Tested: Nvidia GeForce GTX 580 Video Card

By Norman Chan

The green machine comes back swinging.

It was only seven months ago that Nvidia debuted their Fermi-architecture video card at PAX East, the GeForce GTX 480. Plagued by numerous production delays attributed to low chip yield and architectural growing pains, the GTX 480 was somewhat of a disappointment. Though from a performance standpoint, the card was a DX11 powerhouse, reviews compared it unfavorably to AMD's established Radeon 5870, citing faults in noise levels, heat output, and power consumption as reasons to avoid the $500 card. But with the release of the arguably misnamed 6870 GPUs, AMD has seemingly rested on its laurels, allowing Nvidia a chance to make up for its mistakes and catch up in this heated GPU race.


 
Let's dive into what different about the card, and how it performs against its nearest competitors.
 

Architectural Improvements


 
But to run with 512 cores and not suffer the same heat and power problems of the GF100, Nvidia had to re-engineer some the GF110 at the transistor level. More accurately, they reallocated the type of transistors used in different parts of the core, opting for lower "leakage" transistors on low-priority processing paths. These lower leakage transistors aren't as fast as the more wasteful fast-switching transistors, but Nvidia has placed them in non-critical areas to maximize performance while keeping heat down. As a result, the GF110 can run 32 more SM cores than its predecessor, while also using slightly less power; the GTX 580 operates at a max TDP of 244Watts, down from the blistering 250Watts of the GTX 480.
 


Betting Big on Tessellation

 While the GTX 580 is obviously no slouch when it comes to pure core specs, there's more to this card than brute force. In our meeting with Nvidia, they made it clear that their performance goal with the GTX 580 was to not only be the fastest GPU, but take a clear lead in next-gen gaming performance. That means DX11 and tessellation performance. There are 16 dedicated tessellation processors (Nvidia calls them Polymorph Engines) in the GF110, diminishing the performance drop when this feature is turned on in DX11 games like Civilization V and Metro 2033. Tessellation, if you recall, is the ability to dynamically add geometric (physical) detail to in game objects without the use of high-polygon models. This allows for highly-detailed terrain and character detail at a relatively low performance cost.
 

    
   
    
  
We ran the demos on our test bed at 1280x720 for the purposes of recording, and the minor dips in framerate were due to CPU and hard drive limitations. At full-screen 1920x1080, the demos ran perfectly.

Meet the GeForce GTX 580


 

 



 

Speeds and Feeds

 
SpecNvidia GTX 580
Nvidia GTX 480
Nvidia GTX 470
Nvidia GTX 460 1GB
Transistors 3.0B
3.0B
3.0B
1.95B
Die Size
529 mm^2
529 mm^2
529 mm^2
368 mm^2
Compute Cores/
Stream Processors
512
480
448
336
Texture Units 64
60
56
56
ROPs
48
48
40
32
Core Clock
772MHz
700MHz
607MHz
375MHz
Memory Clock
1002MHz
924MHz
837MHz
900MHz
GDDR5 VRAM
1.5GB
1.5GB
1.2GB
1GB
Memory Bandwidth
384-bit
384-bit
320-bit
256-bit
Process
40nm
40nm
40nm
40nm
Max TDP
244W
250W
215W
160W
Idle TDP
47W
47W
33W
22W
 
And here's how the GTX 580 compares to AMD's top cards (note the 6870 is not the high-end AMD part).
 
Spec GeForce GTX 580
Radeon 6870
Radeon 5870
Radeon 5970
 Transistors3.0B
 1.7B 2.15B2x2.15B
 Die Size
529 mm^2
255 mm^2
334 mm^2
334 mm^2
Compute Cores/
Stream Processors
 512
(not comparable)
 1120 16002x1600
Texture Units
64
 56 802x80
ROPs
48
 32 322x32
Core Clock
772MHz
 900MHz 850MHz725MHz
Memory Clock
1002MHz
1200MHz
1200MHz
1000MHz
GDDR5 RAM
1.5GB
1GB
1GB
2x1GB
Memory Bandwidth
384-bit
256-bit
256-bit
256-bit
Process
40nm
40nm
40nm
40nm
Max TDP
244W
151W
188W
294W
Idle TDP
47W
19W
27W
42W
 

Benchmarking Performance

For performance testing, I used the same test bed used in our Radeon 6870 review, a mid/high-end Core i5 system that fairly represents the typical system of most PC gamers. In addition to benchmarking the GTX 580, we included a GTX 460 into the mix-- the card I believe many gamers may be upgrading from if they're considering the GTX 580. Unfortunately, we don't have a GTX 480 in house for a direct comparison, though I'll update our charts when we're able to get one. The other important comparisons are with the Radeon 5870 and 5970 video cards, AMD's current top two performers. Note that the 5970 is a dual-GPU card.
 

Test Bench Specs

CPU: Intel Core i5 750 @ 2.67GHz  
Motherboard: Asus P55  
RAM: 4GB DDR 3  
HD: Seagate 7200.12 1TB @7200RPM  
Monitor resolution: 1920x1080
 

Individual Game Benchmark Settings

Crysis:  DX10 mode, all settings Very High, using the built-in GPU Benchmark.
Dirt 2:  DX11 mode, all settings set to Ultra, using the in-menu benchmark.
Far Cry 2:  Both Action and Ranch (Medium) benchmarks, DX10 mode, all settings maxed out.
STALKER Call of Prypiat:  Stand-alone benchmark demo, all settings maxed out, DX11 mode.
 
We've added two new benchmarks to our suite, which are good indicators of next-gen DX11 gaming. Metro 2033 is very taxing benchmark, while H.A.W.X. 2 is more representative of a next-gen console port that will run fine on most mid-range system but still employs tessellation. I ran both of these benchmarks with tessellation on and off. 

Metro 2033: DX11 and DX10 modes, all settings maxed out, using the built-in GPU benchmark.
H.A.W.X. 2: DX11 and DX9 modes, all settings maxed out, 4xAA, using the stand-alone benchmark demo.
 
Game
GeForce
GTX 580
GeForce 
GTX 460
Radeon 5870
Radeon 5850 1GB
Radeon 5970
 Crysis (no AA)
44.67
27.15
39.3435.1
52.75
 Crysis (8x AA)
40.06
22.43
32.18
29.01
46.92
 Dirt 2 (no AA)
97
76.9
78.8
71.9
100.9
 Dirt 2 (8x AA)
88.9
65.1
69.5
64.9
94
 Far Cry 2 Action (no AA)
44.45
44.06
 62.9762.21
59.72
 Far Cry 2 Action (8x AA)
46.54
46.14
61.86
58.3
57.6
Far Cry 2 Ranch (no AA)
97.29
81.43
96.94
89.1
109.84
Far Cry 2 Ranch (8x AA)
87.54
62.48
66.6
61.45
92.24
Stalker CoP (no AA)
105.4
67.47
92.6
82.7
123.4
Stalker CoP (4x AA)
82.77
49
54
48.9
98.5
 
 
GameGeForce GTX 580
GeForce GTX 460
 Metro 2033 DX 11, No AA
 31.67 17.67
 Metro 2033 DX 11, 4x AA
 26.33 14.67
 Metro 2033 DX 10, 4x AA
Tessellation Off
 36.67 20
 H.A.W.X. 2 DX 11
 149 106
 H.A.W.X. 2 DX 9
 Tessellation Off
 149138
All scores measured in average frames per second.
 

Performance Analysis


 
The GTX 580 couldn't match the performance of the Radeon 5970, though. We asked Nvidia whether the new core would allow for a future dual-GPU card using the GF110, but they couldn't say at this time. My bet is that we'll see a slightly-clocked down GTX 570 model in the future, which may then be used for a dual-GPU part, if at all.
 

The Single-GPU Crown

Nvidia isn't reclaiming the GPU crown with the GTX 580, since the GTX 480 was still faster than the Radeon 5870 in most tests. What this video card release does is make this top performance easier to stomach, since you don't have to worry as much about noise, heat, and power. At $500, this is still the most expensive single-GPU card on the market, but you do get what you pay for. AMD's decision not to launch their 6xxx series GPUs with the high-end has given Nvidia a chance to grab back some of market share (and faith) lost by the GTX 480's blunder, but don't expect AMD to let this stand. If you're brand-agnostic, my recommendation is to wait until AMD reveals their high-end part (and after we benchmark it) before making any purchasing decision. But if you're a devoted Nvidia fan or can't live without PhysX or CUDA apps, the GTX 580 is a bright addition to the GeForce family.