Articles, GPU Battle, Recommendations

GT 650M GDDR5 vs GT 650M DDR3 (vs GTX 660M) revised

don’t expect any pictures, but all the info is here.

The most important part is the comparison of the desktop parts GT 640 and GTX 650 between themselves and their mobile counterparts – GT 650M DDR3 and GT 650M GDDR5. The conclusion is that the GT 650M GDDR5 version deliver as high as 15-20% higher performance compared to the GT 650M DDR3, at least.

 

What is it all about?

 better investigation of the GT 650M GDDR5 and GT 650M DDR3 performance

Few month ago I posted these words about the difference between GDDR5 and DDR3 GT 650M cards. My conclusion was that the GDDR5 cards add to the performance and the impact is probably more obvious in cases where there is a need for higher bandwidth, a.k.a. 900p and especially 1080p and up – better be with AA filters. I want to add now some information, as what I’ve written seems to be insufficient.

Why? Well, the comparison between the two mobile GPUs was mainly based on benchmarks from notebookcheck DB and these benchmarks might be a bit shaky. Sometimes there is a big variation over the restults which cannot be explain by the mere difference in the HW used. Another thing is it seems that in many benchmarks there is some cpu or gpu throttling in motion, probably due the combination of high temperatures and some drivers or bios settings. Moreover, sometimes different driver versions are used – you can see, for example, with the G55VW having very good results compared to other GTX 660M equipped laptops. Barring the possibility of unprofessional benchmarking methods from notebookcheck, the difference is probably caused by throttling or drivers issues.

Saying that, it doesn’t mean the numbers posted are incorrect, but I do think there is a need to add some explaination and some more proofs. First of all, the numbers I used are only in cases there is only a small variation over the results over several machines with GT 650M GDDR5, like in the case of the Mass Effect 3 (47, 48, 50), Skyrim (24, 24, 24, 25), D3 (58, 58, 59), BF3 (15, 15, 16), CoD: MW3 (40, 42, 44, 45) and others. However, it does not enough as the GT 650M DDR3 results aren’t consistent.

Moving next to the more interesting point of this post – we will now investigate the difference in other ways. First, we’ll talk about the expected difference by comparing the GT 650M to the GTX 660M benchmarks to eliminate the difference in core clocks and try to see where does it lead us. Secondly, we’ll talk about the desktop parts GT 640 and GTX 650 which are actually the same chips as the GT 650M and GTX 660M with different clocks. Last -  I hope to make some direct benchmarks comparisons of the GT 650M GDDR5 against GT 650M DDR3 myself, but that will take time.

 

Comparing the GT 650M DDR3 and GTX 660M

“the GK107 core benefits from the GDDR5 speed but also from the higher core clocks, case dependent”

 

As we don’t have canonical benchmarks, we can’t actually rely on the benchmarks, but we will try to compare the GTX 660M and GT 650M DDR3 in order to eliminate the difference in core clocks and so compare the GK107 core with different memory only. Lets see what we get.

Judging by the technical data itself, the 660M and 650M DDR3 are the same GK107 chips, clocked the same with Turbo Boost of 950MHZ. That’s true that the chips are different from one another in their qualities (voltage, OC capabilities and so on), but as far as core model and its stock clocks goes – they are the same, expect the memory in use and its clocks.

Both cores clocked 835MHZ and the difference is that the GTX 660M memory is a 1000-1250MHZ GDDR5 while the GT 650M DDR3 one is … DDR3 clocked 900MHZ. The 660M effective core clock is 4000-5000MHZ compared to the 1800MHZ of the 650M DDR3. That’s huge gap and it’s the only one and it’s reasonable to conclude that a performance difference between these two GPUs is due the memory bandwidth to a very very high degree.

Now, based on the notebookcheck db, no matter GT 650M GDDR5 or DDR3 – the 660M has a good advantage over the GT 650Ms in most games for high settings – an average of 15% can be calculated. That means that the GK107 core benefits from the GDDR5 speed but also from the higher core clocks, case dependent. It remains to be seen in which cases the GDDR5 is more important, though. Not much to say about that now, because as I said – the benchmarks are a somewhat problematic data to rely on.

OK, next to the more interesting part (for now).

 

The desktop parts Nvidia GT 640 and GTX 650 – wow.

” the advantage of the GT 650M GDDR5 over the GT 650M DDR3 can get as high as 15-20%”

That is the more interesting part, if you ask me.

Table I:

Modelcore (stock)memory (stock)
GT 650M DDR3GK107, core@835MHZ – 950MHZ (Turbo boost)DDR3@900MHZ
(desktop) GT 640GK107, core@900MHZDDR3@900MHZ (1800MHZ effective)
(desktop) GTX 650GK107, core@1050MHZGDDR5@1250MHZ (5000MHZ effective)
GT 650M GDDR5GK107, core@735MHZ – 835MHZ (Turbo boost)GDDR5@1000MHZ (5000MHZ effective)

 

The desktop part GT 640 is basically the same chip as the GT 650M DDR3 version, only it has slightly higher base core clocks, but the GT 650M DDR3 make use of the Nvidia “Turbo Boost” feature and in 3D intensive situation, its core clock gets as high as 950% (+ ~15%) which is higher than the desktop GT 640.

The GTX 650 is also the same basic chip with the same basic specifications – same GK107 core, and same 128-bit memory bandwidth. However, it’s equipped with a fast GDDR5 memory clocked 1250MHZ (5000MHZ effective), like the GT 650M GDDR5 model.

Lets see a simple comparison between the dekstop GK107 parts and the mobile GT 650Ms.

Table II:

Modeldifference compared to the desktop counterpart
GT 650M DDR3+50MHZ for core (5-6%), same memory speed/bandwidth
GT 650M GDDR580% core and memory clocks

 

The GTX 650 has 16.67% higher core clock and almost three times the memory bandwidth. Based on several thorough reviews from techpowerup (1, 2), tomshardware and Anandtech we can say that the GTX 650 at stock clocks has a huge performance advantage of something like 50%-60% averagly at least over many games. The numbers of the GT 640 and GTX 650 seem strange? Well, they match other benchmarks too including notebookcheck GTX 660M benchmarks (lower, but matching more or less), pcper or the Anandtech GT 640 benchmarks that use other methods, but the differences between the GT 640 and GTS 450 are the similar – so we know the results are consistent to a high degree.

This difference cannot be explained by the difference in core clocks alone nor the by the driver difference, because the results seem to be consistent more or less. Even if we taking into account the 16-17% higher core clocks, there is still a big advantage of around 30% in gaming performance that cannot be explained by the difference in core clocks and this is probably due to the much higher memory bandwidth.

Now, if we do the math, the conclusion is that the advantage of the GT 650M GDDR5 over the GT 650M DDR3 can get as high as 15-20%. This is, ofcourse, also depends on the system in use, drivers and settings, but the potential is there.

Also, according to the overclocking benchmarks from Techpowerup, OC’ing the GT 640 memory by 25% results in something like 25% performance gain in BF3 (1080p, AAx4, ultra settings). Although these are results from one website and one game only, it’s still another proof that the GT 650M gpu on stock clocks craves for larger memory bandwidth. However, we cannot compare overclocking of the DDR3 to the GDDR5 impact as they are not the same.

Moreover, if the results are a good representation of the GT 650M DDR3 case – it tells us something about the difference between the GT 650M GDDR5 and DDR3. The highest possible gain from the the core clock advantage (13%-14%) is smaller than the gain in the above memory overclocking test (25%) which means that the GT 650M GDDR5 should have a good lead in performance.

Note that the GTX 650 overclocking results in Techpowerup of around +30% memory results in only 10% of performance gain in BF3. That would be interesting to check when the memory bandwidch impact starts to be less important.

 

So, GT 650M DDR3 or GT 650M GDDR5 or GTX 660M?

All counted, the benchmark results and the other indicatives (above) tell us there is a good advantage for the GT 650M GDDR5 over the DDR3 model and I believe that the comparison to the desktop part is viable – there can be an advantage of 15%-20% in favor of the GDDR5 model, in 3D and memory intensive gaming. It may vary from scenario to scenario and from game to game, but overall, you better be accompanied by the GDDR5 model. But, also, it shouldn’t be holly – there probably won’t be a difference in less memory intensive cases, like in 768p resolutions in which case the 650M will be able to deliver a hell of performance non the less gddr5 or not. So, for low 768p resolutions you probably can go quietly with the ddr3 model, even in very 3D intensive games, like BF3.

Next, an issue we should not dismiss is the overclocking. It is now known the GT 650M GDDR5 chips have high overclocking capabilities. So, if you are going to OC your GT 650M, then there is no reason for the GTX 660M almost at all and the GT 650M DDR3 won’t help you. See OC’ing instructions and results from NBR folk - you practically can make your 650M GDDR5 to a super 660M with core overclocked to as high as ~1200MHZ and memory to ~1400MHZ. That’s crazy. The overclocking show considerable performance gains as far as synthetic benchmarks go. One more interesting thing is that some people report that temperatures are almost the same in some cases (!).

OK, and what if I don’t going to overclock my gpu?  - I’d choose the GDDR5 version as the benchmarks are still in favor of the 650M GDDR5 in more memory intensive scenarios (1080p, for example) – as I said before, expect good 15% performance advantage in more intensive scenarios.

But the more context-aware answer is that you don’t have much choice with 650M DDR3/GDDR5 or 660M laptops – most of the laptops equipped with DDR3 version of 650M are in the same price range of 650M GDDR5 laptops and some of the 660M ones. It’s only good to know that you don’t lose much performance (and sometimes not at all) if you really want some specific GT 650M DDR3 model.