Hardware question on HD 5770

[BFG10K]

BFG, valid points. I think 5870 is not memory bandwidth limited since it has plenty of it. I am still not convinced that you can infer the same about 5770 from the firingsquad review you linked.

My inference comes from my own testing of the 5770, along with educated inferences based on the 5870. Here, I’ll inline my 5770 underclocking results to make it easier for all to see:

In each of those games, the yellow bar is the % drop after a 20% underclock is applied to the core, while the green bar shows the same for the memory. You can clearly see that apart from Bioshock, the two are either close, or the core makes a much larger impact.

The overall average for that graph is 11.20% for the core and the 8.46% for memory.

That means it’s actually quite a balanced part with the core being a bit more important. It’s simply not showing memory bandwidth being the primary bottleneck like we’re led to believe.

Another important point: the 5770 is literally a 5870 with the core and bandwidth cut in half. Hence if one doesn’t believe the 5870 is constrained by bandwidth, the same equally applies to the 5770 since everything is reduced in perfect proportion.

Although I am not sure what modern game can be played on 5770 at 2560x1600 8AA?

The game doesn’t have to be modern; all it has to do is hit the memory bandwidth hard, which that setting will do. In Doom 3 and Prey the 5770 outruns my GTX285 at that setting. Given the GTX285 has double the bandwidth, there’s no way the 5770 can be limited by its bandwidth in that situation.

Sure there are some improvements in RV870 core over 770 but this is not an architectural change. Also, 4890 sometimes gives 5770 a beating by such a large, that it's difficult to suggest a 30%-40% performance deficit is driver related since it's more or less the same architecture:

I don’t believe it’s quite that simple, as there have been some changes that could significantly impact performance. For example, the fixed function interpolators were removed and their calculations are now performance by the shaders. This will impact the performance of the AF, and possibly AA too.

The 5770 is sometimes slower than my 4850, and other times it outruns my GTX285. In UT2004 at 2560x1600 with 24xAA for example, it’s less than half the speed of the 4850.

You can’t blame those extremes on memory bandwidth because it has more than the 4850, and half of the GTX285’s. The only possible explanation for such fluctuating performance is the drivers.

Further, since 5870 hardly doubles 4890's performance, despite double the theoretical performance increase (other than memory bandwidth), my instinct tells me there is a bottleneck somewhere else in the architecture itself.

That could be true, but I’m still leaning towards the drivers as a hardware bottleneck can’t explain why it performs exceptionally in some situations, but dismally in others.

 

[BFG10K]

Thanks for correcting my post count.

No worries.

Your 8800ultra benchmarks were years ago and still is core starved for the amount of bandwidth it has.

I’m not understanding the context of this statement. The methodology of my 5770 testing was the same as it was on the 8800, and you accepted those results. In fact, you’re making the statement that the 8800 is core limited precisely because that’s what the results showed.

Or to put it another way, if you don’t accept the 5770’s results, then you can’t accept the 8800’s results either.

So it makes no difference how long ago they were run, or on what card they were run on.

Your Benchmarks were expected. Little more core and a little more bandwidth. Nothing much changed since 4850.

I most certainly didn’t expect them. After what was plastered online, I expected the opposite of the 8800, with the memory showing massively more impact. I think almost everyone was surprised by the results, myself included.

I have a slight different perspective as to say a card is bandwidth or core starved. Instead of looking at "only" average frame rates which usually sway by core because core peaks higher max frame rates to get more average frame rates. The minimum frame rates matter much as average frame rates if not more to a gamer which is swayed by bandwidth. Factor that in to the equation there you'll see the bottleneck.

A minimum/maximum is a single data point by definition, and hence is absolutely useless without a benchmark plot putting it into context. Without such a plot, you could simply be looking at meaningless benchmark noise.

Furthermore, if you reduce the primary bottleneck, then you can increase the min/avg/max just as easily. That is to say, memory can just as easily affect the maximum, and the core can just as easily affect the minimum.

I disagree with your automatically implied correlation between core = maximum and memory = minimum. It might be true in some cases, but the opposite could equally be true too.

That and a higher minimum can just as easily raise the average as a higher maximum can, given the average is by defintion the mid-point of the two.