Rank graphics cards by their reference
A problem that frequently arises is to identify the performance of a card according to the trade name.
This article will explain how to navigate in the classification of graphics cards from ATI and nVidia, and then try to compare a theoretical point of view the performance maps of the 2 manufacturers.
How to identify the performance of an NVidia graphics card by its reference?
For nVidia, the numbering follows an order about logic:
- The thousands digit characterizes to which family the GPU belongs, also the year of release. For example, a 8600 GT, released in 2007, is older than a 9600 GT, released in 2008, and therefore less efficient;
- The hundreds digit is connected to the power of the GPU, the higher it is, the more powerful the GPU is (from there you can benchmark the GPU belonging to the same family). For example, a 8600 GT is significantly less powerful than a 8800 GT;
- Finally, the suffix completes the reference, it is related to GPU and memory frequencies, the number of GPU units GPU (or stream processors), it significantly increases the performance, from lowest to highest (LE, G, GS, GT, GTS, GTX, Ultra). But suffixes do not exist for all cards. For example, a 8800 GT is less powerful than a 8800 GTS, both are all based on the same GPU, but the GTS works with all processing units activated and with GPU/memory frequencies higher than that of the GT version .
But unfortunately, it is sometimes difficult:
NVidia released cards of different generations having the same name: Be careful not to confuse "old" 8800 GTS 320 /640 MB of 2007 with new 8800 GTS 512 MB of 2008, they are not based on the same GPU: The old 8800GTS uses a G80 GPU and the new one uses a G92 GPU. Note that, nVidia launched the 9000 series, based on the same GPU as the 8800 released in 2007-2008, using the same G92 GPU,but having higher frequencies
the 9600 GT, with half of G92 disabled is equivalent to a 8800 GT with limited performance
and the 9800 GTX has the structure of an overclocked 8800 GTS and has the performance of a 8800 GTX.
The 9500 GT is much like a 8600 GTS engraved in 65nm.
End June 2008, nVidia released the GTX 200 series, consisting of the 260 and 280 series: these references have nothing to do with the old cards ...
We can compare card of different generations, but not the ones having a completely different reference: a 7600 GT is less powerful than a 8600 GT (single digit difference), but it's performance can't be compared to a 8600 GTS and a 9500 GT.
How to identify the performance of an ATI card by its reference?
We can say that ATI designations follow roughly the same principle as nVidia, at least for older generations to the year 2007:
- The thousands digit characterizes is the family of GPUs: a 1900 is older than 2900;
- The hundreds digit is connected to the power of the GPU, the higher it is, the more powerful the GPU is: a 1950 Pro is more powerful than 1650 Pro;
- Finally a suffix complete reference. Obviously these suffixes are not the same as those from nVidia. The most common suffixes in 2007 were in ascending order of performance: GT Pro XT. For example:
A 1950 GT is less powerful than 1950 Pro
A 2600 Pro is less powerful than a 2600 XT.
For new generations, (from December 2007), the figure of thousands will point the same thing, but the numbering of the power of the GPU is different and suffixes have been dropped:
- The 4000 series is higher than the 3000 series;
- A 3850 is much more powerful than 3650: the hundreds digit indicates a different GPU different.
- A 4850 is slightly less powerful than 4870: the figure 4870 seems quite unusual (+ 20), this because it is the same GPU running at a higher frequency.
Again, we can compare card of different generations, but not the ones having a completely different reference: a 1600 Pro is less powerful than a 2600 Pro, but it's performance can't be compared to a 1950 Pro and 2600 XT.
Similarly for new generations, we can't determine which one is better between a HD 3850 and a HD 4670 (all digits are different), without looking at the structure of their GPUs.
Theoretical comparison between (same manufacturer) nVidia cards and ATI/AMD cards
To get a clearer idea of theoretical performance of cards from the same manufacturer, we must look at the structure of the GPU. Stream Processors, texturing units and ROP's (Raster Operators).
If we compare, for example, a GeForce 9500 GT with a 8600 GTS, trying to determine which one is better solely by the model name ,is impossible, but looking at the structure of the GPU, we find: The 2 eGPU has 32 stream processors, 16 texturing units ans 8 ROP units. Of course, the internal structure of these units may be slightly different from one card to another, but we can say a priori that the performance of these 2 cards are almost the same. Similarly, between the HD 3850 and HD 4670: HD 3850 (GPU name: RV670) stream processors: 64 v5, often recognized 320 (64x5), 16 texturing units and 16 raster units (ROP's). HD 4670 (GPU name: RV730) stream processors: 64 v5, 32 texturing units and 8 raster units (ROP's).
Here we see the same amount of stream processors, but a difference between texturing units and ROP's. As the stream processors are the most important, we can say that the analysis of GPU performance will be almost the same, but with greater error margin than the GeForce card mentioned above.
In practice, the test showed that the GeForce 8600 GTS and 9500 GT have very similar performance.
For HD 3850 and 4670, the results are close, but the HD 3850 has a slight advantage over the Geforce, for gaming or filtering purposes.
The method of analyzing the GPU will give good results,but of course less accurate than actual testing.
Theoretical comparison of nVidia and ATI/AMD cards
To compare the performance of nVidia cards and ATI is more complicated:
Previously whe trying to compare cards fro the same manufacturer, we compared the number of stream processors, texturing units and ROP's, but nVidia and ATI have quite a different GPU architecture including on stream processors, required to return a little more detail .
Since the first version of the 8800, nVIDIA proposes an architecture called SIMT: "Single Instruction, Multiple Threads," which corresponds in practice to the implementation of an operation on multiple threads as opposed to the SIMD architecture "Single Instruction, Multiple Data, where an of several operations is applied to an item, (technology, adopted by ATI).
The chips from AMD and NVIDIA both offer unified graphics architectures, but the SIMT architecture induces a different accounting of stream processors from an ATI stream processor, which can contain 5 instructions.
For example, the GeForce 8800 GTS will work on 128 items and the GeForce GTX 280 of 240 items, at once, as for the RV770 found in the 4850 card will work on 160 elements, each containing up to five instructions: its units calculation are of a vec5 type, and ATI has proudly announced the release of 5x 160 = 800 stream processors!
Divide by 5 the raw number of stream processors advertised for ATI GPUs to get a number that allows comparison with those given by nVidia: E.g, we must count only160 stream processors for a HD 4850, not 800.
From example above, we see that the performance of the 4850 fall between those of the 8800GTS and GTX 280: 160 stream processors for a HD 4850 card, against 128 for the GeForce 8800 GTS and 240 for the GTX280, the HD4850 is slightly more efficient than the 8800 GTS(confirmed when testing games).
In summary, analysis of the structure allows the GPU to go further in predicting the performance of cards, but with less precision than practical tests. Of course, the best solution is to read the comparative on specialized sites, the real tests will allow you to know exactly what to expect in terms of a perfomance with a particular card.
If no test is available, the analysis of GPUs will provides you a fairly accurate idea of card performance.