Singapore views...: How to select a Graphic Card for desktop?

The need for good graphics…

To non-gamers, Graphic Card (aka VGA card or Video Card) is just purely a link between the computer and the monitor. For serious gamers, it is a matter of life and death because 1 ms delay in refresh may pronounce you KIA. 10 years ago, we have many vendors producing graphic cards and we are spoilt for choices. Now, only 3 major vendors (Intel, AMD/ATI and Nvidia) compete in this arena. Intel, the first vendor, produces many low-end integrated graphic chips which are mainly fixed on-board offering no option to upgrade. Its arch rival AMD acquired ATI which also produces on-board graphics as well as standalone discrete graphic cards. These discrete are in directly competition with Nvidia. If we disregard Intel low-end chips, AMD/ATI and Nvidia nearly capture 100% of the entire worldwide graphic card market. They shall be the primary focuses of our discussion.
If you acquire a budget motherboard, it is likely that an on-board graphic chip is already present. It provides the basic function to operate basic applications, play simple games and watch movies. Intel's Graphic Media Accelerator (GMA) and AMD/ATI's Integrated Graphic Processor (IGP) dominate the market. There are many complaints of poor rendering and inability to play graphic intensive game until AMD/ATI launches 780G chipset (in Apr 2008) which takes the playing field to the next level. Intel is playing catch up while AMD/ATI announced the arrival of 890 chipset in Jan 2010. This is beyond our discussion but we will reference these chips in our discussion.

Reviews, Benchmark and Charts

With so many reviews written on different graphic cards complete with benchmark charts and statistics, choosing a discrete graphic card should be a piece of cake. However, the 2 camps (Nvidia and AMD/ATI) adopt different architectures and specifications in their design and it is not an apple to apple comparison. To confuse us further, they introduce different technologies into their cards and boost about the value-added on the experiences of viewing 3D etc. Nvidia has CUDA, SLI, Optimus, PhysX, PowerMizer and PureVideo. AMD/ATI offers Eyefinity, Avivo, Catalyst, HyperMemory, PowerPlay and CrossFireX. Each claims their technologies are superior and in my opinion, many review sites fail to link what customers really want. With the latest graphic card, they focus entirely on benchmarks, applications and games based on a fixed budget. They fail to provide a complete picture of what are available in the market (especially outside USA). They keep promoting the latest graphic card thinking that it is always superior in design and price against obsolete product lines. Perhaps they are sick of reviewing cards with minor improvement over the old ones or they are paid to promote and give good comments to entice customers to purchase them. None of them clearly tells you what you should look for in the graphic card. I hope to fill in this gap.

Budget and Motherboard slots

The industry defines low-end graphic cards with a price tag of below US$100 (S$145), mid end graphic cards between US$101 to US$300 (S$146 – S$435) and premium cards for more than US$301 (S$435). Since you can invest in 2 graphic cards concurrently with SLI and CrossFireX technologies, it is prudent to know if you are getting 1 card or 2 cards else you will burst your pocket. You also need to know if your motherboard can support multi-cards, check the number of PCIe slots you have and if the PCIe belongs to 1.0 or 2.0 standards.
My advice is simple. If your motherboard has no on-board graphic and you are a non-gamer, choose a low-end graphic card (HD5450 is an excellent card with TDP of only 19.1W). If you are a casual gamer, go for mid-end. If you are a serious gamer, be prepared to burn your pocket or look for a reasonable good mid-end card.

Technical Specifications

In the good old days, we only check on the processor speed to determine if it is a fast processor. For graphic cards, it is not a straight forward story because both camps have different design architecture resulting in AMD having more processors than Nvidia. This does not necessarily mean that AMD is faster than Nvidia. Therefore, I shall list down the considerations in order of priority.

Rendering Output Units (ROP) – MOST IMPORTANT CONSIDERATION
ROP determines the price point. A high-end graphic card has 32 ROP while a low-end graphic card has 4 ROP. Usually, a high ROP commands better price. ROP takes pixel and texel information and processes them in the local memory. If you have a small ROP, everything will be done slowly especially in 3D.
Texture Mapping Units (TMU) and Graphic Processing Unit (GPU)
As the names imply, TMU and GPU handle texture and graphic separately before passing them to ROP. The more TMU and GPU you have, the faster is the process. However, AMD generally has more GPUs than Nvidia so you can compare them only if they are from the same vendors.
Fabrication Process and Heat Dissipation (TDP)
An advanced fabrication production will squeeze more transistors into the same chip resulting in smaller size and less heat dissipation. Graphic chips are producing using 40nm to 65nm technology. Smaller fabrication process produces energy efficient chip to reduce heat dissipation. As you may not be aware, for instance HD5970 chokes up maximum of 294W which require you to have an external power plug to power the chip as PCIe 2.0 can provide only 150W of power. Your power supply may also need to be 700W to ensure normal operation.
Active or passive Heatsink
An active heatsink refers to a fan rotating non-stop during operation. A passive heatsink cools itself via non-mechanical means such as liquid cooler or having an excellent metal conductor mounted to the chip. The latter offers a quieter operation environment but cost more than the formal. You need to examine it before buying.
Memory bus width/type
High-end graphic has 256-bit while low-end cards supports 32-bit data transfer. A higher transfer speed means faster delivery so look for a higher bit memory bus. For memory type, GDDR is preferred over DDR and GDDR5 is faster than GDDR3.
The rest of the specifications
To your sweet surprise, memory size, memory clock rate, GPU core speed, bus width, GPU memory sizes are not in my priority list as touted by many websites. Yes, they are indeed secondary concerns. They are only meaningful if you know how AMD/ATI and Nvidia name their card models which we will discuss in the later paragraphs.

After understanding what to look for in a graphic card, it is nice to know how the vendors name their cards in the next section.

AMD/ATI Cards

For Radeon and Evergreen series, AMD adopts a simple way to name their models – HDxxxx where x are digits. The first digit represents the generation of the cards. The second digit represents the range of the card (3 for low-end, 6 for mid-end or 8 for high-end). The last two digits tell us the capability of that range. Usually, a higher number means better performance.
For example, let's take HD 4870 as an example. It is the 4^th generation of AMD card and it is a high-end version and it performs better than a HD 4850 in the range segment. So we can deduce that HD 5670 is a newer card as compared to HD 4870.

Nvidia Cards

Nvidia has pre-GeForce series, GeForce series to Geforce 9xxx series and GeForce 100 to 400 series. At the moment, only GeForce 9xxx and GeForce 200 and beyond are available in the market.
Unlike AMD, Nvidia employs suffix for additional information according to this order: GS < GT < GTS < GTX. Therefore, we know that GTS is more powerful than GT etc. In contrast, we seldom see suffix in AMD offerings.

A Case Study

If we are looking for a mid-end card with a budget for S$150-200, let's see what is available in Singapore.

Vendor Model	Nvidia GeForce			AMD Radeon
Part number	9800GTX	GT 240	GTS 250	HD 4770 512M	HD 5670	HD 5750 1G
ROP	16	8	16	16	8	16
GPUs	128	96	128	640	400	720
Texture Units	64	32	64	32	20	36
Fabrication (nm)	65	40	65	40	40	40
TDP (W)	140	69	145	80	61	86
Memory Bus width	256-bit/GDDR3	128-bit/GDDR3	256-bit/GDDR3	128-bit/GDDR3	128-bit/GDDR5	128-bit/GDDR5
Retail Price (S$)	185	149	179	149	149	209
Comments	Replaced by GTS250	Eliminated due to 8 ROP			Eliminated due to 8 ROP

We have 3 Nvidia chips and 3 AMD chips. 9800GTX and HD4770 are older chips as compared to the newer series. The first thing we see is that GT240 and HD5670 have 8 ROP which is undesirable. Given that they are priced equally to HD4770, I would rather choose HD4770 over the 2 newer chips which offer lower heat dissipation. You can also see that HD 4770 has more GPUs over HD 5670 so HD 4770 is better.
Next, we see that both 9800GTX and GTS250 have similar specifications but GTS250 is much cheaper. So we shall dismiss 9800GTX.
Comparing HD4770 and HD5750, the latter is superior in specifications with more GPUs, Texture Units with a better and double GDDR. It is priced S$60 more than the formal.
If you are price-conscious, grab HD 4770 512MB. If you are going for performance, both HD5750 and GTS250 are even in specifications so GTS250 will be my choice since it is cheaper but consumes more power. Unless you want support for 3 displays (Eyefinity) and DirectX 11, HD5750 will be your choice.

Summary

I hope the above helps you to choose wisely instead of relying on reviews of individual cards or charts. In office environment, you do not need a discrete graphic card if you buy a motherboard with decent on-board graphic (like AMD 790).
Have you notice that the power consumption of a video card in full load usually exceeds the CPU. The computer is taxed for additional ~100W which may not be necessary in running word processing, emailing and surfing the web. In the long term, you are paying more on electricity bills on top of the defrayment of the discrete card. The bad news is video cards are running at high temperature so they will fail easily after their standard warranty period of 12-24 months, especially if they are cooled down by active fans. This is another reason why I hate to buy overpriced OEM brand desktop which comes with a discrete card.
Nonetheless, the fun is the process and not on the result. J