Review: AMD’s HD 6990 and Nvidia’s GTX 590 - Dueling dual GPUs, and AIBs

Posted by Robert Dow, Alex Garovi, and Jon Peddie on April 12th 2011 | Discuss
Categories: Hardware Review
Tags: gpu nvidia amd market review pmark

The dual GPU AIBs are marvelous products. They are awe inspiring, and not for everyone because of their price, but they are the Ferraris of the market in every sense.

Nvidia officially released the GeForce GTX 590 dual GPU AIB 24 March, 2011, just 16 days after AMD introduced their dual GPU AIB the Radeon HD 6990. That's not a big gap in time relative to what these designs represent, and more of a marketing move between the two companies than a technology gap or indication of any problems. Nvidia is the more clever of the two at marketing and in most cases it's the last story you hear that you remember. No doubt that was part of their strategy. Nonetheless, tactics aside, the two new AIBs will be compared by dozens of organizations and individuals, and those comparisons will be published and live on the web; what you say on the web stays on the web—forever.

The exact logic behind AMD's and Nvidia's recent welter of AIB introductions is obvious. One would expect them to start at the top and work their way down, or vice versa, but that's not how it has happened.

New AIB releases based on new architectures
Segment AMD Introduced Nvidia Introduced
Value < $90 HD 6450
HD 6570
HD 6670
Feb. 2, '11
Feb. 2, '11
Jan. 21, '11
GT 430
GT 440
Oct. 13, '10
Feb. 1, '11
Mainstream $90-$149 HD 6850 Oct. 22, '10 GTS 450
GT 550 Ti
Sept. 13, '10
Mar. 15, '11
Performance $150-$249 HD 6870
HD 6950
Oct. 22, '10
Dec. 15, '10
GTX 560 Ti Jan. 25, '11
Enthusiast $150-$249 HD 6970
HD 6990
Dec. 15, '10
Mar. 8, '11
GTX 580
GTX 590
Nov. 9, '10
Mar. 24, '11

AMD has been traditionally generous with their roadmaps and Nvidia hasn't, since their embarrassing delay with Fermi. Nvidia has reverted to generalized macro family views of their product plans. Given the paranoia in both companies, we expect AMD to revert to this same kind of behavior as well. We will then have to rely on our Asian brothers to leak the OEM and ODM presentations to get a view of what the two companies are planning. Part of the problem is that both companies are now at least par­tial­ly, consumer providers, and compan­ies like Apple, and con­sumer goods companies don't offer roadmaps ever since the Osborne Effect entered our lexicon. Semiconductor and component suppliers do offer roadmaps so their customers can plan future products.

COMPARISON OF specifications: Nvidia GTX 590 and AMD HD 6990
AIB
GPU
GTX 590
GF110
HD 6990
Cayman
Core Clock (MHz) 607 850 MHz
Memory Clock (MHz) 853 1250
Memory (GB) 3 4
Driver    
Power (watts) 365 375
Stream Processors   3072
Cuda Cores 1024  
Transistors (B) 6 5.28
Price $ $730 $700
Pmark 33.0 28.9

The Entry-level Value segment AIBs are mostly OEM parts and the ASP is soft so they are difficult to correctly position in the segment list sometimes.

The point is the introduction of the AIBs does not follow a logical progression by segment. This is partially due to yield and binning, but there are other forces at work like OEM product introductions, and CPU releases.

How many GPUs does it take to light up a pixel?

With the in­tro­­duction of the dual GPU AIBs from AMD and Nvid­ia, we tried to find their value vs. per­formance and other factors (like power and noise) and to compare them to dual AIB solutions. We also tried overclocking the high-end AIBs. All the tests were done on an Intel Core i7 x980, 3.33 GHz, with 3 GB RAM, running Windows 7, 64-bit. You can see the results in the accompanying charts.

AMD recently challenged Nvidia to prove its claim that it was the fastest AIB on the market. Nvidia hoped the reviewers would do that and some have. However, fastest at what is the issue in satisfying such a claim and obviously either company would pick the test and conditions to make them look most favorable. That's a cat fight best avoided; no one wins.

We looked at six programs and three screen resolutions to arrive at an average performance score for the AIBs in order to compute the Pmark under the assumption that no one buys an AIB, especially one selling for $700 or more, simply to run one application.

However, the raw data is also interesting and so we show it here too (in the chart labeled FPS performance).

In a few situations Stone Giant wouldn't run. And since 3DMark 11 doesn't give a FPS score we didn't include it, however we did use it and its results are shown in the accompanying chart.

Not too surprising 3DMark 11 provides a similar view which refutes the claims it is not representative of real world tests. In fact the deviation of the FPS 1920x1200 average score to 3DMark 11 Performance was an average of just 0.9%.

Nvidia GTX 590 and AMD Radeon HD 6990 dual GPU AIBs

The dual GPU AIBs are the Ferraris of the market in every sense: marvelous high-performance products, awe inspiring, and not for everyone because of their price.

The dual GPU AIBs require two 8-pin power plugs which limits the number of systems that can run two of them (giving the power of four GPUs). The AMD AIB is about an inch longer than the Nvidia.

Tweaking the tests.

AMD says Stone Giant is a biased test that favors Nvidia, and looking at the scores one could get that impression. If we pull Stone Giant out of the average score dataset and calculate the Pmark we get the results shown in the following chart.

Pulling Stone Giant out of the mix does improve AMD's Pmark score, but Nvidia still has the best score although the differences are narrowed.

Not the best deal?

On a FPS vs. Dollar basis, could two lower class AIBs in Crossfire or SLI actually be a better deal in price, and in performance? Of all the boards and combinations of boards we tested in this series, dual GTX 570s in SLI gave a highest average games score, and dual GTX 580s in SLI gave the highest 3DMark 11 scores. However, when the Pmark is calculated the GTX 590 is the winner.

Two Nvidia GTX 570s cost ~$660, and a GTX 590 sells for ~$730. Two GTX 580s sell for ~ $998. And two GTX 570s draw 438 W, while the GTX 590 only draws 365 W, and dual GTX 580s need 500 W.

Over clocking

Both dual GPU AIBs are capable of over-clocking. How much to over clock is an individual choice and experience since the upper limit varies by production of the GPUs—some will go faster than others. AMD has added a small slider switch on the top of the AIB near the Crossfire tabs that allows one to switch the BIOS for over clocking from 830 MHz to 880 MHz—a six-percent change. We used that delta, 6%, for the Nvidia AIB as well in the over-clocking tests; however, our experience, and this will vary from one user to another, was the Nvidia board had more over-clocking headroom.

However, even though the clocks were increased 6% on each board, the performance results were not the same, as shown illustrated in the accompanying chart.

Where:
Performance = the average of several tests,
Price = any currency (Dollars, Euros, Yen),
Power = the AIB’s wattage under load.
(Performance is measured in two ways on a graphics AIB, by frame rate (fps), or by an arbitrary score as generated by 3Dmark 11.

What do we think?

The Nvidia dual GPU GeForce GTX 590 is the best overall AIB in every parameter we measured it. Compared to other dual GPU AIBs, and dual AIBs, It has the highest performance/watt, the highest performance/dollar, the lowest power consumption, the lowest noise, the biggest gain from over-clocking, and the best Pmark, regular or over-clocked. It's also the most expensive at $730 compared to the AMD Radeon HD 6990 at $695.

These high-end uber-enthusiast AIBs are the dream machines, the aspirational products that everyone wants and a few can have. And even though they are bit expensive, we think the demand for them is going to surprise AMD and Nvidia.

PMARK COMPARISON of dual GPU AIBs standard and over clocked

PMARK WITHOUT Stone Giant

COMPARISON OF several AIBs in various applications

COMPARISON OF AIBs with 3DMark 11

TWO GTX 570s cost less than one GTX 590 and give better results

PERFORMANCE DIFFERENCE from 6% overclocking

Discuss this entry