AT article on Haswell memory scaling

[CakeMonster]

I can't see that there's been a thread dedicated to this article, so I'm starting one. Forgive me if I missed something.

http://www.anandtech.com/show/7364/memory-scaling-on-haswell

What interests me here are the results for CPU and discrete graphics (non-CF). In the conclusion this is stated:

For discrete GPU users, recommending any kit over another is a tough call. In light of daily workloads, a good DDR3-1866 C9 MHz kit will hit the curve on the right spot to remain cost effective. Users with a few extra dollars in their back pocket might look towards 2133 C9/2400 C10, which moves a little up the curve and has the potential should a game come out that is heavily memory dependent. Ultimately the same advice also applies to multi-GPU users as well as IGP: avoid 1600 MHz and below.

I keep looking at the discrete GPU benchmarks and I fail to see any reason for even considering anything above 1600. Now they did test with a 6950 so it might not act the same way a 7970 or 780 would, but considering that the CF tests show relatively small improvements despite vastly higher bandwidth requirement... I doubt 7970 or 780 would show us anything different. As for CPU, there are a few percentages to gain if you go full 3000-ish, but nothing significant by just one or two steps up in speeds. And most of us will probably consider 1600 vs 1866 or possibly 2133, but not 3000.

I realize the focus of the article was more academical and that IGP was the main focus.. but I think it failed to educate some of the user base that could save lots of money buy not getting memory that is unnecessarily expensive.

Any thoughts?

 

[amitkher]

Yes, I too think the conclusion was wrong. They show prices difference between 1600 and 1866 MHz to be substantial. They also say that always prefer high amount of RAM (an oversimplification) over speed. And, like you say, performance difference between 1600 and 1866 is nothing to write home about, especially in the enthusiast common case of discrete graphics card. Only enthusiasts will read such article.

Given all these, their conclusion of 1866 being the sweet spot doesn't make sense.

 

[Cerb]

Any thoughts?

😕

Not including PCmark, for a regular usage benchmark: 😕

Discrete GPU performance at non-native res for anyone buying new CPUs: 😕

Minimum FPS, with the above, that's high enough that most anyone would ahve increased their eye candy: 😕

Average FPS...well, see above. If the average is 200 FPS...

Normalizing to varying values per chart, based on the lowest recorded actual performance, rather than a common RAM speed, the slowest RAM speed tested, or the maxmimum RAM speed tested: 😕 It makes it harder to follow and compare different charts.

The data itself is OK, but some items should be tested better, and either weighted differently, and/or with different conclusions based on usage. For example, Dirt 3 gets over 144FPS in all configurations (IoW, all RAM you can buy is too fast for their test). This means that either (a) Dirt 3 was a bad choice, (b) it was tested with bad settings (bad meaning not applicable to players of the game, or any other with similar RAM scaling behavior), or (c) 1333@CAS9 is fast enough. The conclusion from that very data that faster RAM is a good value is pointless. To make kind of evaluation matter, the game(s) and settings combinations need to be among those for which the minimum FPS is near or below common monitor refresh rates. The DGPU testing should not be made at a craptop+IGP resolution, if nothing else.

 

[TridenT]

Outside of Winrar... I didn't see any significant differences for my use scenario. I'm most worried about video compression performance and game play performance. It seemed negligible in all those benchmarks. (Within margin of error)

In this case, you can cherry pick data to make it say what you want.

 

[WhoBeDaPlaya]

Well, looks like I'm sitting pretty 🙂 (1866 CL9)

 

[Nec_V20]

As I stated from my own experience in other posts, the price/performance is just not there with regard to RAM above 1600 CL9.

The wonders of RAM are marketdroid hype more than anything else.

 

[mfenn]

I keep looking at the discrete GPU benchmarks and I fail to see any reason for even considering anything above 1600.

I agree wholeheartedly. I read through the article and thought it was decent enough until I got to the conclusion. When I read Ian's analysis, all I could think was "how did you arrive at this conclusion based on your data?"

I find this be an issue with many of Ian's articles to be honest. He's really far out there on the competitive overclocking fringe and gets a lot of free hardware, so he has kind of forgotten what value means to the rest of us.

Side node, kudos to Cerb for paying much more attention to methods than I. I didn't notice the peculiarities he pointed out.

 

[Shivansps]

The IGP gaming part should have done at minimum details, high details shift the bottleneck to the IGP itselft, making memory irrelevant, like the test shows up.

 

[Emulex]

why would you want built in graphics? seems like you give up way too much.

i'd rather have 192gb of ram and a video card (nvidia VGX) than being stuck with 32gb which is archaic these days!

 

[Shivansps]

Its not the point, if you are testing for memory scaling and bring in numbers for IGP it need to be done right.