Original Link: http://www.anandtech.com/show/2584
AMD's Radeon HD 4870 X2 - Testing the Multi-GPU Watersby Anand Lal Shimpi & Derek Wilson on August 12, 2008 12:00 AM EST
- Posted in
Today is all about the Radeon HD 4870 X2, the same card we previewed last month but AMD is quietly announcing a few other products alongside it. The 4870 X2, internally referred to as R700, is a pair of RV770 GPUs on a single card - effectively a single-card, Radeon HD 4870 CrossFire (hence the X2 moniker). Like previous X2 cards, the 4870 X2 appears to the user and the driver as a single card and all of the CrossFire magic happens behind the scenes.
|ATI Radeon HD 4870 X2||ATI Radeon HD 4870||ATI Radeon HD 4850|
|Stream Processors||800 x 2||800||800|
|Texture Units||40 x 2||40||40|
|ROPs||16 x 2||16||16|
|Memory Clock||900MHz (3600MHz data rate) GDDR5||900MHz (3600MHz data rate) GDDR5||993MHz (1986MHz data rate) GDDR3|
|Memory Bus Width||256-bit x 2||256-bit||256-bit|
|Frame Buffer||1GB x 2||512MB||512MB|
|Transistor Count||956M x 2||956M||956M|
|Manufacturing Process||TSMC 55nm||TSMC 55nm||TSMC 55nm|
The benefit of single-card CrossFire is of course that you can use this single card on any platform, not just ones that explicitly support CF. Since CrossFire is supported on both Intel chipsets and AMD chipsets, it's a bit more flexible than SLI and the need for single-card CF isn't nearly as great as the need for single-card SLI.
Unlike most single-card multi-GPU solutions, the 4870 X2 is literally two Radeon HD 4870s on a single card. The clock speeds, both core and memory, are identical and this thing should perform like a pair of 4870s (which is pretty quick if you have forgotten). The only difference here is that while the standard Radeon HD 4870 ships with 512MB of GDDR5 memory, each RV770 on a X2 gets a full 1GB of GDDR5 for a total of 2GB per card.
...which leads us nicely into some of AMD's other products that will be coming out in the next month or so. There will be 1GB versions of both the Radeon HD 4870 and Radeon HD 4850.
Then at $399 we'll see a Radeon HD 4850 X2, which as you can probably guess is a pair of Radeon HD 4850 GPUs on a single card, but with 2GB of GDDR3 and not GDDR5 like the 4870 X2. As interesting as all of these cards are, we only have the 4870 X2 for you today, the rest will have to wait for another time. But it is worth noting that if you are interested in buying a Radeon HD 4870/4850 and keeping it for a while, you may want to wait for the 1GB versions as they should give you a bit more longevity.
Enough with being distracted by AMD's product lineup, let's talk about the competition.
NVIDIA Strikes Back: The GTX Gets a Dose of Reality
NVIDIA was still living in the days of G80 when it launched its 1.4 billion transistor GT200 GPU and the GeForce GTX 280/260 that were based on it. Not only did NVIDIA's own GeForce 9800 GX2 outperform the GTX 280 at a lower price, but once AMD launched its Radeon HD 4800 series it became very clear that NVIDIA's pricing was completely out of whack. NVIDIA was pricing its GPUs for a reality that just didn't exist.
The first step to get things back in line was to drop the price of the GeForce 9800 GTX, which NVIDIA did. Next up were the new GTX cards, the GTX 280 now sells for $450 and the GTX 260 is a $299 part. In the conclusion of our Radeon HD 4800 launch article we wrote:
"The fact of the matter is that by NVIDIA's standards, the 4870 should be priced at $400 and the 4850 should be around $250."
It looks like NVIDIA's standards changed, largely thanks to AMD, and now the key players in NVIDIA's lineup are priced more realistically. Today we'll take a look at how the landscape has been reshaped as a result of NVIDIA's pricecuts. At the same time, AMD's literally hot GPUs have seen their prices fall; the Radeon HD 4870 is now a $270 - $280 GPU, slightly down from $299 and the Radeon HD 4850 is a $170 - $180 card. These are very slight changes in price, but at least they are in the right direction.
AMD Prices the Radeon HD 4870 X2
When we previewed the Radeon HD 4870 X2 we weren't given a target pricepoint, we just knew that it'd be more than $500. Today we have a price: $549.
At $549 the X2 isn't exactly a bargain, it's slightly cheaper than two Radeon HD 4870s but you don't need a motherboard with two PCIe x16 slots to use it, which helps lower overall system costs. With NVIDIA's GeForce GTX 280 price drops, the Radeon HD 4870 X2 is now the most expensive current GPU on the market - pretty impressive for a company that swore off building huge GPUs.
The competing product from NVIDIA is, well, there isn't exactly one. NVIDIA doesn't have a single-card multi-GPU GT200 product, so we have to rely on comparing the 4870 X2 to the GeForce GTX 280 (priced at $450) as well as the GeForce GTX 260 in SLI (priced at $300 x 2).
These Aren't the Sideports You're Looking For
Remember this diagram from the Radeon HD 4850/4870 review?
I do. It was one of the last block diagrams I drew for that article, and I did it at the very last minute and wasn't really happy with the final outcome. But it was necessary because of that little red box labeled CrossFire Sideport.
AMD made a huge deal out of making sure we knew about the CrossFire Sideport, promising that it meant something special for single-card, multi-GPU configurations. It also made sense that AMD would do something like this, after all the whole point of AMD's small-die strategy is to exploit the benefits of pairing multiple small GPUs. It's supposed to be more efficient than designing a single large GPU and if you're going to build your entire GPU strategy around it, you had better design your chips from the start to be used in multi-GPU environments - even more so than your competitors.
AMD wouldn't tell us much initially about the CrossFire Sideport other than it meant some very special things for CrossFire performance. We were intrigued but before we could ever get excited AMD let us know that its beloved Sideport didn't work. Here's how it would work if it were enabled:
The CrossFire Sideport is simply another high bandwidth link between the GPUs. Data can be sent between them via a PCIe switch on the board, or via the Sideport. The two aren't mutually exclusive, using the Sideport doubles the amount of GPU-to-GPU bandwidth on a single Radeon HD 4870 X2. So why disable it?
According to AMD the performance impact is negligible, while average frame rates don't see a gain every now and then you'll see a boost in minimum frame rates. There's also an issue where power consumption could go up enough that you'd run out of power on the two PCIe power connectors on the board. Board manufacturers also have to lay out the additional lanes on the graphics card connecting the two GPUs, which does increase board costs (although ever so slightly).
AMD decided that since there's relatively no performance increase yet there's an increase in power consumption and board costs that it would make more sense to leave the feature disabled.
The reference 4870 X2 design includes hardware support for the CrossFire Sideport, assuming AMD would ever want to enable it via a software update. However, there's no hardware requirement that the GPU-to-GPU connection is included on partner designs. My concern is that in an effort to reduce costs we'll see some X2s ship without the Sideport traces laid out on the PCB, and then if AMD happens to enable the feature in its drivers later on some X2 users will be left in the dark.
I pushed AMD for a firm commitment on how it was going to handle future support for Sideport and honestly, right now, it's looking like the feature will never get enabled. AMD should have never mentioned that it ever existed, especially if there was a good chance that it wouldn't be enabled. AMD (or more specifically ATI) does have a history of making a big deal of GPU features that never get used (Truform anyone?), so it's not too unexpected but still annoying.
The lack of anything special on the 4870 X2 to make the two GPUs work better together is bothersome. You would expect a company who has built its GPU philosophy on going after the high end market with multi-GPU configurations to have done something more than NVIDIA when it comes to actually shipping a multi-GPU card. AMD insists that a unified frame buffer is coming, it just needs to make economic sense first. The concern here is that NVIDIA could just as easily adopt AMD's small-die strategy going forward if AMD isn't investing more R&D dollars into enabling multi-GPU specific features than NVIDIA.
The lack of CrossFire Sideport support or any other AMD-only multi-GPU specific features reaffirms what we said in our Radeon HD 4800 launch article: AMD and NVIDIA don't really have different GPU strategies, they simply target different markets with their baseline GPU designs. NVIDIA aims at the $400 - $600 market while AMD shoots for the $200 - $300 market. And both companies have similar multi-GPU strategies, AMD simply needs to rely on its more.
General Performance at 2560x1600
We've already established that in games with CrossFire support, the Radeon HD 4870 X2 is the grand poobah of gaming performance. Along with such a title comes a general requirement: if you're dropping over $500 on a graphics card on a somewhat regular basis, you had better have a good monitor - one of many 30" displays comes to mind. Without a monitor that can handle 2560x1600, especially with 2x 4870 X2 cards in CrossFire, all that hard earned money spent on graphics hardware is just wasted.
Since the target resolution here is 2560 x 1600, let's see how the 4870 X2 stacks up in our suite of games at this resolution:
The 4870 X2 holds its own against its major competition from NVIDIA, the GTX 260 SLI, in 4 of our 7 benchmarks, while the 4870 X2 CrossFire leads the pack in all but one game (Assassin's Creed). It is always tough to pick the games we want to test, as the games we pick end up deciding what we think of performance. We try to pick games that are both interesting to the community and/or show interesting performance differences between hardware. In this case, what we have here is a pretty good picture of the general case: the 4870 X2 and GTX 260 SLI are pretty well matched.
Of course, the GTX 260 SLI option is two cards which requires an NVIDIA motherboard. This puts it at a bit of a disadvantage to a single card solution that is platform agnostic. And we also have the fact that at the target resolution of 4870 X2 CrossFire AMD does have the most powerful option that fits into two slots. These are very important factors for AMD, but as we've seen before 4-way solutions aren't the value option (you don't get the same return on your money as you might with other solutions even if you do get high performance).
Taking into account the price drops, we also see the single GPU arena looking very competitive with the GTX 260 and HD 4870 about on par. In these tests, the HD 4870 looks to have an advantage, but again we could add a few more benchmarks and see the GTX 260 do better. This is what we like to see: real competition. Of course, the tough question to answer here is whether or not the GTX 280 is worth 1.5x either the GTX 260 or the 4870.
Age of Conan Performance Scaling
This benchmark is one we first looked at with our R700 preview, and it does seem to favor AMD's architecture. The 4870 is on par in performance with the GTX 280, and the GTX 260 gets left way behind. When we look at SLI and CrossFire we see the same thing: CrossFire wins out here and thus the 4870 X2 is king.
4-way CrossFire isn't so fortunate, as we see no scaling beyond 2 GPUs for AMD's cards. Regardless, right now, at any resolution an AMD 4800 series part is the way to go if Age of Conan is your game. This may change when the DX10 version of the game makes its way out, but we'll have to wait to find out.
Crysis Performance Scaling
Ahh Crysis. Our familiar friend. Not the greatest game in the world, but it looks really good and still absolutely kills graphics hardware.
In this case things are a little strange. We see the 4870 X2 and 2x 4870 X2 CrossFire solutions very system limited at below 2560x1600. The NVIDIA SLI options provide a marked performance advantage at 1920x1200 and lower, but drop off very steeply at 2560.
The major factor in a purchasing decision here should be monitor size. If you have a 30" monitor, you might want to consider the 4870 X2 soluiton, but other wise it is a much better bet to stay with a single card (and lower settings) or to go with an NVIDIA SLI system.
Not that anyone would want to build a system just to play Crysis. Right?
The Elder Scrolls IV: Oblivion Performance Scaling
Going a different way than the Crysis test, the 4870 X2 system perform better at lower resolutions than at 2560x1600 compared to the competition. The 4-way CrossFire solution did lead the pack at 2560x1600, but it was fairly bottlenecked at lower resolutions.
These numbers are all very high and playable, and we didn't notice any stuttering issues this time around. Of course, this is a subjective test and we haven't yet taken a quantitative look at it, but our subjective test didn't give us any cause for concern.
We do see NVIDIA hardware with a bit of an advantage, but it still seems like the GTX 280 is ever so slightly over priced giving between a 20% and 40% advantage over the 4870 for a roughly 50% increase in price. As the GTX 260 out performs the 4870 yet costs the same, the value here is even clearer.
Enemy Territory: Quake Wars Performance Scaling
Once again, at all but the highest resolution the 4870 X2 CrossFire is system limited. Of course, at 2560x1600 it does take the lead even in this traditionally NVIDIA and SLI dominated benchmark.
The single 4870 does hold it's own against the GTX 260 and even matches the pace of the GTX 280 until we hit 2560x1600. This makes the 4870 a great value for games based on Id's engine.
SLI scaling still wins out here over CrossFire scaling, and the advantages the single card 4870 had are removed when we look at the 4870 X2 compared to GTX 260 SLI.
Race Driver GRID Performance Scaling
Memory is king in the game at 2560x1600. Our 1GB cards handle the stress the best, and lower framebuffer sizes mean worse performance (especially during loading screens which ran incredibly slow). Our 512MB 4870 cards tanked in CrossFire at 2560x1600, while the 1GB per GPU 4870 X2 delivered wonderful performance. At all other resolutions, the single 4870 performed really well, but we will be interested in seeing if the 1GB version bumps performance up another notch here.
This game, like Age of Conan, runs incredibly well on AMD hardware. Both CrossFire and SLI scale very well, with 4-way CrossFire once again delivering a flat line of performance across all resolutions.
The 4870 X2 is an incredible power hog. Yes, it does perform well. Yes it costs a bit less than two 4870 cards. But will we be able to afford the energy bill?
In both idle and load power the 4870 X2 is near the bottom of our chart. The only thing more inane than the 4870 X2 is plugging two of them into one system. Keep in mind that this power measurement is taken running a 3dmark pixel shader test using almost no other resources. In actual game play power draw is much higher as the CPU, memory, and hard drive can come under load at the same time.
Welcome to the kilowatt era. We've been telling you these huge power supplies weren't that necessary until this generation, and we absolutely mean it. If you are thinking about a multi GPU solution involving the latest hardware you are going to want something upwards of 1KW. If you want GTX 280 SLI or 4870 X2 CrossFire you'll want to head on up to the 1200W deparment.
Oh, and, don't forget to turn it off when you aren't gaming.
I've never felt totally comfortable with single-card multi-GPU solutions. While AMD reached new levels of seamless integration with the Radeon HD 3870 X2, there was always the concern that the performance of your X2 would either be chart topping or merely midrange depending on how good AMD's driver team was that month. The same is true for NVIDIA GPUs, most games we test have working SLI profiles but there's always the concern that one won't. It's not such a big deal for us benchmarking, but it is a big deal if you've just plopped down a few hundred dollars and expect top performance across the board.
Perhaps I'm being too paranoid, but the CrossFire Sideport issue highlighted an important, um, issue for me. I keep getting the impression that multi-GPU is great for marketing but not particularly important when it comes to actually investing R&D dollars into design. With every generation, especially from AMD, I expect to see a much more seamless use of multiple GPUs, but instead we're given the same old solution - we rely on software profiles to ensure that multiple GPUs work well in a system rather than having a hardware solution where two GPUs truly appear, behave and act as one to the software. Maybe it's not in the consumer's best interest for the people making the GPUs to be the same people making the chipsets, it's too easy to try and use multi-GPU setups to sell more chipsets when the focus should really be on making multiple GPUs more attractive across the board, and just...work. But I digress.
The Radeon HD 4870 X2 is good, it continues to be the world's fastest single card solution, provided that you're running a game with CrossFire support. AMD's CF support has been quite good in our testing, scaling well in all but Assassin's Creed. Of course, that one is a doubly bitter pill for AMD when combined with the removal of DX10.1 support in the latest patch (which we did test with here). That has nothing to do with CrossFire support of course, but the lack of scaling and the fact that 4xAA has the potential to be free on AMD hardware but isn't really doesn't stack up well in that test.
In addition to being the fastest single card solution, the 4870 X2 in CrossFire is also the fastest 2 card solution at 2560x1600 in every test we ran but one (once again, Assassin's Creed). It is very important to note that 4-way CrossFire was not the fastest solution at lower than 2560x1600 in as many cases. This is generally because there is more overhead associated with 4-way CrossFire which can become the major bottle neck in performance at lower resolution. It isn't that the 4870 X2 in CrossFire is unplayable at lower resolutions, it's just a waste of money.
We do have yet to test 3-way SLI with the newest generation of NVIDIA hardware, and the 3-way GTX 260 may indeed give 2x 4870 X2 cards a run for their money. We also have no doubt that a 3x GTX 280 solution is going to be the highest performing option available (though we lament the fact that anyone would waste so much money on so much unnecessary (at this point in time) power).
For now, AMD and NVIDIA have really put it all in on this generation of hardware. AMD may not have the fastest single GPU, but they have done a good job of really shaking up NVIDIA's initial strategy and forcing them to adapt their pricing to keep up. Right now, the consumer can't go wrong with a current generation solution for less than $300 in either the GTX 260 or the HD 4870. These cards compete really well with each other and gamers will really have to pay attention to which titles they desire greater performance in before they buy.
The GTX 280 is much more reasonable at $450, but you are still paying a premium for the fastest single GPU solution available. In spite of the fact that the price is 150+% of the GTX 260 and the 4870, you just don't get that return in performance. It is faster than the GTX 260, and most of the time it is faster than the 4870 (though there are times when AMD's $300 part outperforms NVIDIA's $450 part). The bottom line is that if you want performance at a level above the $300 price point in this generation, you're going to get less performance per dollar.
When you start pushing up over $450 and into multi-GPU solutions, you do have to be prepared for even more diminished returns on your investment, and the 4870 X2 is no exception. Though it scales well in most cases and leads the pack in terms of single card performance when it scales, there is no gaurantee that scaling will be there, let alone good, in every game you want to play. AMD is putting a lot into this, and you can expect us to keep pushing them to get performance impovements as near to linear as possible with multi-GPU solutions. But until we have shared framebuffers and real cooperation on rendering frames from a multi-GPU solution we just aren't going to see the kind of robust, consistent results most people will expect when spending over $550+ on graphics hardware.