GTX480 / GTX470 Reviews and Discussion

[TheGreatGrapeApe]

As it mentioned in the VR-Zone post, 336 doesn't make sense, it's 56 clusters, which means it's 7 groups of 8.

How does that work into 6:4 or 8:6 ratio of the memory interface too?

Sound like it would need to be a crippled something or other, and it's doubtful that with all that nV put into the design of Fermi and the movement of something GTX class, that they would even bother designing for anything more than a 256 shader core, it would seem over-designed at the time of inception, whereas now it makes sense only because we've seen final product come out from the GF100.

As a mid-range new product 256 shader core seems more logical (although 640 seem more logical than the 800 in the HD4870 too).

 

[TheGreatGrapeApe]

isnt that just one game out of 4. in metro, avp and dirt it seems 480 slis are ahead .
even in badcompany it is close.

Yes, but there is no other link to another test, and he didn't mention other pages or other tests just including the original.

Also, when he said that it's tested with lower AA and get worse results, like I said.... not in the case linked, which like I also said... is the minority, not the norm.

And look at Metro, AA is exactly the same, and settings are exactly the same, and while performance is different that wasn't what was stressed in the reply, and in DiRT2 nV has a bug and the GTX480SLi actually has the lower AA settings not the higher since 16XAA was not available. So it doesn't fit his statement or your theory either.

Only AvP would match the statement and that makes it as you say "1 game out of 4" thus the minority, not the norm for settings, while perhaps being the norm for performance and BC2 being the minority for performance.

Once again for the original statement it does provide support for them being pretty close in performance if not price & power consumption, with each having their advantage.

Still PR points for both sides, but still doesn't affect SI, NI nor the GTX4 refresh.

 

[shadow187]

TGGA;
Do you think 240Shaders is enough to beat the HD5770?

 

[TheGreatGrapeApe]

Let's try to stay within the GTX4xx lines....

As a simple stopgap, I think SI is meant to completely replace evergreen.

I doubt it.

I would say SI will replace the HD5770 (needs it most due to the number of people whining about the performance compared to the HD4870) and also gives them an easier die to work with, and then also maybe replace the high end if they do care enough about the top spot.

I would also guess that the HD5770 replacement comes out similar to the G92 launch with a very performing mid-range part that challenges top spot, this will put added pressure on the GTX460 and give them a moving target to hit, which may be a bad thing if they are having any kind of difficulty hitting the HD5770 or HD5830.

A Full on GTX485 is supposed to be about 15% faster at best if they can hit full clocks with full cores. The 512 cores only provide a 6.6% boost on their own, so there isn't much there in just opening the cores, they need to be able to push the clocks too on a full die, even restoring the target clocks on a 480 chip would produce nearly that 15% on it's own (about 13.5%).

So if an SI 5870 refresh was only a minor tweak it would compete well with the GTX480 but likely not change the positioning versus the GTX485 if the 15% target were correct. But if it were even just a change of SPUs to 2000, that's a 25% change there, and that's without adding the change to uncore and cache which should offer efficiencies of their own in the things the Evergreen series is noticeably weaker at versus the GTX4xx series.

However no point in counting your die before they cool, as we've seen many times before, not just from Fermi, but alot of ATi too.

 

[TheGreatGrapeApe]

TGGA;
Do you think 240Shaders is enough to beat the HD5770?

Depends on how they are clocked and what they have to cut on the back-end which bothers me more than just the shader (everything points to a very crippled rear from early reports). If that's the case I don't think 240 with reasonable clocks would beat an HD5770 except in texture heavy and memory restrictive titles, nor would a 256 shader version either. Couldn't even speak to tessellation either without seeing the back end due to the surface requirements of tessellated objects being thrown out the back.

They will likely trade blows very well, and remember it's not the front end where the HD5770 is a little underwhelming, it's the rear, it has much more shader power than the HD4890, but if the GTX460 is equally crippled then they will both show this limitation in the same area, plus likely not see any advantages in AA as a differentiator either.

 

[notty22]

Hocp just released a new review today of the SOC gigabyte 5870. Its the highest o/c 5870 available and is priced at 500 dollars. Since we are talking about Metro 2033 #'s , here is another take. The card is tested against a gtx 480 (197 drivers) , a stock 5870.
Performance is on par with the gtx 480 at those clock speeds. The conclusions people can draw for themselves. IMO, the wattage is what I expected, but I like having the graph.
http://www.hardocp.com/article/2010/06/14/gigabyte_hd_5870_super_overclock_edition_review/1

These temp #'s don't scream , one is a beer cooler and the other a frying pan ?

 

[shadow187]

Like I've been saying, a $500 HD5870 ~= a GTX480. A $600GTX480 < HD5970 though...I seriously doubt that it can reach HD5970 levels on air.

Topic...odds of a GTX485?

 

[AMW1011]

The GTX 480 can gain 20% pretty easily since it gains so much from overclocking and it overclocks well. A GTX 480 at 800-850 MHz on the core should equal or be slightly faster than a stock 5970, and the GTX 480 can do 900 MHz on air too, if you are lucky. That said, the 5970 can overclock really well too so its a moot point.

 

[area61]

http://tpucdn.com/reviews/NVIDIA/GeForce_GTX_480_Fermi/images/perfrel.gif

The GTX 480 can gain 20% pretty easily since it gains so much from overclocking and it overclocks well. A GTX 480 at 800-850 MHz on the core should equal or be slightly faster than a stock 5970, and the GTX 480 can do 900 MHz on air too, if you are lucky. That said, the 5970 can overclock really well too so its a moot point.

Overclocking should not be considered as something that ONLY MATTERS. It should be considered as an added bonus. Comparing overclocked and stock cards are like comparing apples to oranges, unless the card is factory overclocked with a decent price tag

 

[welshmousepk]

Overclocking should not be considered as something that ONLY MATTERS. It should be considered as an added bonus. Comparing overclocked and stock cards are like comparing apples to oranges, unless the card is factory overclocked with a decent price tag

personally i tihnk overclcoking is very important. especially on high end cards liek this where the majority of buyers will be enthusiasts who know how to overclock.

but thats why i like to see two cards compared at their highest stable OCs. its pointless to compare one OCd card to another at stock. but with the 5970 at least, not taking OCing into account puts the card at a massive disadvatage. i can't imagine buying one and not overclocking it, that would seem such a waste.

 

[Luckyfox]

Are there any new coolin' solutions from 3rd party companies?
In march-april i saw on web some announcements about special fans for gtx480, like those triple fans coming with OC Zotac 480 premium.
Is there smth worth our attention on market today?

 

[area61]

get the gtx480 with the dangerden water block and water cool it...those are worth looking..

 

[Luckyfox]

@area61
well, i was asking about air cooling solutins..I know about dangerden waterblock.
I think i am not ready for watercoolin thing. It comes with additional expences(buying rad,res and all the stuff) and it needs every year maintance cycle, which for me, lazybone, is kinda hard.

 

[RealityRush]

Are there any new coolin' solutions from 3rd party companies?
In march-april i saw on web some announcements about special fans for gtx480, like those triple fans coming with OC Zotac 480 premium.
Is there smth worth our attention on market today?

Eh, the stock 480 coolers are petty much the best air cooling you're going to get, just reapply the TIM on it and apply custom fan profiles and you're golden. Any of the aftermarket air coolers really aren't as good, which is fine 'cause the stock cooler is great anyways and you save money if you want to air cool 😛

If you want to watercool to try to hardcore OC the cards, than yeah, the Dangerden blocks are always worth looking into. I think EK made some too but don't quote me on that.

 

[RealityRush]

Not to piss anyone off here, but I'd like to point out the stock 5870 is only 4fps behind the 480, which is more like 7-8fps with the 256 drivers (why' they use 197, wtf?), and the SOC 5870 is like 5-6fps behind. Proportionally speaking, it isn't all that much closer to a 480 considering the extra power it uses, although I guess you can only gain so much from OCing. So shadow, a SOC 5870 isn't really ~= to a 480 at all. It's a bit closer than a stock 5870, but certainly doesn't match the 480.

That's like saying, if a 480 is getting 80fps and the SOC 5870 is getting 70fps and the regular 5870 gets 60fps, the SOC 5870 is ~= to the 480. It's the same proportions as above, but as you see it is more obvious that it isn't really equivilant. Not to mention you're comparing an equally priced card with less performance and now 0 OCing headroom as it is already OC'd whereas the stock 480 has much more room.

But that bench shows them all so close, and with no AA and without the latest drivers no less, so I'm disappointed as it is all a little underwhelming. I want to see some more polarized benches with higher fps numbers so it is easier to tell the difference, cause there's a lot of things that can play into that few fps difference in Metro 2033. Other benches might help the 5870 do better than Metro. Although adding AA might nullify that for all I know.

From what I see right now, I'd say the SOC 5870 should be more like $450 to be a decent buy, actually more like $420 'cause of the loss of the ability to OC it more.

Again though, we need to see more benches to compare them other than Metro, and regardless of that fact, notty, why are you bringing up the SOC 5870 in the 470/480 thread? Kinda the wrong place, especially seeing as it didn't beat the 480 so it isn't an important note as of right now 😵

 

[RealityRush]

@area61
well, i was asking about air cooling solutins..I know about dangerden waterblock.
I think i am not ready for watercoolin thing. It comes with additional expences(buying rad,res and all the stuff) and it needs every year maintance cycle, which for me, lazybone, is kinda hard.

Stick with the stock air coolers then, you can easily OC to 900MHz on them if you have a decent case and fan profiles. Haven't bothered with going above that, but yeah, the stock air coolers are fine. Anything aftermarket would be a huge waste of money.

 

[Luckyfox]

ok, thank you for your answers, i will probabily reapply TIM.
the heat from card is ok for me, but the noise at full load is really horrible.

 

[RealityRush]

ok, thank you for your answers, i will probabily reapply TIM.
the heat from card is ok for me, but the noise at full load is really horrible.

Download the newest version of Afterburner that works with the 480s (Mine is a beta 5 version or something, the final might be out now..), and setup a custom fan profile. I have mine start at around 40% at 40C or below--and during idle my cards never go above 40C--then just make the fans ramp up as an initially steep sloped curve that levels off as the temps get hotter, and set the max to whatever fan speed you tolerate noise wise, which is about 80% for me, mine jumps to 90% only if the temps get really excessive (which they don't except for Furmark 😛).

http://www.nvnews.net/vbulletin/showthread.php?t=149400

There, that link, look at the bottom of it for the guy's custom fan profile.

Mine has a steeper more aggressive curve at the start, but even his is miles ahead of nVidia's crap one, which wont ramp up past 44% until you hit like 80C.......

 

[shadow187]

@RR
If you're making the proportions bigger, every card looks like it's worse compared to the better card. You can't say, "These drivers add XXX FPS" because it just doesn't work like that.

@AMW
You're not going to OC a GTX480 to catch up to a 5970 that easily. That 20% is across all resolutions.

 

[area61]

Not to piss anyone off here, but I'd like to point out the stock 5870 is only 4fps behind the 480, which is more like 7-8fps with the 256 drivers (why' they use 197, wtf?), and the SOC 5870 is like 5-6fps behind. Proportionally speaking, it isn't all that much closer to a 480 considering the extra power it uses, although I guess you can only gain so much from OCing. So shadow, a SOC 5870 isn't really ~= to a 480 at all. It's a bit closer than a stock 5870, but certainly doesn't match the 480.

That's like saying, if a 480 is getting 80fps and the SOC 5870 is getting 70fps and the regular 5870 gets 60fps, the SOC 5870 is ~= to the 480. It's the same proportions as above, but as you see it is more obvious that it isn't really equivilant. Not to mention you're comparing an equally priced card with less performance and now 0 OCing headroom as it is already OC'd whereas the stock 480 has much more room.

But that bench shows them all so close, and with no AA and without the latest drivers no less, so I'm disappointed as it is all a little underwhelming. I want to see some more polarized benches with higher fps numbers so it is easier to tell the difference, cause there's a lot of things that can play into that few fps difference in Metro 2033. Other benches might help the 5870 do better than Metro. Although adding AA might nullify that for all I know.

From what I see right now, I'd say the SOC 5870 should be more like $450 to be a decent buy, actually more like $420 'cause of the loss of the ability to OC it more.

Again though, we need to see more benches to compare them other than Metro, and regardless of that fact, notty, why are you bringing up the SOC 5870 in the 470/480 thread? Kinda the wrong place, especially seeing as it didn't beat the 480 so it isn't an important note as of right now 😵

not to piss you off, but i dont think a gain about 4 or 7 fps matters much. That gain might translate to 3 -5 % difference which, actually is insignificant.....

 

[Mousemonkey]

As it mentioned in the VR-Zone post, 336 doesn't make sense, it's 56 clusters, which means it's 7 groups of 8.

How does that work into 6:4 or 8:6 ratio of the memory interface too?

Sound like it would need to be a crippled something or other, and it's doubtful that with all that nV put into the design of Fermi and the movement of something GTX class, that they would even bother designing for anything more than a 256 shader core, it would seem over-designed at the time of inception, whereas now it makes sense only because we've seen final product come out from the GF100.

As a mid-range new product 256 shader core seems more logical (although 640 seem more logical than the 800 in the HD4870 too).

336 Shaders may not make sense but I can't help but add fuel to the fire, sorry mate! :lol:

I'd like to know where they get their drivers from though.

 

[TheGreatGrapeApe]

Not to piss anyone off here, but I'd like to point out the stock 5870 is only 4fps behind the 480, which is more like 7-8fps with the 256 drivers (why' they use 197, wtf?), and the SOC 5870 is like 5-6fps behind.

The reason [H] tested with the 197.75 is because they have a long standing policy for regular hardware reviews to test WHQL versus WHQL unless they are specifically testing Beta versus Beta or a new beta feature. As it is a hardware review and there is no ATi Beta, so it's WHQL vs WHQL. The reason for this should be clear to anyone with a history of nVidia and [H].

So for that it may not be a global test, but until the drivers go WHQL, then there was really no option... until TODAY, when the 257, not 256 package got certified. And obviously there was something not up to snuff if they sent the 257 candidate for certification.
Now you will see that in their (and others) later tests since it's the one certified. But Metro is definitely on the big # changes.

As for proportionality, it's really a question of are the differences significant enough to you, does the performance difference warrant other differences, does overclocking give diminishing returns in some areas or is it more global. The 20% scenario is best case, and that 800-850Mhz number is 50-100Mhz above the target frequencies they couldn't hit for launch, do you think they are really going to get reliable enough yields for that now for a retail product?
People can hit 1Ghz on air with the HD5870 or 5970, but I don't expect to see a model anywhere near that speed for retail, even the special edition overclocked models.

It's a little unrealistic to expect that "20%" as a retail product even in the premium space, and I doubt a major refresh is coming this year to aide that either, I think they will simply refine the GTX480 with the 4th or 5th SPIN or whatever they'll be on to fix the issues holding the SPUs back and re-launch that, since the 512 SPU part would better help the Tesla and Quadro market long term rather than trying to push insane clocks (which dramatically increases power consumption).

GTX 485 I would expect to be a 512 model, and expect clocks to be near original target @ 750Mhz with the factory OC'd models at 800Mhz top on air (whatever they can on water).

 

[shadow187]

I thought drivers stopped at 257?

 

[area61]

seems fudged to me tho..that 336 shaders probably some deactivated just like the gtx480...just my wild guess

 

[TheGreatGrapeApe]

336 Shaders may not make sense but I can't help but add fuel to the fire, sorry mate! :lol:

I'd like to know where they get their drivers from though.

It's possible, but then it's definitely a disabled cluster, and like I mentioned earlier, I'm really more interested in the back-end since it's the are where on first blush it looked like they cut too much away trying to fit inside package size. IIRC it was like 1/2 ratio of the GTX470/480, essentially 1/4 depending on clocks. The HD5770 is also crippled on the back end, but only about 1/2 the HD5870, so this will be a closer battle IMO, with the GTX460 having alot of shader power but not much in the back-end. This makes sense for TESLA and Quadro products if they are fully unleashed, but the retail model will likely be equally crippled as the GTX480 (or else it would outpace the GTX480) for GPGPU tasks. Of course if they just wanna sell a boatload, then don't cripple the GTX460 and let it be a Folding and GPGPU value play (now that would be interesting to test with some Quadro/Cuda hacks). I think that depends a bit on yields, if they think they can produce a boatload, then it makes sense to open up the DP rate for folders, etc.

I thought drivers stopped at 257?

I think those are either leaked betas many generations/revisions out (or meant to look like betas if faked).
The device ID info is interesting, doesn't give detail on what die it's based on but looks very GF104-esque based on the positioning of those features specs.