Date: June 25, 2008 - Author: Rob Williams
Proper competition from AMD in the mid-range scheme of things might have taken a while to happen, but it does happen with the HD 4000 series. We are taking a look at the smaller of the two new models, which offers exceptional performance for the price of $200.
Right this moment, the GPU industry is at one of its most exciting points in history. Both AMD and NVIDIA, and even Intel, are going head-to-head-to-head with different ideas of what makes a GPU great. Well, for most enthusiasts, being fast is what makes a GPU great, and right now, AMD has a reason to gloat.
It's been quite a while since AMD has released a part to equal or at least come close to one of NVIDIA's higher-end offerings, but today marks the day when that happens. The HD 4850 and HD 4870 are proper mid-range cards, and right now, they happen to be the fastest single-GPU cards out there - at least out of those that are commercially available.
NVIDIA has their upgraded 9800 GTX, the 9800 GTX+, right around the corner, and once that's released, the situations might change a little bit. It's supposed to be priced above the $200 point of the HD 4850, but is supposed to show consistent improvements all-around. Initial results show consistent 15% performance increases, and for $30, that's not so bad.
It will all come down to pricing, as usual, but that's a good thing. Both companies will fight for your dollars, and as a result, you will wind up with a product that feels like it's worth a lot more than what you paid for it.
AMD haven't released a light upgrade, but rather a significant one with the HD 4000 series. The transistor count has been bumped up to 956 million, from 666 million, still based on the 55nm process. Both of the new cards, along with the future models, support DX 10.1 along with PCI-E 2.0. One new feature lays with the HD 4870, GDDR5. These chips are still fresh and come direct from Qimonda.
Arguably the most significant upgrade is the bump in stream processors. While the HD 3870 / 3850 both featured 320, the new cards feature an incredible 800. This is likely most responsible for the transistor increase.
The HD 4870 clocks in at 750MHz, while the HD 4850 is a bit more modest at 625 MHz. Even with the lower clock speed, though, the card is capable of 1.0 TeraFLOPS worth of math crunching ability. That makes AMD the first, and boy have they been touting the fact. Now we just need a reason on our home computers to exploit such power. We can't just sit around and crunch advanced math problems all day!
Model |
Core MHz |
Mem MHz |
Memory |
Bus Width |
Stream Proc. |
| HD 3870 X2 | 825 |
900 |
2x512MB |
256-bit |
320 |
| HD 3850 X2 | 666 |
828 |
2x512MB |
256-bit |
320 |
| HD 4870 | 750 |
900 |
512MB |
256-bit |
800 |
| HD 4850 | 625 |
993 |
512MB |
256-bit |
800 |
| HD 3870 | 775 |
900 |
512MB |
256-bit |
320 |
| HD 3850 | 666 |
828 |
512MB |
256-bit |
320 |
| HD 3650 | 725 |
800 |
256,512MB |
128-bit |
120 |
| HD 3450 | 600 |
500 |
256MB |
64-bit |
40 |
All the stats aside, what matters is that these cards are fast, and at $200 for the HD 4850 and $299 for the HD 4870, they almost seem like relative steals.
While we aren't taking a look at the HD 4870 today, we are taking the ASUS EAH4850 for a spin. Their card, like most of the others, follow the reference design perfectly, so you can expect our results to closely mimic any other HD 4850 out there. What the purchase will come down to is whatever heatsink you like most, or the warranties that are available.
The card features dual DVI ports, which has pretty well been the standard for a while, and also a TV-Out. For those who want HDMI output, ASUS includes such an adapter in the box.
This card is designed to compete head-to-head with NVIDIA's 9800 GTX, but I already see one killer advantage of the HD 4850 - one 6-pin power connector. In our review of the 9800 GTX, my prime complaint was the fact that it required two separate power connectors, when it really wasn't needed.
There's a lot going on in the back, with a total of twelve screws that need to be removed in order to see what's underneath the cooler.
As for accessories, ASUS has included an adapter to convert a molex power connector into a PCI-E 6-Pin, TV-Out/Component, TV-Out RCA, DVI-to-HDMI adapter, DVI-to-VGA adapter and also a Crossfire bridge. They've pretty much covered all of the bases here.
On the next page, we'll tackle our testing methodology. If you haven't read one of our GPU reviews before, we highly recommend you read through as we conduct testing differently than most other sites.
Regardless of the operating system or product being reviewed, there are a few conditions that are met prior to testing to assure we receive accurate, repeatable results.
Below is our testing machine, which remains untouched throughout all testing except for the graphics card. AMD didn't coincide the HD 4000 series' launch with a new Catalyst release, but rather released a 'hotfix' driver as soon as the cards began showing up in retailers. We are using that driver for all testing, but will retest when the Catalyst 8.7 releases, to see if performance at all improves.
In previous GPU reviews, we've used Windows XP Professional due to its stability (when compared to Vista), but as Vista becomes increasingly popular and the choice for many, it makes sense for us to make the switch as well. We choose to use the 64-Bit version of the OS due to it being the logical choice for gamers who want to use more than 2GB of RAM in their machine.
Depending on the graphic card being reviewed, we split up models into two different categories: Low-End to Mid-Range and Mid-Range to High-End. The former will see the GPUs tested using 1280x1024 and 1680x1050 resolutions, since those are the most common resolutions for gamers looking to purchase a GPU in that price-range.
For our Mid-Range to High-End category, we test GPUs at 1680x1050, 1920x1200 and also 2560x1600 to better reflect the resolutions for those looking for a solid GPU offering.
We do not use time demos in our reviews except where necessary, and in the case of our current GPU reviews, the only game to be subject to a time demo is Enemy Territory: Quake Wars. This is due to that game disallowing greater than 60FPS without the use of a time demo. But since the game is a popular choice for multiplayer gamers, it should be included in some form or another.
Manual Benchmarks
In an attempt to deliver "real-world" results, all games except the above mentioned title are played through manually, with the average FPS recorded with the help of FRAPS 2.9.4. In our personal tests, we have found that manual benchmarks are the best way to deliver accurate results, since time demos rely heavily on the CPU.
In order to deliver the best results, each title we choose is explored to find the best possible level for our benchmarking. Once a level is chosen, we play through in order to find the best route, and then in future runs, we stick to that route as close as possible. We are not robots, so we cannot make sure that each run is identical, but they will never be far off from each other. As we see in our results, scaling is good, so we are confident that our methodology is a good one.
1680x1050 |
1920x1200 |
2560x1600 |
  |
 |
 |
1680x1050 |
1920x1200 |
2560x1600 |
 |
 |
 |
1680x1050 |
1920x1200 |
2560x1600 |
 |
 |
 |
1680x1050 |
1920x1200 |
2560x1600 |
  |
 |
 |
1680x1050 |
1920x1200 |
2560x1600 |
 |
 |
 |
1680x1050 |
1920x1200 |
2560x1600 |
  |
 |
 |
1680x1050 |
1920x1200 |
  |
 |
1680x1050 |
1920x1200 |
2560x1600 |
  |
  |
 |
Note that the reason we do not test Need for Speed at 2560x1600 is because it's a resolution not supported by the game. EA tends to be a little slow when it comes to supporting high-end hardware.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
It's not often that a game comes along that truly pushes our hardware to the utmost limit. Crysis is one of those few games, and that will be the case for at least the next year. Don't believe me? Boot up your top-end machine, max out your resolution and set the graphics to "Very High". I guarantee tears will be shed within a few seconds of loading a level.
The level we chose here is Onslaught, also known as level five. We begin out in a tunnel, but what's important is that we are in control of a tank. What could be more fun? Our run through consists of leaving the tunnel and hitting the other side of the battlefield, killing six or seven enemy tanks along the way.
It goes without saying that any level in Crysis would make for a great benchmark, but this one in particular is gorgeous. Using the "Medium" settings, the game looks spectacular and is playable on all of our graphic cards, so we stick with it. Throughout the level, there is much foliage and trees and also large view-distances. Explosions from the tanks is also a visual treat, making this one level I don't mind playing over and over, and over.
Settings: Due to the intensiveness of the game, no AA is used at any resolution, and the secondary settings are all left to Medium. Please also note that our HD 3870X2 results are outdated, as they have not changed since our original review. Newer drivers will improve the scores, and we are in the process of re-benchmarking that card through all the tests.
These results impressed me, only because I expected to see the HD 4850 exceed the performance of the 9800 GTX, but it didn't happen. At our highest resolution, both performed exceptionally close, however.
Please note that the ASUS EN8800GTS 512 card used is pre-overclocked, and as a result it's faster than a stock 9800 GTX. Seeing it faster than the HD 4850 and 9800 GTX will be a common theme throughout all of our results.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
While Crysis has the ability to bring any system to its knees with reasonable graphic settings, Call of Duty 4 is a title that looks great no matter what setting you choose, even if you have it running well! It's also one of the few games on the market that will benefit from having more than one core in your machine, as well.
The level chosen here is The Bog, for the simple fact that it's incredibly intensive on the system. Though it takes place at night, there is more gunfire, explosions and specular lighting than you can shake an assault rifle at.
Our run consists of proceeding through the level to a point where we are about to leave a building we entered a minute before, after killing off a slew of enemies. The entire run-through takes about four minutes on average.
Settings: High details are used overall throughout all tests, although 4x AA is used for our 1920x1200 setting. That AA is removed in our 2560x1600. As we can see in the graphs below, both of those settings are quite similar in performance.
We heard rumors before the HD 4850 launched that its performance was to equal or exceed the 9800 GTX, and they couldn't have been more true. At our 1680x1050 setting, the HD 4850 proved just 0.107FPS slower on average. Ouch, it hurts to contemplate that small of a difference...
Moving our resolutions skyward changed things a little bit. At our 1920x1200 resolution, it took the top spot, surprisingly beating out the HD 3870 X2. Things changed at 2560x1600, as dual-GPU setups thrive on ultra-high resolutions, especially 2560.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
If there is one game in our line-up that most everyone has played at some point, it would be Half-Life 2. The most recent release is Episode Two, a game that took far too long to see the light of day. But despite that, it proved to be worth the wait as it delivered more of what fans loved.
We are using the Silo level for our testing, which is a level most people who haven't even played the game know about, thanks to Valves inclusion of it in their Episode Two trailers during the year before its release. During our gameplay, we shoot down a total of three Striders (their locations are identical with each run, since we are running a saved game file) and a barn is blown to smithereens.
Overall it's a great level, but the Strider's minions can prove a pain in the rear at times - most notably when they headbutt you. Nothing a little flying log won't solve, however! This levels graphics consist mostly of open fields and trees, although there is a few explosions in the process as well, such as when you blow the Striders apart with the help of the Magnusson Device.
Settings: High graphic settings are used throughout all three resolutions, with 4x AA and 8xAF.
Like the Crysis results, these ones here surprised me. I expected to see much better performance out of the HD 4850, but to say that is somewhat undermining what the card is capable of. Compared to the 9800 GTX that it's competing with, it still came quite close, but still fell short by 2 or 3 FPS.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
Western FPS games are not common, so when one hits, people notice. Luckily for FPS fans, Call of Juarez delivered great graphics, solid gameplay and a very high difficulty. It's a great game to benchmark due to its ability to run in DX10 mode, under Windows Vista. This mode is far more demanding than the DX9 mode, but the results are better.
We take the role of Billy Candle in the level we chose, which is rather simple in concept. We begin out at the end of a linear path that we must follow in order to reach a ravine that we must cross.
The goal of the level is to sneak through a farm and ride off with a horse in order to make the jump, but since that process takes far too long, our run through consists of following the exact same path each time, which ends up on the opposite side of the farm near an edge with water below.
Settings: Very high graphic settings are used here, although AA is never used. The fact that the game uses DX10 is enough to drag performance down.
There's no clear-cut winner here, but the HD 4850 didn't win any of the contests, except the last one, where it was only out-performed by the HD 3870 X2 and pre-overclocked 8800 GTS 512 card.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
Post-apocalyptic FPS games have been done over and over, but S.T.A.L.K.E.R. Shadow of Chernobyl was unique in many ways. First was the fact that the story was loosely based off of a real-life tragedy, the Chernobyl nuclear plant explosion, with the player starting out post-disaster working to survive in the now very brutal world.
One of the areas where the game excelled was with the depth. It was an open world with non-linear gameplay. AI was not top-rate, but reacted in a mostly realistic way, so it's pretty much impossible to just stroll through the game and not expect to die. Coupled with the ability to keep an inventory and sell artifacts you find along your journey makes this game an immersive experience.
The level we use for our testing is a "Thumb Drive" mission that occurs earlier in the game. The premise is simple... walk into a small camp that's being inhabited by enemy Stalkers, wipe them out and go deliver a thumb drive to a lone Stalker huddled around a campfire. The entire quest takes between four and five minutes from our starting point.
Settings: Static lighting and medium quality is used for our lowest resolution here, while 1920 and 2560 use full dynamic lighting along with high quality settings.
The 1680x1050 result here struck me a little, because 293FPS at that resolution is rather low for cards of that calibre. It might sound a bit ridiculous to say that 300FPS is low, because in reality, it's far from it, but compared to cards on the same level, it was held back.
Things of course change as we increase our resolution, where the cards put on almost the same level as the higher-end 9-series cards from NVIDIA.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
The Unreal series has always been one that's pushed graphics to the next level. Surprisingly, though, as the graphics improve, the game still remains playable on a reasonable machine, with good FPS. How often is that the case?
"Gateway" is our level of choice for a few different reasons. The first and most notable is the fact that it's a great level, and chock-full of eye-candy. The entire level consists of three different areas that can be accessed through portals, or "gateways". The area we begin out in is a snow-filled wonderland, similar to Lost Planet's winter levels, with a futuristic city and waterfall area also being accessible.
Settings: All in-game settings are maxed out, with physics and smooth frame rate disabled.
UTIII is a great example of a game that the HD 4850 takes full advantage of. It proved the fastest in both our 1680 and 1920 tests, but came an inch behind our 2560 run. The benefits of a dual-GPU setup can be seen here also, with the HD 3870 X2.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
Electronic Arts is one of the largest game publishers in the world, and because of that, they have plenty of fans and plenty of enemies. Even if you don't like them, it's hard to dispute the fact that many of their games are solid, one being anything from the Need for Speed series.
"Pro Street" received rather poor reviews upon launch, and for mostly good reason. It removes the freedom of being able to explore a city at your leisure, which to many, is a huge step backwards. But despite that fact, it's still a great game if you enjoy the series and want an offering that's a little more realistic than previous versions (in terms of money and damage).
Our run through consists of racing through two laps at the Chicago Airfield, something that takes about three and a half minutes to accomplish from the moment we begin recording frames. The beginning of each race shows an automated camera fly-by over the cars in the race - we begin recording our FPS as soon as this clip begins.
Settings: Our lowest resolution uses fully default settings, while the 1920 resolution ups the AA to 4x and enables Anisotropic texture filtering.
These results surprised me, but in a good way. Most ATI cards I've tested in recent memory haven't run that well in this game. Take the HD 3870 X2, for example. It should conquer these charts, but didn't. The HD 4850 was a joy to use, though. The game ran very smooth and had enough brawn to take the lead in our 1920x1200 run.
Each graph for our benchmarking results are labeled with the resolution that the game was played at, while omitting secondary settings such as Anti-Aliasing, Anisotropic Filtering, texture quality, et cetera. To view all specific settings that we used, please refer to our testing methodology page, where screenshots show the exact settings used.
The last game we will be using in our benchmarks is ET: Quake Wars. This is also the only game in our testing that's executed as a time demo, as opposed to the manual play through like the rest of our games. The reason for this is twofold.
The first reason is that we like to include at least one time demo, despite its CPU-boundedness, in order to see how our cards scale when run in such a situation. The second is the fact that this game caps its FPS at 60, except during time demos.
Our time demo takes place in the Area 22 level, with the main goal to destroy the jamming generator. The actual play through took around five minutes, but the time demo goes far quicker, as is the case with most time demos.
Settings: Maxed settings are used here for the most part. Our 1680 resolution uses 2x AA while 1920 and 2560 use 4x.
I've said it before, but I'll say it again... the results here struck me as odd. In all of our other tests, the HD 4850 either sat right behind the 9800 GTX, or came inches ahead. Here, the card flies past the rest, minus the dual-GPU card in our 2560x1600 run.
It's a little strange. The only test where the HD 4850 truly shines is with our lone timedemo. Luckily, our 3DMark Vantage tests are up next, and since they are of a timedemo nature as well, we can see if similar results are to be had.
3DMark doesn't need to be explained to most anyone, because if you've been benchmarking or PC gaming for a while, you have no doubt heard of Futuremark and their tools. Vantage is the newest of the bunch, and its tests are as hardcore as they come. The benchmark properly stresses a GPU, and spits out an overall score for you to munch on.
The overall use of these scores is constantly debated, because real gameplay matters far more than canned benchmarks. However, they are still fun to use for the sake of competition. In no way should they be the sole factor of your GPU purchasing decision, however.
Hmm. As we saw on the previous page with ET: QW, the HD 4850 proved superior. This was not the theme throughout the entire review, however. Here again we see the same thing with our Vantage results. It seems this card can keep right up to NVIDIA's 9800 GTX in real-world gameplay, but excels in timedemos and canned benchmarks.
I can't verify right now that this is always going to be the case, but I plan to do some tests to make sure that all results are constant and accurate. But as it stands right now, it does seem that the HD 4850 showcases better performance in timedemo-type benchmarks than it does in real-world testing.
In testing power consumption for our graphic cards, the system components are kept consistent to help keep accurate results. To capture wattage, a Kill-a-Watt is used. It is plugged straight into the wall and the PSU is plugged in directly to it. After the computer boots into Windows and is left idle for five minutes, the idle wattage is captured.
To capture the average, a run of 3DMark 2006 is run while keeping an eye on the voltage for the first two minutes. I record the value that the Kill-a-Watt reports the majority of the time. Sometimes the wattage might go higher, but scale right back down, and vice versa.
You have got to love this! The HD 4850 uses less power at load than its predecessor, the HD 3850, yet is far faster. Another interesting comparison is the HD 4850 to 9800 GTX. The latter uses 32W more at full load and 12W more at idle. Not significant, but differences none-the-less.
It goes without saying, although the HD 4850 didn't always surpass the 9800 GTX's performance, it's still right up there, and thanks to its price, it's still the best choice for a new mid-range GPU at around the $200 mark. AMD has a winner, any way you look at it.
What I did find interesting was the fact that the card performed far better in canned benchmarks than in our real-world tests, though. Out of the seven manually-played games, the HD 4850 didn't dominate any. Rather, its performance either met the 9800 GTX, or came close. In a few instances, such as our games run at higher resolutions (namely 2560x1600), we did see increases. But it wasn't congruent across the chart.
Instead, it was with our lone timedemo (ET: QW) and 3DMark Vantage where the real differences were seen. The HD 4850 didn't just beat the competition, it obliterated it in some cases. So I'm led to believe the card performs better in canned benchmarks than in real-world, but take that as you will. Aside from that odd fact, the card still offers killer performance, and that's what matters.
Did I mention that the HD 4850 is one heck of a hot card? In the picture above, you see the result of stressing the GPU with RTHDRIBL. The diode is pushed into the cooler to get as close to the GPU core as possible. I estimate the diode was no further than half of an inch away.
What this means is that you may as well forget about overclocking, unless you have some killer airflow or after-market cooling. The reason I didn't include any overclocking reports is because any overclock that I found reasonable caused the card to overheat, which resulted in even worse performance.
This is the reason we don't see pre-overclocked cards right now. I've been told by two companies that they are unsure if they will even be able to release the cards they want to, because the overclocks they want to hit, just can't be done without overheating. Rumor has it that AMD will be releasing a revision of the core, however, which should allow higher clocks to be put in place. Those might even happen as early as three weeks to a month, so stay tuned.
I might have been focusing a lot on the bad the past few paragraphs, but the fact remains, this card is amazing and is the huge step-up that AMD needed. Before it came out, the 9800 GTX was still priced at around $300, but now it's closer to $235. Even at that price, the HD 4850 is still a better value, if what you are looking for is the best bang for your buck.
The question now is whether you care about overclocking. If you do, the 9800 GTX might be the smarter purchase. As we saw in our results, our pre-overclocked 8800 GTS 512 out-performed the 9800 GTX in almost all of the tests, and that overclock is nowhere near top-end. So overclockers might want to look that way instead or wait until the HD 4000 series revisions come out.
If you have a comment you wish to make on this review, feel free to head on into our forums! There is no need to register in order to reply to such threads.
Copyright © 2005-2008 Techgage Networks Inc. - All Rights Reserved.