HD5770 bottlenecked by 9800GT for PhysX

[InvisiBill]

I have a few different cards around here, so I thought I'd test this out.

My system had a GTX285 for graphics and a 9800GT for PhysX. I found a good deal so I decided to buy an HD5870. While that was in shipment, I decided to pick up an HD5770 locally to make sure that I didn't have issues with a hybrid ATI/Nvidia system.

This is all running on my Classified E759/i7-920 system. Overclocked to 3.3GHz by turning the BCLK up to 166. Turbo and HT enabled.

Here are the video cards I'll be using for the main graphics.

1. The HD5770 that I'm using temporarily.
2. My new HD5870. You can see that it's almost exactly two times the 5770 in every spec.

Here are the PhysX cards that I'll be using.

1. My 9800GT PhysX card. I put an Accelero on it and overclocked it a bit via the BIOS (600/900 -> 650/1000).
2. My friend's GTX275 that I borrowed for benchmarking. Bone stock.
3. My GTX285, my previous graphics GPU. Bone stock.

I'm testing with Batman: Arkham Asylum. These are the settings I used to play through the game (minus VSync). No tweaking done in the ATI control panel. Note that they do recommend a 9800GTX+ for the High PhysX setting.

[SIZE=+3]Results[/SIZE]
I ran each test three times to ensure that I got consistent results. Numbers listed are minimum, average, and maximum FPS.

          [color=#00FF00]9800GT     GTX275      GTX285[/color]
        31-76-126  29-79-130   30-81-134
[color=#FF0000]HD5770[/color]  31-77-126  30-70-129   31-83-135
        31-76-127  30-79-130   28-82-136

        59-98-130  63-106-156  65-107-157
[color=#FF0000]HD5870[/color]  62-97-130  63-105-156  63-108-165
        62-98-132  60-104-156  61-107-163

You can see that even with just the HD5770 for graphics, upgrading from the 9800GT for PhysX does increase the average and maximum framerates. Especially with VSync locking you to 60fps, you probably wouldn't notice the difference, but the game does perform better with a beefier PhysX card. Even without an uber top of the line card, a weak PhysX card can hold you back.

In reality, there's not a huge difference in performance between a budget 9800GT and a rather-expensive GTX285 though. The GTX285 is definitely better in all of the tested scenarios, but probably not enough to justify the extra ~$200 for most people. You'd definitely be better off upgrading your graphics card from a 5770 to a 5870 rather than upgrading your PhysX card from a 9800GT to a GTX285.


[SIZE=+3]Without PhysX[/SIZE]
For comparison, here are the results without PhysX. I tested both disabling PhysX in the Nvidia Control Panel while leaving the game set to High, as well as turning off the PhysX effects in-game. Note that turning off PhysX effects dramatically changes the game's appearance. Example 1 Example 2

        [color=#00FF00]Disabled     Off[/color]
        11-17-38  74-125-162
[color=#FF0000]HD5770[/color]  12-18-39  76-128-167
        11-18-40  78-128-166

        11-18-42  102-216-296
[color=#FF0000]HD5870[/color]  12-18-37  103-217-301
         9-16-42  106-218-310

[SIZE=+3]Single-card Graphics/PhysX (Newer GPU - Feb. 2015)[/SIZE]

I noticed my HD5870 was almost 5 years old and decided to upgrade. I got a GTX970 (specifically the EVGA GeForce GTX 970 SSC ACX 2.0+), and figured I'd do some testing with the new card. Honestly, nothing else relevant to this comparison has really changed in my system. I installed the latest Nvidia drivers and replaced my Radeons (I had also been using a secondary HD5830 for BitCoin mining and second monitor duty) with the GTX970. I used all the same in-game settings. My Nvidia Control Panel settings are fairly high, as I don't really have any new games, so the quality is more beneficial than extra (wasted) performance.

         [color=#FF0000]PhysX Off[/color]     [color=#00FF00]GPU PhysX[/color]
        122-280-400   78-141-197
[color=#00FF00]GTX970[/color]  115-276-412   74-140-189
        115-277-388   78-142-205

Just to explicitly state the results, the GTX970 by itself is approximately 30% faster than the HD5870/GTX285 combo, averaging 140+ rather than ~107 (similar percentage compared to just the HD5870 too, without PhysX). Newer hardware is faster than older hardware, in case you didn't get the memo. What surprised me is just how much of the GPU the PhysX processing takes, even on a beefy new card (since I didn't previously test a single card with and without PhysX). Enabling PhysX cuts the average and max framerates in half (min is only about a 1/3 drop).

 

[MuZI]

Do you have results of just the HD5770 alone?

 

[InvisiBill]

I'll run the tests without PhysX too, just for comparison. I don't want to mess with all the hacks to get CPU PhysX (i.e. the other way to get PhysX on an ATI system).

Keep in mind that the game is very different without PhysX.
http://www.youtube.com/watch?v=luSAnouAFJs&hd=1
http://www.youtube.com/watch?v=iWlmIjNiqWM&hd=1



Ok, test results have been added to the original post.

 

[ChucklesC6]

I have a 5870 in my rig and a spare 9800GTX+ laying around. How did you get the physx to work with ATi card?

 

[InvisiBill]

I have a 5870 in my rig and a spare 9800GTX+ laying around. How did you get the physx to work with ATi card?

http://www.google.com/search?q=GenL+PhysX+mod. I've also got some details of my own personal issues with it at http://hardforum.com/showthread.php?p=1035542384#post1035542384.

 

[MuZI]

Any updates on this with the HD5870??

 

[InvisiBill]

My original 5870 deal was, as I suspected, too good to be true. I ordered one elsewhere, and it will be delivered tomorrow. I should have the numbers up in the next few days.

 

[Aaron11]

I'm interested in seeing how the 5870 would do. I'm thinking of maybe buying one and a 9800gt. Would a 9600gt be at all adequate for a physx card?

 

[InvisiBill]

I'm interested in seeing how the 5870 would do. I'm thinking of maybe buying one and a 9800gt. Would a 9600gt be at all adequate for a physx card?

This thread shows that a 5770 is bottlenecked by a 9800GT for PhysX. The bottlenecking will only be worse with the 5870, since the graphics card is faster while the PhysX card is the same. Stepping down to a 9600GT would make it even worse, since the 9800GT has almost twice the shaders (112 vs. 64).

I'm using a 9800GT because I had the card from before. If you're specifically going out and buying one, you'd probably be better off spending a little more and finding deal on a GTX260, if you want better PhysX performance. I guess it depends on the deals you can find.

 

[AMD_Gamer]

would you be interested in doing a review for my website, http://www.gamephys.com ?

let me know.

 

[DeanEH]

I have a HD5870 + 9800GT. When I run the in game bench mark the Max. for normal PhysX is 25 so I would say that a 9800GT may be to slow for a HD5870. I have run this twice and get the same score. Would a 9800GTX+ be any faster for PhysX?

 

[MuZI]

Is it just me or are people asking the same question over and over again just worded differently...

 

[InvisiBill]

I have a HD5870 + 9800GT. When I run the in game bench mark the Max. for normal PhysX is 25 so I would say that a 9800GT may be to slow for a HD5870. I have run this twice and get the same score. Would a 9800GTX+ be any faster for PhysX?

Look up a few posts. The 9800GT is too slow for a 5770, so it's only logical that it's going to be too slow for a 5870, which is even faster.

The 9800GT has 112 shaders at 1500MHz. The GTX+ has 128 shaders at 1836MHz. Yes, it will be faster. For comparison, the GTX285 has 240 shaders at 1476MHz (over twice as many as the 9800GT and running only a hair slower). You'll have to determine if that's good enough for your own needs though.

You stated that you got 25 max with Normal PhysX. What exactly do you mean by that? As you can see from the numbers in the original post, the 5770 got 31-127fps (76 average) with the 9800GT on High PhysX. With the PhysX card disabled but the game still trying to use High PhysX, it got 11-40fps (18 average). If you got a maximum of 25fps on any test, I don't think your PhysX is working properly.

 

[vjcsmoke]

How important/big of a hit is PhysX performance in Batman:AA?

Let's consider the most ridiculous scenario possible.

2x GTX 480 + 1x GTX 480 dedicated physx processor. That's probably the best you can get...

But would having the dedicated physx processor beat using 3x GTx480 in SLI without using a dedicated PhysX processor?

Where is the point where, the dedicated physX processor reaches a point of diminishing importance vs how much power is dedicated to the rendering cards?

Lastly, should we care enough to spend $200+ for a dedicated PhysX processor? After all, a GTX 260 is $200, a GTX 275 is $250, and a GTX 285 is $320-400. Meanwhile a GTX 470 is $350 and a GTX is $500 (if you can find them without a jacked up price).

At this point I don't think there's any reason to buy a 285 for physx processing, when a 470 is the same price or less but faster.

 

[InvisiBill]

Ok, got the numbers from the 5870 with the GTX285 up. Haven't had time to swap the 9800GT back in for those tests yet.

With PhysX set to High in-game but the Nvidia card disabled, the 5870 posts almost the exact same numbers as the 5770. The PhysX chugging along seems to be the bottleneck here.

Using the 285 for PhysX, the 5870 is basically 30fps higher across the board.

With PhysX effects turned off in-game, the 5870 is 30fps higher on minimums, 90fps higher on averages, and 130fps higher on maximums. This basically has nothing to do with PhysX, but gives you an idea of how the 5770 and 5870 compare to each other in general.

 

[vjcsmoke]

I'm looking forward to seeing the numbers for 5870+285 vs 5870+9800GT. Obviously I think that the 9800GT is going to be MUCH slower. What might be interesting is using a value card like a 250GTS for PhysX instead. That seems to perform rather well as a dedicated processor, nearly as well as a GTX 260, and of course the price point is lot more attractive than any of the GTX series for dedicated PhysX use. Though perhaps GTX 480 GTX SLI + GTX 470 dedicated PPU might be the way to go for high end physX. [Although your ears might bleed! ]

 

[MuZI]

Ok, got the numbers from the 5870 with the GTX285 up. Haven't had time to swap the 9800GT back in for those tests yet.

With PhysX set to High in-game but the Nvidia card disabled, the 5870 posts almost the exact same numbers as the 5770. The PhysX chugging along seems to be the bottleneck here.

Using the 285 for PhysX, the 5870 is basically 30fps higher across the board.

With PhysX effects turned off in-game, the 5870 is 30fps higher on minimums, 90fps higher on averages, and 130fps higher on maximums. This basically has nothing to do with PhysX, but gives you an idea of how the 5770 and 5870 compare to each other in general.

I find this to be a bit strange...

As a general rule of thumb would be be wise to say if you want to use PhysX get a PhysX card on the same level as the main card?

 

[InvisiBill]

With PhysX set to High in-game but the Nvidia card disabled, the 5870 posts almost the exact same numbers as the 5770. The PhysX chugging along seems to be the bottleneck here.

I find this to be a bit strange...

Yeah, the game even gives a popup warning when you select that if it doesn't find an Nvidia card. I'm not sure exactly what it does without a PhysX card - the crippled CPU PhysX I guess? Trying to run High PhysX without a PhysX card is painful to watch (16-18fps average), so I wasn't surprised when the two ATI cards put up essentially identical numbers. The 5870 is a bit higher on maximums, which is probably from an area with almost no PhysX stuff going on, so the extra power of the 5870 allows it to pull ahead a bit.

As a general rule of thumb would be be wise to say if you want to use PhysX get a PhysX card on the same level as the main card?

I've seen it recommended (on the Nvidia forums I believe) that you should use a PhysX card with half the shaders of your main card. I know that the number of shaders doesn't translate directly between ATI and Nvidia, but it gives you an idea of what to aim for.

However, the 5770 test shows that there's noticeable improvement going from a 9800GT to a GTX285. I've seen the 5770 compared to the GTX260, and Tom's chart this month put it in the same level as the 8800GTS/GTX, one step above the 9800GT. I would therefore say that the 9800GT is more than half of a 5770, and this test shows that you get improvements going even higher than that.

If I had more PhysX cards to test, I might be able to find a sweet spot for performance/cost. If you want to add on a good PhysX card for maximum performance, I'd say get the best card you can afford. Even if PhysX is replaced by an OpenCL competitor, you should still be able to run it on the dedicated card just as you do with PhysX (so it's not like it would be a complete waste). As far as standalone goes, bigger == better.

A more important question might be whether it's better to use one new card for graphics and PhysX, or use the new card for graphics and keep the old card for PhysX. This question has already come up, and there don't seem to be a lot of hard numbers out there. What level of dedicated PhysX card does it take to outperform running it all on a single GTX480 - GTS250, GTX260, GTX285? Offloading the PhysX to any card will reduce the load on the main GPU and allow it to focus on just the graphics, but if the PhysX card is limiting the system to 70fps, then it's pointless for your main card to jump up from 300fps to 400fps.

As long as your PhysX card isn't bottlenecking it below your desired minimum, then I guess you could say that any card that meets that minimum is "good enough". If you're VSynced at 60fps, then it doesn't really matter if the GTS250 gives you 70fps while the GTX285 gives you 90fps, other than planning for heavier workloads in future PhysX games.

Also, I just realized that my friend has a GTX275. I might be able to swap him my 285 for a couple days to get some more numbers. I'm sure they'll be just a bit slower than the 285's numbers, but it may help pinpoint that perfect value spot. I wish I knew someone who had a GTS250 that I could test for a bit, as that seems to be a popular choice and a lot closer to the middle of the 9800GT-GTX285 gap than the GTX275 would be.

 

[drnilly007]

Isnt the physx card supposed to help the main gpu. I would think that if you put an 8600gt it would help with the processing unless the main card is just way better at working by its self than with help?

 

[MuZI]

Isnt the physx card supposed to help the main gpu. I would think that if you put an 8600gt it would help with the processing unless the main card is just way better at working by its self than with help?

http://hardforum.com/showthread.php?t=1492813

Read the posts made by Unknown-One. It should help explain what's going on.

 

[drnilly007]

I find this to be a bit strange...

As a general rule of thumb would be be wise to say if you want to use PhysX get a PhysX card on the same level as the main card?

Why then would anyone enable physx wouldnt just running in sli/crossfire be more FPS effective?

 

[MuZI]

Why then would anyone enable physx wouldnt just running in sli/crossfire be more FPS effective?

It would be much more FPS effective but you wouldn't get the cool effects of PhysX.

Or did you mean SLI and then have PhysX enabled like that?

 

[LAiN_OfTheWired]

OK, so you don't get the raw frames... but you get a better visual experience without the horrid framerate drop... hmm... I'll be doing some testing with crossfired 5770s and an OCed 9800GT next weekend.

 

[InvisiBill]

Got the HD5870/9800GT numbers posted. The drops between the GTX285 and 9800GT seem to be pretty similar whether you've got a 5870 or 5770 as your main GPU. There's not a huge increase upgrading your PhysX card from a 9800GT to a GTX285, but there is an improvement, if you're one of those people that has to have the best of the best.

OK, so you don't get the raw frames... but you get a better visual experience without the horrid framerate drop... hmm... I'll be doing some testing with crossfired 5770s and an OCed 9800GT next weekend.

I played through the game with my GTX285/9800GT setup first, simply because that's what I had. I always had PhysX, so I had no idea what the game was like without it. Example 1 above shows good close-up comparisons with and without. Watching this, it's hard to say that PhysX really changes the game much - you can kick papers around and break tiles and there's some fog. Big deal. However, Example 2 is a split-screen version of the benchmark run with PhysX off on the left and on on the right. You can see exactly how the environment is different between the two. The steam/fog in the tunnel scenes is the biggest ambient factor for me. With PhysX, it seems dark and spooky. Without PhysX, it seems like something from 10 years ago. It's high res and good artwork and stuff, but it's just a plain tunnel like in any other game.

Assuming that CF 5770's would be equal to a single 5870, I think that for the overall experience I'd rather have a single 5770 with a 9800GT for PhysX than a pair of 5770's for the higher framerate. That's not exactly a fair comparison of cards, but I really think the PhysX effects add a lot to the game. I'd rather have "good" performance with PhysX than "great" performance without. Obviously this is going to vary person to person, taking into account what old hardware they already have, what deals they can get, what resolution they run, etc. With just my 5870, I can get 100-200-300. Using the 9800GT for PhysX, it slows it down to 60-100-130. Since I'm stuck at 60fps with VSync anyway, the extra performance of not using PhysX is wasted. With the 9800GT, it's still "good enough" performance and the game has a lot more ambiance to it.


P.S. While I'm a big fan of the PhysX effects in this game, don't consider that an endorsement of all aspects of PhysX itself. I think Nvidia's "ATI lockout" and essentially bribing devs to use PhysX are bad things. Even in this game, there seem to be artificial limits imposed for these effects (since reportedly poeple were able to get PhysX running on the CPU with 30+fps). I can't wait for an OpenCL replacement for PhysX to be embraced (or even for an OpenCL wrapper for PhysX to come out).

Generally speaking, I think realistic physics will help move things forward. Remember when you were just a dot in a video game? Remember when they added sprites so that you actually looked like a little guy? Remember when they went to actual 3D models? Remember when they went from being a few 3D cubes with a picture of a guy stuck on it to an actual skeletal model? Remember when they went from being a 3D model of a stick figure to actually having modeled details like hands and fingers? Realistic physics will move us further forward. The more accurately we can model the molecular interactions of the real world, the more realistic the resulting animations will be. In my eyes, our goal is the photorealistic renderings that are done on server farms for movies (or even better than that), but being able to do that dynamically in realtime as you play your game.

I <3 physics. I like PhysX only because it's basically the only option at the moment.

 

[InvisiBill]

Borrowed my friend's GTX275 and posted those numbers up. They're actually closer to the GTX285 than I thought they'd be.

 

[LAiN_OfTheWired]

That shouldn't really be a surprise considering physx cares more about the shader count and speed than memory amount or bandwidth.

Also, I actually ended up with a 9800GTX+ instead of a 9800GT for physx so I'll post up some comparison numbers with my 5770 CrossfireX setup this weekend probably. I think the only real physx-based anything I have benchmark-wise is 3DMark Vantage. I don't suppose there's a stand-alone Batman benchmark? I guess it wouldn't hurt to download fluidmark, either...

 

[InvisiBill]

Necro-bump! I updated the first post with a few numbers from my new GTX970. There's not really anything amazingly new here, but it gives some idea as to how newer hardware compares to the older cards. It also includes a direct on/off comparison of PhysX on a single card, which I didn't do before (I didn't want to mess with removing my Radeons from primary GPU duty).

 

[LaCuNa]

Necro-bump! I updated the first post with a few numbers from my new GTX970. There's not really anything amazingly new here, but it gives some idea as to how newer hardware compares to the older cards. It also includes a direct on/off comparison of PhysX on a single card, which I didn't do before (I didn't want to mess with removing my Radeons from primary GPU duty).

I wasn't aware hybrid physx was functional in 2015

 

[davidm71]

I wasn't aware hybrid physx was functional in 2015

It's more of a novelty to run Hybrid Physx these days. I do with my 290X paired with a 450gts single slot card. I only keep it around to play Borderlands 2 when I'm in the mood. The most recent game that got a boost from it was Metro 2033 so I think it's still somewhat useful but your limited to an old Physx version.

 

[bigyellocup]

It's more of a novelty to run Hybrid Physx these days. I do with my 290X paired with a 450gts single slot card. I only keep it around to play Borderlands 2 when I'm in the mood. The most recent game that got a boost from it was Metro 2033 so I think it's still somewhat useful but your limited to an old Physx version.

I concur.

 

[davidm71]

Now that Nvidia released their physx libraries maybe some better mind can figure out a newer method to run hybrid physx side by side to AMD hardware. Not that it's all that relevant.