5770 is NOT good as 4870 or GTX260

[rancherlee]

I'm kinda thinking its still more of a driver issue than a memory bandwidth issue on the 5770 cards. Granted the best I can test is 1920x1080 but cranking up the memory by 250mhz (1200->1450) didn't really show any gains in benchmarks I did, at least not the gains it should have had if the bandwidth was the issue. A small 50mhz bump on the core had more gains than bumping up the memory 250mhz. Ya, they are a bit overpriced for the performance but I'm In the boat of only upgrading every few years so DX11 was important to me and when I need more performance later on I can always toss another 5770 in the other slot.

 

[AzN]

The thing is you didn't really test what I'm implying in the same thread. You tested average frame rates that only has small difference with bandwidth. Okay I understand some of you 5770 owners want to justify your purchase but let's not sway from fact from fiction.

It's not like 5000 series is a new architecture. It's basically 4000 series with minor improvements to SP. There's isn't a magic driver that will make your 5770 faster. Perhaps in SP hungrier games that are the exceptions and few at that.

 

[Meaker10]

I'm sorry but several things point to it not being memory bandwidth:

1. Poor scaling when clocking memory large amounts.
2. Good scaling when clocking core.
3. Excellent scaling when adding a second card in crossfire.

I really think the drivers need to be sorted.

Remember that per SP the 5770 has more bandwidth than the 5850....

 

[Schmide]

Yes. The same kind of results as Far Cry2.

I think we're missing each others point. Clarifying.

Either we accept that the 5770 follows the general frame rate of the other two cards (H graphs), and the frame dips are very instantaneous, or we declare the 3 card's frame rates uncompariable and the dips of the 5770 are dramatic and long lasting.

I take the first stance.

 

[AzN]

I'm sorry but several things point to it not being memory bandwidth:

1. Poor scaling when clocking memory large amounts.
2. Good scaling when clocking core.
3. Excellent scaling when adding a second card in crossfire.

I really think the drivers need to be sorted.

Remember that per SP the 5770 has far more bandwidth than the 5850....

1. Which points to the Original post. I gave clear examples how the bandwidth effect minimum fps and not so much on avg fps.
2. Again also points to my original post. Core clocking peak maximum fps to get better avg frame rates.
3. No doubt about that. You are doubling everything.

Of course there's an equilibrium between core and bandwidth but in case of 5770 bandwidth has dramatic impact in minimum fps in these cases I've shown. I'm not saying 5770 in not unbalanced card. I'm just saying it's not good as 4870/4890/GTX260 with only 60% bandwidth and expect driver optimization to the rescue.

Radeon drivers are pretty mature and only need optimizing when new games come out that hasn't had the luxury. I think you probably meant 4850 which 5770 outperforms regardless of any driver optimization as it's clocked higher.

 

[Meaker10]

You do not double bandwidth with crossfire, you have to double all your writes for a start etc. There is an improvement, but you get FAR more shader power than memory bandwidth,

 

[AzN]

I think we're missing each others point. Clarifying.

Either we accept that the 5770 follows the general frame rate of the other two cards (H graphs), and the frame dips are very instantaneous, or we declare the 3 card's frame rates uncompariable and the dips of the 5770 are dramatic and long lasting.

I take the first stance.

I stopped reading Hardocp long time ago. The frame dips are spontaneous on H results even on their fastest cards vs slower cards. It seems like they never get rid of margin of errors to hard drive seeks.

This ideology is nothing new. It's been ongoing. G80 vs G92 comes to mind.

 

[AzN]

You do not double bandwidth with crossfire, you have to double all your writes for a start etc. There is an improvement, but you get FAR more shader power than memory bandwidth,

Ha you've mistaken. You double bandwidth not vram size.

 

[Meaker10]

No I meant 5850.

5850 = 1440 SP, 128 GB/sec bandwidth
5770 = 800 SP, 76.8 GB/sec bandwidth

True the 5850 is clocked lower, so overall they have similar bandwidth in relation to their processing power.

 

[Meaker10]

Ha you've mistaken. You double bandwidth not vram size.

Again you keep assuming I have made the mistake, it acts a bit like a raid array, when one card makes a write to memory, so does the other.

So for each write I make I have to write twice. There is also some extra overheads to make sure they do it right. Therefore you do not get twice the bandwidth.

 

[Schmide]

I stopped reading Hardocp long time ago.

He He. You don't have to read it, just look at the graphs. However, I don't know of another review that does it, so it may just be the way things are.

This ideology is nothing new. It's been ongoing. G80 vs G92 comes to mind.

The G80 across the segments, did have more raster units.

 

[AzN]

Again you keep assuming I have made the mistake, it acts a bit like a raid array, when one card makes a write to memory, so does the other.

So for each write I make I have to write twice. There is also some extra overheads to make sure they do it right. Therefore you do not get twice the bandwidth.

It is amusing when someone argues that bandwidth doesn't double in crossfire when it does.

 

[hans030390]

Tell you what. I run a loaded system with GTX 260 on a 450watt power supply capable of 30A to 12v. I have 2 hard drives, 2 pci cards, game controllers, etc and overclocked it for good measure and not a single hickup with power issues. Unless you want to run 300 watt dell systems I doubt it's an issue for most of us.

Wow, good for you. Not everyone has powerful and/or high quality PSUs.

 

[AzN]

Wow, good for you. Not everyone has powerful and/or high quality PSUs.

I hardly call a 450watt a powerful supply.

It's not even high quality. I bought it for $15.

 

[xenolith]

I always wondered about this...

When you guys refer to "bandwidth" in this discussion, are you referring to the card's internal bus?

With crossfire/SLI, I know memory size does not double but does the internal memory bus effectively double?

5770 + 5770 = 128bit + 128bit = 256bit overall bandwidth?

 

[AzN]

You get 2 128bit bus. It has the same bandwidth as 256bit.

 

[xenolith]

Then that could explain why 5770's performance in CF scales 100% in most benchmarks.

Sorry, I'm just trying to catch up...

 

[hans030390]

I hardly call a 450watt a powerful supply.

It's not even high quality. I bought it for $15.

Still, not everyone has that powerful of a power supply, regardless of how much you can buy one for.

Still, I got the card for DX11. No one actually knows if it'll be good enough a couple years down the road or not (that includes you). I'd rather have it than not if I can get similar performance to the 4870 or GTX 260.

 

[Daedalus685]

You get 2 128bit bus. It has the same bandwidth as 256bit.

You are totally ignoring what meaker10 said.

The memory on each card is a mirror. The writing has to be done twice (once to each card). You can only theoretically improve bandwidth on reads, all writing cycles will only provide you with the bandwidth of a single card, less in reality due to latency et al.

 

[Meaker10]

You will also have the cards accessing the same data twice too where as a single card may buffer it.

 

[T2k]

Remember that per SP the 5770 has more bandwidth than the 5850....

Word.

 

[dug777]

There are still some odd results that really don't fit with the memory bandwidth theory, in that xbitlabs test...

How do you explain the S.T.A.L.K.E.R.: Clear Sky results, where the 5770 beats the 4890 as far as min fps are concerned in all except one resolution?

Or NFS: Shift, where the min fps are faster at every resolution tested?

I suppose it could be driver issues...

 

[AzN]

There are still some odd results that really don't fit with the memory bandwidth theory, in that xbitlabs test...

How do you explain the S.T.A.L.K.E.R.: Clear Sky results, where the 5770 beats the 4890 as far as min fps are concerned in all except one resolution?

Or NFS: Shift, where the min fps are faster at every resolution tested?

I suppose it could be driver issues...

5770 although same clock as 4890 has upgraded shader and slightly better than previous gen.

this was discussed prior. Some games are more prone to SP. Yes these games do exist. NFS loves shader. Clear skys I don't know this engine that well but I suppose it is also have greater effects with shader in some aspects when 5770 is getting better results than 4890 when it comes to minimum frame and avg frame rates. But there are far more games out there that bottom out harder because of bandwidth.


RE5, Crysis, Far Cry2, COD, WIC, etc.

I'm not saying some games do not show some effects of SP and get not better results. Stalker has dx11 mod or something like that and swayed by SP and it shows but the difference is very little. 1fps difference isn't what I call an advantage now.

Fallout 3 is also 1 of those games that sway on SP to some degree. We can think back in time to oblivion. Need for speed engine as well. This game loves SP.

When it's all done the difference on SP forefront is very little difference if not subtle but the bandwidth isn't. It shows huge disadvantages where some games needs a certain bandwidth and 5770 just doesn't have enough and start bottoming out like in WIC, COD, Far Cry 2.

 

[dug777]

I think it would be very insteresting for us all for BFG to rerun his excellent bottleneck test, only watching minimum fps

 

[BFG10K]

1 game showing architectural differences does not mean drivers.

I don’t think you looked at the benchmarks properly, specifically the percentages.

That and architectural differences would make no difference in that situation if the card is as starved for bandwidth like you claim

You can't conclude on just on avg. fps alone.

Sure you can, because minimums affect averages by definition.

You should update your investigation further and show minimum fps of the 5770 where bandwidth has the most effects on instead of only showing avg fps that is swayed by core clocks to come up with that conclusion.

If I show minimums that prove you wrong, will you retract your claims and apologize to the entire forum for being wrong?

Or will you simply continue to tip-toe around any results that don’t back your incorrect notions?

I mean you don’t accept [H]’s benchmarks that clearly show plot points putting minimums into a context that don’t back your claims. Why should I waste my time if you’ll do the same to my results?