12-Way Radeon Gallium3D GPU Comparison

[Kivada]

I think the issue is that there's no standard way for PTS to obtain this info.

Yet this is not a problem the rest of the hardware sites on the internet have, they put up all the info about the card in all of their graphs.

Also, don't make crap up about there being no way to garner this information, since GPU-Z gets all of this information and more about every type of GPU, there is no reason PTS cannot get this same information.

[Kivada]

A better than normal article, but I still would suggest you include a bit more information about the specific cards being tested. For instance, the Radeon HD 4670 you used in your test is not the same as the Radeon HD 4670 I am using in my machine, and I doubt the Radeon HD 6450 in the article is the one I recommended to my brothers. Maybe include a mention of the brand used, such as Sapphire, Diamond, MSI, etc...

The card manufacturer brand doesn't make any difference on performance, it only determines what the warranty support will be like or after market cooling if they make a model that uses something other then the AMD reference design cooler.

What does make a world of diference on low end cards is the memory type and it's interconnect bit width. On the low to lower mid range cards you will find all combinations of GDDR, GDDR2, GDDR3, GDDR5 64-bit, 128-bit, etc. For a given GPU core, a card with 64-Bit GDDR5 is going to perform about the same as a card with 128-bit GDDR3 as they will have about the same total VRAM bandwidth.

Now though you will also see the same GPU core being paired up with an absolutely pitiful 64-bit GDDR which no matter what will always be much slower because all the life is being choked out of it by the lack of VRAM bandwidth.

Rule of thumb on low end dedicated GPUs, never buy anything with less then 128-bit GDDR5. Memory size doesn't matter pretty much at all since these cards wont have the grunt to handle most heavier games that would need allot of vram anyways, but you still want the card to be able to fully stretch it's legs instead of tying an anvil to it's neck..

[Sidicas]

I think the issue is that there's no standard way for PTS to obtain this info.

Yes, it would be nice to have a universal standard way to obtain GPU memory configuration info from *ANY* graphics card.

You can actually compare AMD vs. nVidia cards on their memory performance based on memory bitwidth and memory clock which is something you can't do with the GPUs themselves without benchmarking because they have different designs.. nVidia has always historically made very high bitwidth cards because their product releases often happen at awkward times, which is whenever the Taiwan fab is able to manufacture the GPUs with a decent yield... AMD has historically delayed their graphics card releases to correspond with new video memory technology so that they don't need to make their cards high bitwidth, which then means more performance for your money and it's easier to get a lot of board manufs. making AMD graphics cards since the board manuf.'s costs are lower.

[Hamish Wilson]

I'd say it's logical to put memory configurations (VRAM capacity+VRAM Clock+VRAM Bitwidth) of the cards in the labeling.. I don't think brand is necessary as it doesn't really have any relation to performance. So long as you have *ALL* of the big 3 measurements of video card memory (capacity, clock, bitwidth) you have something meaningful.. Without one of them, it really doesn't say anything.. You could have a lot of capacity without any throughput or vice-versa, so if it's going to be done, I hope it's done right..

That is fair enough; the main reason I have brought up brand is I that I had found it to be the easiest way to distinguish between my Radeon HD 4670 and the others that were available in the past. A more thorough system would of course lend well to better results.

[bridgman]

I think the issue is that there's no standard way for PTS to obtain this info.

[Sidicas]

I was more referring to differences in VRAM and memory configuration, not overclocking. Case in point:

http://ncix.ca/products/?sku=73221&vpn=R6450MD2GD3%2FLP&manufacture=MSI%2FMicroStar&promoid=1016

http://ncix.ca/products/?sku=72269&v...cture=Gigabyte

I'd say it's logical to put memory configurations (VRAM capacity+VRAM Clock+VRAM Bitwidth) of the cards in the labeling.. I don't think brand is necessary as it doesn't really have any relation to performance. So long as you have *ALL* of the big 3 measurements of video card memory (capacity, clock, bitwidth) you have something meaningful.. Without one of them, it really doesn't say anything.. You could have a lot of capacity without any throughput or vice-versa, so if it's going to be done, I hope it's done right..

Too many benchmarking websites (80%+) don't understand how video memory works and leave out at least one of the three and make benchmarks that are impossible to draw any real conclusions from.

IMO, it's better to not do it at all, unless it can be done right.. And to do it right, you need all 3. Only with all 3 can you really draw conclusions about whether a card is video memory bottlenecked by memory capacity or memory performance or not. Not sure if phoronix wants to get into all that video memory stuff, but if they do, I hope they do it right, that's all.

It is common for some card manufs. to compensate for low memory clocks by adding bitwidth to their cards and it always works.. It's also common for card manufs. to lower the bitwidth (reduce board manuf. costs) but then use newer memory (higher memory clock) to obtain the same performance at a lower bitwidth (cost).. So reporting just memory clock or memory bitwidth alone is meaningless, but some benchmarking websites do exactly that!

[Hamish Wilson]

Brand really doesn't have anything to do with performance.. Sure, some brands sell factory overclocked cards that *MAY* have been binned to be more expensive because they overclock 5% better, but I really don't think these cards should be used in benchmarking since they end up being a tiny minority of cards from that model sold on the market.

I was more referring to differences in VRAM and memory configuration, not overclocking. Case in point:

http://ncix.ca/products/?sku=73221&vpn=R6450MD2GD3%2FLP&manufacture=MSI%2FMicroStar&promoid=1016

http://ncix.ca/products/?sku=72269&vpn=GV-R645-1GI&manufacture=Gigabyte

[agd5f]

Going through OpenBenchmarking.org you can find the PCI IDs and other information. Unfortunately AFAIK there is no embedded information readable from any of the AIB vendor cards for automatically being able to determine the brand from the software side.

You should be able to get the AIB vendor from the pci subsystem ids in most cases.

[bridgman]

Memory configuration (DDR2 vs DDR3 vs GDDR5) usually makes the biggest difference between cards using the same GPU SKU. I've seen almost a 2:1 performance difference between slowest and fastest memory configurations.

[Michael]

A better than normal article, but I still would suggest you include a bit more information about the specific cards being tested. For instance, the Radeon HD 4670 you used in your test is not the same as the Radeon HD 4670 I am using in my machine, and I doubt the Radeon HD 6450 in the article is the one I recommended to my brothers. Maybe include a mention of the brand used, such as Sapphire, Diamond, MSI, etc...

Going through OpenBenchmarking.org you can find the PCI IDs and other information. Unfortunately AFAIK there is no embedded information readable from any of the AIB vendor cards for automatically being able to determine the brand from the software side.