assembler's Quick Guide ...
2007 March -- work in progress (WIP) / quick, conceptual dump
There's not a day that goes by without someone e-mailing me about PC components and assembly. The overwhelming majority of these are home consumers, although some are for some high I/O throughput server designs. I've addresses many basic server concepts in both my print and blog articles, so I will offer this blog article for the home consumer crowd. I'll try to release a new blog article every six (6) month with major changes, and these articles will be -- more or less -- a "Quick Guide" and not too in-depth.
NOTE: This article is not trying to sell you on PC self-assembly or why you should do it. If you're not sure you should, don't read this article, it's for those who already self-assemble their PCs. In fact, I personally use a sub-$1,000 17" 'desktop replacement' notebook (with a dual-core CPU, GeForce Go 7600 256MB and 1680x1050 video display, 2GB of memory and 160GB disk) as my primary PC now, so my self-assembled system is really for when I need more power.
Overview:
There's not a day that goes by without someone e-mailing me about PC components and assembly. The overwhelming majority of these are home consumers, although some are for some high I/O throughput server designs. I've addresses many basic server concepts in both my print and blog articles, so I will offer this blog article for the home consumer crowd. I'll try to release a new blog article every six (6) month with major changes, and these articles will be -- more or less -- a "Quick Guide" and not too in-depth.
NOTE: This article is not trying to sell you on PC self-assembly or why you should do it. If you're not sure you should, don't read this article, it's for those who already self-assemble their PCs. In fact, I personally use a sub-$1,000 17" 'desktop replacement' notebook (with a dual-core CPU, GeForce Go 7600 256MB and 1680x1050 video display, 2GB of memory and 160GB disk) as my primary PC now, so my self-assembled system is really for when I need more power.
Overview:
- Rule #1: Seek at least 2 reviews for every product
- Everything starts with the mainboard (aka motherboard)
- GPUs (video/graphics) change twice as fast as CPUs
- Power, cooling and enclosures
- The lesser known things to know about memory
- Storage options, what not to do, and redundancy
- The extra frills, and budgetary mistakes
Side Discussion:
- Free/Fake RAID (FRAID) dos, donts and conversions
Rule #1: Seek at least 2 reviews for every product
With the exception of warranty and service (or reseller in the case of DoA, Dead on Arrival), vendor brand name means squat today. Everyone outsources, even Sony to OEM Chinese fabs, Seagate to Maxtor (now the same company), etc... That means you need to research the exact model (and sub-model/revision if need be) and I recommend you seek at least 2, in-depth reviews for each and every component, possibly combinations of components if at all possible.
This also means you don't merely want to rely on "favorable reviews" on the reseller's site. You want to see more technicial expertise at times, or at least look at the explanations in the negative reviews to find some lesser know facts. After all, you're going to find some "lesser known facts" in this article, among others out there. So look for those "lesser known" facts that -- "hey, it works!"
Everything starts with the mainboard (aka motherboard)
We're not talking servers or even workstations here, so throw those out. We're talking about desktops, systems that will run 8 hours a day, maybe 14 hours in a full day (unless you have a serious gaming addiction). So we're talking about mainboards (aka motherboards) as inexpensive as consumers want to pay for their needs.
Key mainboard attributes:
With the exception of warranty and service (or reseller in the case of DoA, Dead on Arrival), vendor brand name means squat today. Everyone outsources, even Sony to OEM Chinese fabs, Seagate to Maxtor (now the same company), etc... That means you need to research the exact model (and sub-model/revision if need be) and I recommend you seek at least 2, in-depth reviews for each and every component, possibly combinations of components if at all possible.
This also means you don't merely want to rely on "favorable reviews" on the reseller's site. You want to see more technicial expertise at times, or at least look at the explanations in the negative reviews to find some lesser know facts. After all, you're going to find some "lesser known facts" in this article, among others out there. So look for those "lesser known" facts that -- "hey, it works!"
Everything starts with the mainboard (aka motherboard)
We're not talking servers or even workstations here, so throw those out. We're talking about desktops, systems that will run 8 hours a day, maybe 14 hours in a full day (unless you have a serious gaming addiction). So we're talking about mainboards (aka motherboards) as inexpensive as consumers want to pay for their needs.
Key mainboard attributes:
- Size and expansion
- Integrated features and video
- AMD v. Intel, single or multiple core
Size and expansion
In the PC self-assembly world, ignoring Intel's OEM-centric BTX, there are two (2) common standards: ATX and MicroATX. ATX has seven (7) expansion slots spread over a 12" width, and is typically 9.6" deep (Extended ATX is 13" and typically found in only workstation or server mainboards), but can be as shallow as 8". MicroATX has four (4) expansion slots spread over is 9.6" wide and can be as deep as a square 9.6", but is commonly 8" or less. Both ATX and MicroATX share a common 6" port cutout, so they really only differ in size and expansion.
At first glance, MicroATX and ATX seem to only have a 2.4" difference in size. But in reality, the actual, total "area" as typically implemented really adds up. MicroATX cuts corners not merely in expandability, but advanced feature set because it may be a single chip chipset, resulting in a smaller area and fitting into some near small form-factor (SFF) cases (e.g., MicroATX towers, desktops and cubes can be quite small, and almost as small as a more limited expansion chassis, but with 2 extras slots and drive bays). ATX tends to sport additional expansion, including newer multi-video cards options, more ports, extra on-board options, etc...
Integrated features and video
Features are really about what you are going to do with your system. If you're like me, I like either the on-board video or a single video card (we'll get into this later on GPUs), two hard drives, one optical drive and maybe one "device bay" (see the final frills section). That's my typical desktop. I don't like something that is big, I prefer something that is portable, so I've gone MicroATX since 2004 because the feature sets in MicroATX systems now sport almost everything. I still put a MicroATX board into a chassis that can take a full ATX power supply (and not a more limited MicroATX, which will talk about when we get to power). And despite the common belief, smaller cases can actually house solid cooling, power and have other, positive considerations besides just portability.
The more features lacking, the more you may end up going to an expansion board for them if you need them. If you don't need them, then you don't need the expansion, so MicroATX becomes more of a consideration. Some key, integrated features include ...
In the PC self-assembly world, ignoring Intel's OEM-centric BTX, there are two (2) common standards: ATX and MicroATX. ATX has seven (7) expansion slots spread over a 12" width, and is typically 9.6" deep (Extended ATX is 13" and typically found in only workstation or server mainboards), but can be as shallow as 8". MicroATX has four (4) expansion slots spread over is 9.6" wide and can be as deep as a square 9.6", but is commonly 8" or less. Both ATX and MicroATX share a common 6" port cutout, so they really only differ in size and expansion.
At first glance, MicroATX and ATX seem to only have a 2.4" difference in size. But in reality, the actual, total "area" as typically implemented really adds up. MicroATX cuts corners not merely in expandability, but advanced feature set because it may be a single chip chipset, resulting in a smaller area and fitting into some near small form-factor (SFF) cases (e.g., MicroATX towers, desktops and cubes can be quite small, and almost as small as a more limited expansion chassis, but with 2 extras slots and drive bays). ATX tends to sport additional expansion, including newer multi-video cards options, more ports, extra on-board options, etc...
Integrated features and video
Features are really about what you are going to do with your system. If you're like me, I like either the on-board video or a single video card (we'll get into this later on GPUs), two hard drives, one optical drive and maybe one "device bay" (see the final frills section). That's my typical desktop. I don't like something that is big, I prefer something that is portable, so I've gone MicroATX since 2004 because the feature sets in MicroATX systems now sport almost everything. I still put a MicroATX board into a chassis that can take a full ATX power supply (and not a more limited MicroATX, which will talk about when we get to power). And despite the common belief, smaller cases can actually house solid cooling, power and have other, positive considerations besides just portability.
The more features lacking, the more you may end up going to an expansion board for them if you need them. If you don't need them, then you don't need the expansion, so MicroATX becomes more of a consideration. Some key, integrated features include ...
- Audio: 2, 4 or 6 channel, maybe even optical output
- Network: 10/100 or 10/100/1000 (although I wouldn't use the 1000Mbps in most desktop systems for a server)
- SATA and other expansion ports
- Integrated video: ATI, Intel, nVidia, SiS or ViA, PCIe slot or dual/tripple PCIe (aka "SLI" or "CrossFire")
Integrated audio varies wildly. More and more mainboards today, even MicroATX, come with a 6-channel audio solution. The question is how much CPU off-load (or increased CPU load) does the on-board DSP/codec give you. A lot of on-mainboard audio, regardless of board size, leaves much to be desired, and you may look to a PCI or, fairly new for audio, PCIe x1 audio card. I have personally put in Creative SoundBlaster Audigy 2ZS PCI cards even on expansion-limited MicroATX mainboards, because most audio DSPs/codecs put a lot of strain on the I/O-CPU for more than 2-channel output. It is very rare for a mainboard to have a quality DSP/codec, and most go for low-cost RealTek ALC6x0/8x0 solutions.
Network is in the same boat as audio, you can get mixed results on either ATX or MicroATX. Many desktop GbE (gigabit ethernet, 1000Mbps) would croak under a server load and result in little better than FE (fast ethernet, 100Mbps). Others actually put a good amount of SRAM (48-64KiB/port) and have decent performance. It all depends on the MAC+PHY combination.
SATA and other ports can be a consideration. Most mainboards are now coming with at least four (4) SATA ports and one (1) ATA channel. You'll want at least (1) ATA channel for an optical drive (see the storage section on why), more than one if you have legacy ATA hard drives. If you really need more than four (4) SATA ports, you might want to invest in a real hardware controller anyway, although I won't tell you what you need.
Video is really the main differentiator, and for home consumers (and gamers), probably the #1 consideration.
For people that want cheap -- around $50 -- the nVidia C51/C61 with integrated NV44 (GeForce 61x0) also known as the nForce 410/430 chipsets have an integrated video powerful enough to run Vista and even older DirectX 8 games (although they support Shading 2.0 for DirectX 9, their performance will not do well). In fact, these integrated Graphics Processor Unit (GPU) chipsets will beat the GeForce FX 5200/5500/5700LE, so it's cheaper to upgrade to a new $50 C51/61 mainboard (with a future PCIe upgrade option) than pay $50-100 for a GeForce FX AGP card. These chipsets are only available for AMD Socket-754/939 and newer Socket-940/AM2, although AMD's new (ATI-based) G690 chipset for Socket-AM2 is also comparable in performance.
If you want integrated graphics, Intel, SiS and ViA are lackluster, and not Vista capable. Yes, the newer Intel G955 options say they are Vista ready, have partial Shader support and look good in artificial benchmarks, but they are not up to the standard of the nVidia C51/C61, the AMD G690 or even the aged ATI Xpress 200/1100 integrated graphics.
If you want to go with the upper-end, then you need a mainboard with a PCIe graphics slot -- which is everything today. You may want to look towards the newer nVidia nForce 500 series (than the older nForce4 or GPU-integrated nForce4x0), which is available for either AMD or Intel. Go SiS or ViA at your own risk. Intel's own chipsets work well with an add-on PCIe graphics too -- the 975X is the premier professional, although the 965 is solid for consumers too. If you want to use more than video card, then you'll definitely need to look to a full ATX mainboard for either nVidia's SLI or ATI's CrossFire (MicroATX options exist, but I don't recommend them as they wipe out virtually all other expansion). And the price of those options will be over $100, sometimes close to $200.
In all honesty, I don't think much of SLI/CrossFire unless you are going to buy two (2) of the top-end graphics cards. The reason why is because two (2) lower-end graphics cards rarely equal the price-performance of one (1) higher-end graphics card, as I'll discuss later. That's the reason why I've stuck with MicroATX personally, but if you're willing to pay for the top-end mainboard and top-end pair of video cards, then you'll want ATX.
AMD v. Intel, single or multiple core
First off, forget the Intel Celeron D, it is not dual-core, and it's a slouch, period. there are far better options unless you find a really cheap mainboard+CPU deal (I'm talking under $50) with AGP and you have a very high-end AGP card (e.g., GeForce 6800) that you want to reuse. Pentium D, although dual-core, not much more desirable either, despite the over-clocking potential (for you enthusiasts out there), with the new "price war" between AMD and Intel.
In fact, the reason why this is under the "mainboard" section has everything to do with cost and performance and your choice of CPU as a result of the mainboard.
If you want "cheap," get a $50 nVidia C51/61 (GeForce 61x0/nForce 4x0) mainboard, put a $50 Sempron/Athlon 64 in it, and you'll get damn fine performance for most desktop applications and even aged (pre-DirectX 9/Shader) games, as well as Vista. Upgrade to dual-core for $50 more, which will be more responsive in multiple applications, although will help gaming very little. Choose Socket-939 if you want to reuse any existing DDR SDRAM memory, or just go Socket-AM2 with DDR2 SDRAM if you have to buy memory.
If you want to put in a solid game performer, buy a $100 nForce 570 or Intel 945 series, put in an $200 Intel Core 2 Duo -- preferably a 6000 series (the 4000 series isn't holding itself against AMD Athlon x2 at the price points) -- and at least a GeForce 7600GT or ATI X1950Pro or faster (see the forthcoming video section). Intel is matching AMD's price cuts, even if delayed, and the Intel Core 2 Duo 6000 series is clearly the "bang for the buck" on performance. Quad core gives you little, unless you are really building software and want to throw "make -j8" or something (if you don't know what that means, don't worry about it).
Your mainboard really defines your video and CPU choices. For cheap, Vista (or even Linux Compiz) capability, the nVidia C51/61 (or newer AMD G690, although for Linux it's still an "unknown" quantity) with a cheap Athlon 64 or entry Athlon x2 is a $100 combo -- especially in a smaller MicroATX form-factor. If you want to game, look to a $100+ mainboard with an nForce 500 series or Intel 965/975, add your own accelerator, and you can still often find then in MicroATX, although you'll definitely need to go ATX for SLI/CrossFire.
Linux Notes: Of all chipsets, I've found the nVidia nForce4/4x0 series to be the most Linux compatible, GPL/MIT driver ready chipset. As far as integrated video, as long as you don't need 3D acceleration, the C51/61 (NV44 GPU) is also excellent for X-Windows releases in the last 2 years. If you want 3D acceleration without binary drivers, Intel is an option, but understand the "open" drivers lag Intel's own proprietary drivers heavily in performance and features (especially the total lack of hardware accelerated OpenGL 2.0 support, shaders, etc...), and are only available on Windows anyway (unlike nVidia's which are a "shared object" for all OSes -- FreeBSD/Linux/X, MacOS X, Windows).
GPUs (video/graphics) change twice as fast as CPUs
With the mainboard+CPU question out-of-the-way, we can talk graphics. The problem with graphics is that they double in performance twice as fast as CPUs. If Moore's (Intel's) Law is that CPUs double in performance every 18 months, nVidia's Law is that GPUs (graphical processor units) double in performance every 9 months. So in 3 years, where CPUs increase in performance 4x over, GPUs will increase in performance 16x over.
Which is why your graphics is what you will upgrade the most, or just not care about the most. ;)
If you're still looking at or want to "recycle" a nVidia GeForce FX 5200/5500/5700LE or ATI Radeon 9200/9600 (pre-X) card, forget it. With nVidia's C51/61 chipset-integrated GPU in a $50 mainboard, you're already behind it. Today, you can get a nVidia GeForce 7600GS for $50 when on-sale, or even select ATI Radeon X1600 products for not too much more. And with the forthcoming nVidia G81+ products and ATI R600 series, prices will drop again.
Here's the deal ...
- Under $50 -- get out of here, you might as well use chipset-integrated (and save money)
- $50-60 -- the GeForce 7600GS on-sale, is also often passively cooled (0 noise)
- $90-100 -- the GeForce 7600GT or possibly ATI Radeon X1650Pro
- $130-150 -- the GeForce 7900GS or ATI Radeon X1950Pro (some would recommend the latter)
Things to avoid ...
- nVidia GeForce 7300LE (junk, maybe 50% better than chipset integrated)
- nVidia GeForce 7300GT (mega-underclocked 7600GS)
- ATI Radeon lower than X1600 (except maybe earlier X800/850 products)
If you're wanting to spend over $150, I would wait. nVidia is going to release its G81/82/etc... GeForce 8600/8900 and other series products shortly. These are die-shrinks of the already "big/costly" G80 GeForce 8800GTS/GTX. I don't like the performance of the 8800GTS compared to a cheaper GeForce 79x0GTO/GTX (or ATI's high-end options for that matter), and the 8800GTX is big and power-hungry at its 130nm size for $600 (ouch). In all cases, prices will come down, so the most I advocate is the 7900GS/X1950Pro at this time, and the 7600GT is a good start (and faster than an older GeForce 6800GT).
Okay then, what about SLI or CrossFire? What about them?
If you're going to be pushing 400W just for the video and dropping $500+ on video cards, you might as well just get an 8800GTX for $600 and save a little power usage. After all, you could go SLI by buying a 2nd 8800GTX at a latter date when they come down in price.
In fact, that's a repeat theme I see. Don't buy two (2) cards for SLI (or CrossFire) immediately. One strategy is to buy a top-end card now, and then another later when it drops in price because it's no longer the best. But then again, even when you do that, the top-end card tends to be just as good as your SLI configuration on its own in many cases. Hence the problem with SLI/CrossFire. It seems to be only the "best" when you literally spend $1,000 just for two (2) of the very top end video cards.
I mean, one (1) GeForce 7900GS typically bests two (2) GeForce 7600GT cards in SLI (let alone the 7600GS or laughable 7300GT) for less money. And one (1) GeForce 79x0GTO/GTX will typically best two (2) GeForce 7900GS cards in SLI. So I have trouble recommending SLI at all myself. Which means I have difficulty putting in an ATX mainboard instead of a MicroATX, hence my preference for the latter. But that's just myself.
Power, cooling and enclosures
Power is all about delivering current on specific voltage rails. When you don't, because your power supply is inadequate or inefficient, then you get voltage drop, which results in instability. Modern mainboards require ATX 2.0, which has two sets of +12V inputs on its standard 24-pin (not 20-pin) connector, as well as the additional +12V of the 4-pin "P4" connector separate from the "additional 4 pins" on the 24-pin mainboard connector. Your ATX 2.0 power supply should have at least two, independent +12V "rails" -- many, even ones that offer a 24-pin connector -- sometimes only have a single +12V rail. This will be noted with a +12V(1) and +12V(2), each with their own current, possibly as well as the total aggregate offered between them.
Note this is only looking at ATX 2.0, not the EPS12V or EEB SSI standards for servers/workstations (which have various other 6-pin workstation and 8-pin server connectors that are not ATX/PCIe or otherwise compatible power connectors for consumers).
Additionally, you need to consider the following, additional connectors ...
Network is in the same boat as audio, you can get mixed results on either ATX or MicroATX. Many desktop GbE (gigabit ethernet, 1000Mbps) would croak under a server load and result in little better than FE (fast ethernet, 100Mbps). Others actually put a good amount of SRAM (48-64KiB/port) and have decent performance. It all depends on the MAC+PHY combination.
SATA and other ports can be a consideration. Most mainboards are now coming with at least four (4) SATA ports and one (1) ATA channel. You'll want at least (1) ATA channel for an optical drive (see the storage section on why), more than one if you have legacy ATA hard drives. If you really need more than four (4) SATA ports, you might want to invest in a real hardware controller anyway, although I won't tell you what you need.
Video is really the main differentiator, and for home consumers (and gamers), probably the #1 consideration.
For people that want cheap -- around $50 -- the nVidia C51/C61 with integrated NV44 (GeForce 61x0) also known as the nForce 410/430 chipsets have an integrated video powerful enough to run Vista and even older DirectX 8 games (although they support Shading 2.0 for DirectX 9, their performance will not do well). In fact, these integrated Graphics Processor Unit (GPU) chipsets will beat the GeForce FX 5200/5500/5700LE, so it's cheaper to upgrade to a new $50 C51/61 mainboard (with a future PCIe upgrade option) than pay $50-100 for a GeForce FX AGP card. These chipsets are only available for AMD Socket-754/939 and newer Socket-940/AM2, although AMD's new (ATI-based) G690 chipset for Socket-AM2 is also comparable in performance.
If you want integrated graphics, Intel, SiS and ViA are lackluster, and not Vista capable. Yes, the newer Intel G955 options say they are Vista ready, have partial Shader support and look good in artificial benchmarks, but they are not up to the standard of the nVidia C51/C61, the AMD G690 or even the aged ATI Xpress 200/1100 integrated graphics.
If you want to go with the upper-end, then you need a mainboard with a PCIe graphics slot -- which is everything today. You may want to look towards the newer nVidia nForce 500 series (than the older nForce4 or GPU-integrated nForce4x0), which is available for either AMD or Intel. Go SiS or ViA at your own risk. Intel's own chipsets work well with an add-on PCIe graphics too -- the 975X is the premier professional, although the 965 is solid for consumers too. If you want to use more than video card, then you'll definitely need to look to a full ATX mainboard for either nVidia's SLI or ATI's CrossFire (MicroATX options exist, but I don't recommend them as they wipe out virtually all other expansion). And the price of those options will be over $100, sometimes close to $200.
In all honesty, I don't think much of SLI/CrossFire unless you are going to buy two (2) of the top-end graphics cards. The reason why is because two (2) lower-end graphics cards rarely equal the price-performance of one (1) higher-end graphics card, as I'll discuss later. That's the reason why I've stuck with MicroATX personally, but if you're willing to pay for the top-end mainboard and top-end pair of video cards, then you'll want ATX.
AMD v. Intel, single or multiple core
First off, forget the Intel Celeron D, it is not dual-core, and it's a slouch, period. there are far better options unless you find a really cheap mainboard+CPU deal (I'm talking under $50) with AGP and you have a very high-end AGP card (e.g., GeForce 6800) that you want to reuse. Pentium D, although dual-core, not much more desirable either, despite the over-clocking potential (for you enthusiasts out there), with the new "price war" between AMD and Intel.
In fact, the reason why this is under the "mainboard" section has everything to do with cost and performance and your choice of CPU as a result of the mainboard.
If you want "cheap," get a $50 nVidia C51/61 (GeForce 61x0/nForce 4x0) mainboard, put a $50 Sempron/Athlon 64 in it, and you'll get damn fine performance for most desktop applications and even aged (pre-DirectX 9/Shader) games, as well as Vista. Upgrade to dual-core for $50 more, which will be more responsive in multiple applications, although will help gaming very little. Choose Socket-939 if you want to reuse any existing DDR SDRAM memory, or just go Socket-AM2 with DDR2 SDRAM if you have to buy memory.
If you want to put in a solid game performer, buy a $100 nForce 570 or Intel 945 series, put in an $200 Intel Core 2 Duo -- preferably a 6000 series (the 4000 series isn't holding itself against AMD Athlon x2 at the price points) -- and at least a GeForce 7600GT or ATI X1950Pro or faster (see the forthcoming video section). Intel is matching AMD's price cuts, even if delayed, and the Intel Core 2 Duo 6000 series is clearly the "bang for the buck" on performance. Quad core gives you little, unless you are really building software and want to throw "make -j8" or something (if you don't know what that means, don't worry about it).
Your mainboard really defines your video and CPU choices. For cheap, Vista (or even Linux Compiz) capability, the nVidia C51/61 (or newer AMD G690, although for Linux it's still an "unknown" quantity) with a cheap Athlon 64 or entry Athlon x2 is a $100 combo -- especially in a smaller MicroATX form-factor. If you want to game, look to a $100+ mainboard with an nForce 500 series or Intel 965/975, add your own accelerator, and you can still often find then in MicroATX, although you'll definitely need to go ATX for SLI/CrossFire.
Linux Notes: Of all chipsets, I've found the nVidia nForce4/4x0 series to be the most Linux compatible, GPL/MIT driver ready chipset. As far as integrated video, as long as you don't need 3D acceleration, the C51/61 (NV44 GPU) is also excellent for X-Windows releases in the last 2 years. If you want 3D acceleration without binary drivers, Intel is an option, but understand the "open" drivers lag Intel's own proprietary drivers heavily in performance and features (especially the total lack of hardware accelerated OpenGL 2.0 support, shaders, etc...), and are only available on Windows anyway (unlike nVidia's which are a "shared object" for all OSes -- FreeBSD/Linux/X, MacOS X, Windows).
GPUs (video/graphics) change twice as fast as CPUs
With the mainboard+CPU question out-of-the-way, we can talk graphics. The problem with graphics is that they double in performance twice as fast as CPUs. If Moore's (Intel's) Law is that CPUs double in performance every 18 months, nVidia's Law is that GPUs (graphical processor units) double in performance every 9 months. So in 3 years, where CPUs increase in performance 4x over, GPUs will increase in performance 16x over.
Which is why your graphics is what you will upgrade the most, or just not care about the most. ;)
If you're still looking at or want to "recycle" a nVidia GeForce FX 5200/5500/5700LE or ATI Radeon 9200/9600 (pre-X) card, forget it. With nVidia's C51/61 chipset-integrated GPU in a $50 mainboard, you're already behind it. Today, you can get a nVidia GeForce 7600GS for $50 when on-sale, or even select ATI Radeon X1600 products for not too much more. And with the forthcoming nVidia G81+ products and ATI R600 series, prices will drop again.
Here's the deal ...
- Under $50 -- get out of here, you might as well use chipset-integrated (and save money)
- $50-60 -- the GeForce 7600GS on-sale, is also often passively cooled (0 noise)
- $90-100 -- the GeForce 7600GT or possibly ATI Radeon X1650Pro
- $130-150 -- the GeForce 7900GS or ATI Radeon X1950Pro (some would recommend the latter)
Things to avoid ...
- nVidia GeForce 7300LE (junk, maybe 50% better than chipset integrated)
- nVidia GeForce 7300GT (mega-underclocked 7600GS)
- ATI Radeon lower than X1600 (except maybe earlier X800/850 products)
If you're wanting to spend over $150, I would wait. nVidia is going to release its G81/82/etc... GeForce 8600/8900 and other series products shortly. These are die-shrinks of the already "big/costly" G80 GeForce 8800GTS/GTX. I don't like the performance of the 8800GTS compared to a cheaper GeForce 79x0GTO/GTX (or ATI's high-end options for that matter), and the 8800GTX is big and power-hungry at its 130nm size for $600 (ouch). In all cases, prices will come down, so the most I advocate is the 7900GS/X1950Pro at this time, and the 7600GT is a good start (and faster than an older GeForce 6800GT).
Okay then, what about SLI or CrossFire? What about them?
If you're going to be pushing 400W just for the video and dropping $500+ on video cards, you might as well just get an 8800GTX for $600 and save a little power usage. After all, you could go SLI by buying a 2nd 8800GTX at a latter date when they come down in price.
In fact, that's a repeat theme I see. Don't buy two (2) cards for SLI (or CrossFire) immediately. One strategy is to buy a top-end card now, and then another later when it drops in price because it's no longer the best. But then again, even when you do that, the top-end card tends to be just as good as your SLI configuration on its own in many cases. Hence the problem with SLI/CrossFire. It seems to be only the "best" when you literally spend $1,000 just for two (2) of the very top end video cards.
I mean, one (1) GeForce 7900GS typically bests two (2) GeForce 7600GT cards in SLI (let alone the 7600GS or laughable 7300GT) for less money. And one (1) GeForce 79x0GTO/GTX will typically best two (2) GeForce 7900GS cards in SLI. So I have trouble recommending SLI at all myself. Which means I have difficulty putting in an ATX mainboard instead of a MicroATX, hence my preference for the latter. But that's just myself.
Power, cooling and enclosures
Power is all about delivering current on specific voltage rails. When you don't, because your power supply is inadequate or inefficient, then you get voltage drop, which results in instability. Modern mainboards require ATX 2.0, which has two sets of +12V inputs on its standard 24-pin (not 20-pin) connector, as well as the additional +12V of the 4-pin "P4" connector separate from the "additional 4 pins" on the 24-pin mainboard connector. Your ATX 2.0 power supply should have at least two, independent +12V "rails" -- many, even ones that offer a 24-pin connector -- sometimes only have a single +12V rail. This will be noted with a +12V(1) and +12V(2), each with their own current, possibly as well as the total aggregate offered between them.
Note this is only looking at ATX 2.0, not the EPS12V or EEB SSI standards for servers/workstations (which have various other 6-pin workstation and 8-pin server connectors that are not ATX/PCIe or otherwise compatible power connectors for consumers).
Additionally, you need to consider the following, additional connectors ...
- PCI-Express (PCIe), 6-pin, dual +12V for video cards
- SATA, 15-pin, +12V, +5V and +3.3V for hard drives
The PCIe slot on the mainboard is supposed to deliver up to 150W of power, far more than the 25W of the AGP slot (and still less than even 25-50W or 50-100W of the lesser known "AGP Pro" slot). Prior to PCIe, many AGP cards used an extra 4-pin Molex connector for additional +12V input. With the advent of PCIe, and the reality that some video cards are pulling 250W+ -- which is over 20A on a +12V rail -- and it's not uncommon to pull more current on a Molex line than the wire gage is rated in some cases when its attached to other devices. The dedicated PCIe connector is designed for this, and you'll need two (2) for SLI/CrossFire.
Even more greatly misunderstood is the 15-pin SATA power connection. Although some SATA drives offer a Molex connector, the 7-pin data + 15-pin power SATA design is a fixed location, single connector attachment (SCA), with staggered pins for power transient. The 15-pin power connector contains three (3) power rails, +3.3V in addition to +5V and +12V. As such, it is best not to use Molex adapters with newer SATA drives as their logic boards may not be regulated to adapt +5V or +12V to +3.3V. Get newer Y-cables that split SATA to SATA itself, including providing the +3.3V that power supplies with SATA connectors do.
Cooling comes next, and it's often a major issue. Bigger heatsinks and more fans doesn't always mean better, just more protruding and louder. Retail CPU fansinks and stock GPU fansinks aren't necessarily bad, and they are warrantied. The bigger issue is the ambient temperature in your case. How much air you can allow in your case, flow over the warmest components and push out of your case is the main issue, as well as the location of components that may or may not be cooled.
Large fans spinning slower can move more air with less noise than smaller fans spinning louder. 120mm at 1000rpm has become the universal ideal, even in smaller MicroATX designs. Dual ball bearing fans are quieter than sleeve fans, and last longer.
The three (3) biggest issues in your case are in order of importance ...
Even more greatly misunderstood is the 15-pin SATA power connection. Although some SATA drives offer a Molex connector, the 7-pin data + 15-pin power SATA design is a fixed location, single connector attachment (SCA), with staggered pins for power transient. The 15-pin power connector contains three (3) power rails, +3.3V in addition to +5V and +12V. As such, it is best not to use Molex adapters with newer SATA drives as their logic boards may not be regulated to adapt +5V or +12V to +3.3V. Get newer Y-cables that split SATA to SATA itself, including providing the +3.3V that power supplies with SATA connectors do.
Cooling comes next, and it's often a major issue. Bigger heatsinks and more fans doesn't always mean better, just more protruding and louder. Retail CPU fansinks and stock GPU fansinks aren't necessarily bad, and they are warrantied. The bigger issue is the ambient temperature in your case. How much air you can allow in your case, flow over the warmest components and push out of your case is the main issue, as well as the location of components that may or may not be cooled.
Large fans spinning slower can move more air with less noise than smaller fans spinning louder. 120mm at 1000rpm has become the universal ideal, even in smaller MicroATX designs. Dual ball bearing fans are quieter than sleeve fans, and last longer.
The three (3) biggest issues in your case are in order of importance ...
- Hard drive
- CPU
- GPU
Despite more recent marketing, you should never let the ambient temperature around your hard drive exceed 40C (104F). It should be warm to the touch, hot is a bad sign and to singe your finger on mere touch is deadly. Airflow must occur over your hard drives. It doesn't matter if the drive is enterprise rated for "24x7," it will die rather quickly without airflow. Your case design and options for cooling the hard drive bays (unless you relocate them to external 5.25" with a mount option, taking up those slots) is of the utmost concern.
Next is the CPU, which needs to get cool air as well. Intel addressed this in the BTX mainboard design by putting the CPU at the middle-front of the mainboard, and having tower cases blow air from the middle over the CPU first. Ironically, the best cases in the ATX are those that have little more than a side "fan duct" directly over the CPU tower's fan, and then with a 120mm or at least a 92mm exhaust fan out the back. It's effective, doesn't rely on any cooler air from the front of the case (which may have been warmed by hard drives or expansion slots below it in a tower), and just works.
Lastly is the GPU. The GPU is quickly becoming the biggest power gobbler and thermal generator in the system, but the PCIe expansion slot -- by its very nature -- forces GPU packagers and thermal designers to accomodate. As such, it is typically designed to run at much higher temperatures and with less airflow in its far more constrained spaces than your hard drive or CPU. Even just a set of air vents above the expansion area lets more cooler air in. And the hottest GPUs take up two slots, using the headroom to enclose the card and push warm air directly out the back of the second slot.
Without radical enclosures like BTX or the "flipped ATX" design, there's not much we can do about the GPU in a PCIe expansion slot. Although these solutions do lower thermals by as much as 10C. The BTX flips the mainboard orientation so the expansion card's "main IC side" are "facing up" as heat rises. The "Flipped ATX" design also does the same, and can even put the CPU at the lowest part of the case for the ultimate cooling. There are even "Flipped MicroATX" towers.
Furthermore, there is a newfound interest in the "normal standing" desktop or "cube" case, especially for MicroATX. The becoming common 9" x 11" x 14-15" MicroATX "cube" (e.g., Chenming 118 series or the $39 Ultra Microfly) offers some excellent airflow designs. First, they have a 120mm outtake fan just above the CPU area, which is also directly behind the hard drive area (which may have some "side slits" for cooler intake air). Second, they have the power supply right above the card area, and its outtake fan (especially if mounted internally, like some power supplies with large 120mm fans do) can "pull up" warm air directly from the slot area. With two 120mm outtake fans (one in the power supply, another directly above the CPU area), these new MicroATX "cube" designs can effectively cool all components, even though the designs are tight and cramped for their components.
Other variations exist as well.
I have covered both the "Flipped ATX" and "MicroATX cubes" in my earlier blog articles, and the options have increased since them. Enclosures are always a personal taste, but the key is to look for options, and price compare. Even just a cheap ATX enclosure with a 120mm outtake option near the CPU area, the CPU duct/openings and a 80mm intake option around the lower hard drive area is adequate for most users. The Ultra Wizard cases have these options and can be had for free after rebate regularly (provided you add the fans, possibly some felt feet as well).
The lesser known things to know about memory
Memory is probably the most misunderstood aspect of PC assembly. To start, the synchronous timing/signaling used is hardly the performance metric. To continue, virtually all vendors promote violation of JEDEC specifications.
Socket-754 (DDR) is a single channel, direct, 184-pin trace from the CPU, so Socket-754 runs into this limitation directly. Socket-939 (DDR) offers two, 184-pin traces from the CPU, so has limits that are double. Socket-940/AM2 offers two DDR2 channels, so it can only support two (2) DIMMs, period, under JEDEC specification (for stability). Despite marketing, Socket-478 (DDR) runs into single channel issues. LGA-775 is dual channel, so it can only support two (2) DIMMs, period, for JEDEC like Socket-940/AM2.
Timing of the DRAM cells in DRAM ICs on a DRAM DIMM is the biggest performance factor, by far.
The first thing to understand about simple, Dynamic Random Access Memory (DRAM) cells, even those on Synchronous DRAM packages, is that they do not have a clock that runs synchronously with the channel. The MHz signaling does not correspond to the actual cell "latency." That is, a 667MHz (PC2-5300) DDR2 SDRAM module does NOT have 1.5ns (1/666,666,667 of a second) timing. In fact, it doesn't mean it has 3.0ns timing (since two bits are transferred per clock) either.
DRAM cells are typically between 20 to 60ns latency (only 16 to 50MHz -- yes sixteen to fifty). That means it takes several cycles when reading to even get the "first bit of data." Synchronous DRAM attempts to mitigate these great latencies by bursting transfer in pages, typically 4K on x86/x86-64 (what Intel calls IA-32/IA-32e) processors.
Writes are virtually synchronous because they are "send and forget." They will move to the SRAM (static RAM) cache of the processor and/or board, and eventually be sent out to memory. No real latency impact. But reads are devastating if they are not in the SRAM cache already, as the great latency difference between the processor/board and its SRAM and the system DRAM adds great delay. A "cache miss" in modern, superscalar microprocessors and their board interconnects causes an exponential performance hit (of typically over a magnitude).
The SPD chip on each DIMM reports to the channel its various CAS, RAD, RP and RAS timings for the synchronous clock rate. For any DIMM product, you will typically see four (4) timings like 2.5-3-3-6 DDR400/PC-3200 or 5-5-5-15 on DDR2-800/PC2-6400 SDRAM modules (which are pretty good timings). Taking in the synchronous timing, you can figure out the CAS, RAD, RP and RAS of the SDRAM module. If you're intersted more on each definition, see the Wikipedia page (I will not go into them here).
A RAS of 6 on DDR400/PC-3200 means 6 cycles at 200MHz DDR = 6 * 5ns = 30ns (33.3MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
A RAS of 15 on DDR2-800/PC2-6400 means 15 cycles at 400MHz DDR = 15 * 2.5ns = 37.5ns (26.6MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
A RAS of 15 on DDR2-667/PC2-5300 means 15 cycles at 333MHz DDR = 15 * 3ns = 45ns (22.2MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
Yes, latency is not the only factor, as increased, synchronous singling means increased Data Transfer Rate (DTR) -- especially since pages are fetched 4K at a time. But if you're reading more than writing, have a smaller CPU cache (not just L2, but also L1 -- remember, AMD processors have 4x the L1 cache of Intel processors -- long story), etc... latency can have a far greater impact than DTR.
This is why even cheap DDR2 memory can be outperformed by quality DDR, resulting in overall system performance even being 10-20% better. Having directly memory channels on the CPU (e.g., AMD Athlon 64/x2/Opteron), let alone a larger L1 cache (which is completely synchronous with the CPU), can make a huge latency difference as well. This is why AMD did not gain much performance in moving from the DDR Socket-939 platform to the DDR2 Socket-940/AM2 platform.
Over-voltage requirements of DDR2 DIMMs is becoming an increasing issue. It's beyond just many DDR2 DIMMs requiring greater than the JEDEC spec 1.8V -- such as 1.9V, 2.0V, 2.1V or even 2.2V in some cases, to reach the optimal/fastest timing. Many DDR2 DIMMs -- especially higher PC2-6400 (DDR2-800), but even some PC2-5300 (DDR2-667) DIMMs require them to even function at all. So beware when purchasing higher signaling DDR2 DIMMs these days, and verify they will work at the speeds and timings at the stock JEDEC DDR2 1.8V specification. Check the reviews on forums, or even just the negative reviews on reseller sites for feedback with those key discoveries. Over-voltage significantly reduces product life, assuming you even have a mainboard that can offer over-voltage for your DIMMs in the first place.
It's rare, but older DDR DIMMs can require over-voltage as well. Virtually all are stock 2.8V, although newer DIMMs may only require lower voltages of 2.6V (or even 2.1V in a few cases). Putting 2.6V DIMMs in a mainboard that only supports 2.8V is typically not an issue, as over-volting often improves stability (although it can shorten the life of the product), unlike the issues with DDR2 (and its commonly required over-volting). Again, it's far less common with DDR than DDR2, but there are still select products out there.
Storage options, what not to do, and redundancy
The extra frills, and budgetary mistakes
Next is the CPU, which needs to get cool air as well. Intel addressed this in the BTX mainboard design by putting the CPU at the middle-front of the mainboard, and having tower cases blow air from the middle over the CPU first. Ironically, the best cases in the ATX are those that have little more than a side "fan duct" directly over the CPU tower's fan, and then with a 120mm or at least a 92mm exhaust fan out the back. It's effective, doesn't rely on any cooler air from the front of the case (which may have been warmed by hard drives or expansion slots below it in a tower), and just works.
Lastly is the GPU. The GPU is quickly becoming the biggest power gobbler and thermal generator in the system, but the PCIe expansion slot -- by its very nature -- forces GPU packagers and thermal designers to accomodate. As such, it is typically designed to run at much higher temperatures and with less airflow in its far more constrained spaces than your hard drive or CPU. Even just a set of air vents above the expansion area lets more cooler air in. And the hottest GPUs take up two slots, using the headroom to enclose the card and push warm air directly out the back of the second slot.
Without radical enclosures like BTX or the "flipped ATX" design, there's not much we can do about the GPU in a PCIe expansion slot. Although these solutions do lower thermals by as much as 10C. The BTX flips the mainboard orientation so the expansion card's "main IC side" are "facing up" as heat rises. The "Flipped ATX" design also does the same, and can even put the CPU at the lowest part of the case for the ultimate cooling. There are even "Flipped MicroATX" towers.
Furthermore, there is a newfound interest in the "normal standing" desktop or "cube" case, especially for MicroATX. The becoming common 9" x 11" x 14-15" MicroATX "cube" (e.g., Chenming 118 series or the $39 Ultra Microfly) offers some excellent airflow designs. First, they have a 120mm outtake fan just above the CPU area, which is also directly behind the hard drive area (which may have some "side slits" for cooler intake air). Second, they have the power supply right above the card area, and its outtake fan (especially if mounted internally, like some power supplies with large 120mm fans do) can "pull up" warm air directly from the slot area. With two 120mm outtake fans (one in the power supply, another directly above the CPU area), these new MicroATX "cube" designs can effectively cool all components, even though the designs are tight and cramped for their components.
Other variations exist as well.
I have covered both the "Flipped ATX" and "MicroATX cubes" in my earlier blog articles, and the options have increased since them. Enclosures are always a personal taste, but the key is to look for options, and price compare. Even just a cheap ATX enclosure with a 120mm outtake option near the CPU area, the CPU duct/openings and a 80mm intake option around the lower hard drive area is adequate for most users. The Ultra Wizard cases have these options and can be had for free after rebate regularly (provided you add the fans, possibly some felt feet as well).
The lesser known things to know about memory
Memory is probably the most misunderstood aspect of PC assembly. To start, the synchronous timing/signaling used is hardly the performance metric. To continue, virtually all vendors promote violation of JEDEC specifications.
- JEDEC banking limits for PC (DDR) and PC2 (DDR2)
- Timing is everything
- Over-voltage requirements and JEDEC spec violations
DDR JEDEC Spec DIMMs DDR2 JEDEC Spec DIMMsSeeing a repeat theme here? Only 1 DIMM per channel for any DDR2 speed as well as PC-3200 DDR400. On any JEDEC compliant DDR memory controller, using more than 1 PC-3200 (DDR-400) DIMM per channel will result in the signaling slowing down to PC-2700 (DDR-333), which people regularly complain about. These are doubled for "Registered/Buffered" DIMMs, but desktop systems rarely offer them (let alone the newer, "Fully Buffered (FB)" DDR2 DIMMs in newer servers/desktops).
---- ---------- ----- ---- ---------- -----
200 PC-1600 3 400 PC2-3200 1
266 PC-2100 2 533 PC2-4200 1
333 PC-2700 2 667 PC2-5300 1
400 PC-3200 1 800 PC2-6400 1
Socket-754 (DDR) is a single channel, direct, 184-pin trace from the CPU, so Socket-754 runs into this limitation directly. Socket-939 (DDR) offers two, 184-pin traces from the CPU, so has limits that are double. Socket-940/AM2 offers two DDR2 channels, so it can only support two (2) DIMMs, period, under JEDEC specification (for stability). Despite marketing, Socket-478 (DDR) runs into single channel issues. LGA-775 is dual channel, so it can only support two (2) DIMMs, period, for JEDEC like Socket-940/AM2.
Timing of the DRAM cells in DRAM ICs on a DRAM DIMM is the biggest performance factor, by far.
The first thing to understand about simple, Dynamic Random Access Memory (DRAM) cells, even those on Synchronous DRAM packages, is that they do not have a clock that runs synchronously with the channel. The MHz signaling does not correspond to the actual cell "latency." That is, a 667MHz (PC2-5300) DDR2 SDRAM module does NOT have 1.5ns (1/666,666,667 of a second) timing. In fact, it doesn't mean it has 3.0ns timing (since two bits are transferred per clock) either.
DRAM cells are typically between 20 to 60ns latency (only 16 to 50MHz -- yes sixteen to fifty). That means it takes several cycles when reading to even get the "first bit of data." Synchronous DRAM attempts to mitigate these great latencies by bursting transfer in pages, typically 4K on x86/x86-64 (what Intel calls IA-32/IA-32e) processors.
Writes are virtually synchronous because they are "send and forget." They will move to the SRAM (static RAM) cache of the processor and/or board, and eventually be sent out to memory. No real latency impact. But reads are devastating if they are not in the SRAM cache already, as the great latency difference between the processor/board and its SRAM and the system DRAM adds great delay. A "cache miss" in modern, superscalar microprocessors and their board interconnects causes an exponential performance hit (of typically over a magnitude).
The SPD chip on each DIMM reports to the channel its various CAS, RAD, RP and RAS timings for the synchronous clock rate. For any DIMM product, you will typically see four (4) timings like 2.5-3-3-6 DDR400/PC-3200 or 5-5-5-15 on DDR2-800/PC2-6400 SDRAM modules (which are pretty good timings). Taking in the synchronous timing, you can figure out the CAS, RAD, RP and RAS of the SDRAM module. If you're intersted more on each definition, see the Wikipedia page (I will not go into them here).
A RAS of 6 on DDR400/PC-3200 means 6 cycles at 200MHz DDR = 6 * 5ns = 30ns (33.3MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
A RAS of 15 on DDR2-800/PC2-6400 means 15 cycles at 400MHz DDR = 15 * 2.5ns = 37.5ns (26.6MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
A RAS of 15 on DDR2-667/PC2-5300 means 15 cycles at 333MHz DDR = 15 * 3ns = 45ns (22.2MHz equivalent) maximum delay to access a select row of data in a DRAM IC.
Yes, latency is not the only factor, as increased, synchronous singling means increased Data Transfer Rate (DTR) -- especially since pages are fetched 4K at a time. But if you're reading more than writing, have a smaller CPU cache (not just L2, but also L1 -- remember, AMD processors have 4x the L1 cache of Intel processors -- long story), etc... latency can have a far greater impact than DTR.
This is why even cheap DDR2 memory can be outperformed by quality DDR, resulting in overall system performance even being 10-20% better. Having directly memory channels on the CPU (e.g., AMD Athlon 64/x2/Opteron), let alone a larger L1 cache (which is completely synchronous with the CPU), can make a huge latency difference as well. This is why AMD did not gain much performance in moving from the DDR Socket-939 platform to the DDR2 Socket-940/AM2 platform.
Over-voltage requirements of DDR2 DIMMs is becoming an increasing issue. It's beyond just many DDR2 DIMMs requiring greater than the JEDEC spec 1.8V -- such as 1.9V, 2.0V, 2.1V or even 2.2V in some cases, to reach the optimal/fastest timing. Many DDR2 DIMMs -- especially higher PC2-6400 (DDR2-800), but even some PC2-5300 (DDR2-667) DIMMs require them to even function at all. So beware when purchasing higher signaling DDR2 DIMMs these days, and verify they will work at the speeds and timings at the stock JEDEC DDR2 1.8V specification. Check the reviews on forums, or even just the negative reviews on reseller sites for feedback with those key discoveries. Over-voltage significantly reduces product life, assuming you even have a mainboard that can offer over-voltage for your DIMMs in the first place.
It's rare, but older DDR DIMMs can require over-voltage as well. Virtually all are stock 2.8V, although newer DIMMs may only require lower voltages of 2.6V (or even 2.1V in a few cases). Putting 2.6V DIMMs in a mainboard that only supports 2.8V is typically not an issue, as over-volting often improves stability (although it can shorten the life of the product), unlike the issues with DDR2 (and its commonly required over-volting). Again, it's far less common with DDR than DDR2, but there are still select products out there.
Storage options, what not to do, and redundancy
The extra frills, and budgetary mistakes
Free/Fake RAID (FRAID) dos, donts and conversions
