17 entry and mid level graphic cards (page 3: GPU, cards, AGP)

Specifications

5 GPUs and 1 core currently make up the PCI entry level market: the ATI RV370, RV380 and RV410; NVIDIA NV44 and NV43 and Intel GMA 9x0.

RV410 : Radeon X700

The current middle range of the Canadian manufacturer supports Shader 2.0b like the X800 and features 6 vertex shader pipelines and 8 complete pixel pipelines (shader + ROP). This 110 nm GPU includes 120 million transistors on a 150 mm² surface.

RV380 : Radeon X600

A previous middle range product, this is simply the RV350/360 PCI Express version (Radeon 9600 Pro/XT). It supports Shader 2.0 and features 2 vertex shader pipelines and 4 complete pixel pipelines. This 130 nm GPU has 75 million transistors on a 98 mm² surface.

RV370 : Radeon X300

Strictly identical to the RV380, this 110 nm GPU is cheaper to produce as it only has a 83 mm² surface. The downside is that frequency increases aren’t as easy.

NV43 : GeForce 6600, 6200

This GPU is used by NIVIDIA for a wide range of products from entry level to middle range. It supports Shader 3.0, HDR and features 3 vertex shader pipelines, 8 pixel shader pipelines and 4 ROPs. The low amount of ROPs generally isn’t restrictive, because in practice memory bandwidth prohibits all ROPs (which mainly write data in memory) from working simultaneously. This 110 nm GPU has 146 million transistors on 160 mm².

Only the 6600 & 6600 GT have a “full option” NV43. Indeed, the 6600 LE only has 4 functional pixel shader pipelines. It’s also the case for the GeForce 6200, but it doesn’t support data compression (Z and color) and HDR.

NV44 : GeForce 6200 TurboCache

The last NVIDIA entry level GPU, the NV44 supports the Shader 3.0 but not HDR. It also doesn’t support the data compression system (Z and colors) which significantly improves other GPU performances once the FSAA is activated. It features 3 vertex shader pipelines, 4 pixel shader pipelines and 2 ROPs. The 4 pixel shader pipelines are not grouped together, but are separated in two groups of two. NVIIDA then has the possibility of deactivating 2 for some products. The GPU surface is 105 mm².

GMA 900/950 : Intel i915G/i945G

This integrated core supports the DirectX9 and Shader 2.0. It includes 4 pixel shader pipelines and 4 ROPs but not any vertex shader pipelines. All geometric calculations must be made by the GPU. It doesn’t support FSAA. It is hard to know exactly how many transistors this core requires, probably 30 million.

One important detail is that the GMA 9x0 features 2 MAD units per pipeline just like the G70 or the GeForce 7800. This of course is the only thing they have in common! The GMA 9x0 pipelines are relatively short and don’t give the possibility of efficiently masking texture access latency. These accesses are shifted to mask their latency. This is very efficient except for complex textures accesses (which are more and more common) which result in a serious performance reduction.

The GMA 900, integrated in the i915G is clocked at 333 MHz as compared to 400 MHz for the i945G GMA 950. Intel integrated cores don’t have any local memory bandwidth and have to share the entire system memory bandwidth with the rest of the system.

Graphic cards

We selected a wide range of entry level graphic cards and we added the current ATI and NVIDIA middle line for comparison.

* The TurboCache and HyperMemory cards can access local and system memory in parallel. Effective bandwidth can be slightly superior to this figure. The theoretical maximum (never reached in practice) is the addition of the total PCI Express bandwidth.

Non respected official specifications

When a new NVIDIA or ATI graphic card hits the market, it features quite strict specifications for the GPU, memory frequency and the memory bus size. Unfortunately, all manufacturers do not exactly respect these specifications and we usually find products with castrated performances. (Yes, it is possible to castrate entry level products!). For example, we found a GeForce 6200 built by Point of View with a 64 bits memory clocked at 166 MHz instead of 128 bits at 200 MHz. The graphic card’s price was even higher than other graphic cards with the correct specifications. Of course, performances are clearly lower than what they should be.

TurboCache and HyperMemory cards also have problems. Some TC 32 MB cards use only one 32 MB chip instead of 2x16 MB chips. As one chip has a 16 bit data bus, the bandwidth is divided by 2! The X300 SE HM cards are supposed to use memory clocked at 300 MHz instead of 200 MHz for a standard X300 SE. Some manufacturers, with the change to HyperMemory, have apparently decided to change the memory frequency, which is still at 200 MHz for some of the X300 SE HM. For both cases, performances are significantly reduced. Of course, these differences are never clearly indicated on the product box or characteristic page…

What about AGP ?

You have probably noticed that we only included PCI Express solutions. The ageing AGP 8x isn’t dead yet and several interesting products for the low budget and previously tested are still on the market.

For NVIDIA, the 6600 GT and 6600 exist with this interface thanks to HSI. Be careful, GeForce 6200 graphic cards based on a NV44 (the NV44A), which natively support AGP also exists. This GPU has a 64 bit memory bus and, consequently, much lower performances.

For ATI we will also remind you of the 9600 Pro, the AGP equivalent to the X600 Pro but also and mainly the 9800 Pro. It is as expensive as the 6600 and provides higher performances. Unlike the 6600, however, it doesn’t support Shader Model 3.0, HDR, or WMV9 decompression.