 | |||
|
|
|
|
|
 |
ATI Radeon 4670 Performance Preview September 09, 2008 Brandon Sandman Bell |
|
 |
|
| |
Summary: Thanks to its 320 shaders ATI's Radeon 4670 is capable of giving the Radeon 3850 a run for its money at significantly lower price tag: just $80. As a result, NVIDIA's countered with price cuts for the 9600 GSO and 9600 GT. See how the 4670 compares to all of these GPUs in this article!
 Radeon 4670 Performance Preview | Page:: ( 1 / 18 ) |
[image]
| <% print_image("01"); %> |
Their RV770-based Radeon HD 4850 and 4870 both deliver excellent performance in their respective price segments. Their arrival forced NVIDIA to slash prices on their entire GeForce lineup, including their flagship GeForce GTX 260 and GTX 280 GPUs. Despite the GeForce price cuts, this hasn’t stopped ATI from taking share from NVIDIA. And in the high-end of the market, ATI’s dual GPU Radeon HD 4870 X2 is the fastest graphics card in the world right now.
But how well does this strategy scale down? Does ATI’s architecture continue to perform well with half of its functional units (including its shaders), removed? That’s what we’re here today to find out!
[image]
| <% print_image("02"); %> | <% print_image("03"); %> | <% print_image("04"); %> |
Finally, a new ATI GPU architecture for the mainstream segment
Based on ATI’s RV730 GPU, the Radeon 4600 series is designed to replace the Radeon 3600 series in ATI’s lineup. Despite their new name, the 3600s weren’t dramatically different than ATI’s older Radeon 2600 line. The Radeon 3650’s RV635 GPU boasted the same 120-shader architecture as the Radeon 2600, with 8 texture units, 4 ROPs, and a 128-bit memory interface, just like Radeon 2600. The big additions ATI incorporated into RV635 were DirectX 10.1 support and a more energy efficient 55-nm manufacturing process. Every where else the Radeon 3600 GPU was quite similar to its predecessor, and as a result, performance between the two was quite comparable. As you can imagine this put ATI at a performance disadvantage in comparison to NVIDIA’s GeForce 8500 and 8600 lines of GPUs.
NVIDIA must have taken note of this, as their GeForce 8500/8600 replacements, the GeForce 9400 GT and 9500 GT are largely based on their respective predecessors; both the GeForce 8500 GT and GeForce 9400 GT sport 16 stream processor architectures, while the 9500 GT and 8600 GT/GTS are all based around a 32-shader architecture. NVIDIA has incorporated a few tweaks to improve clock-for-clock performance, and the GeForce 9400 and 9500 GT are clocked to higher speeds, but NVIDIA didn’t start with a blank sheet of paper when designing these parts. This lack of innovation in the mainstream segment has kept some gamers on a budget from upgrading their trusty graphics cards like the Radeon X1950 Pro and GeForce 7900 GT.
Fortunately this isn’t the case for RV730, it’s an entirely new GPU that’s designed to address the flaws in RV635…
| RV730 Architecture | Page:: ( 2 / 18 ) |
To accomplish this ATI has dramatically reduced the number of shading units inside RV730. While RV770 boasts up to 800 stream processors, just 320 stream processors are found within RV730. Interestingly enough, this is the same number of shaders found in ATI’s previous high-end architecture, RV670/R600.
As a result there are cases where the Radeon 4670 comes awfully close to matching the performance of Radeon 3850, if not outperforming it, but more on this later.
[image]
| <% print_image("05"); %> | <% print_image("06"); %> | <% print_image("07"); %> |
In comparison to RV770, ATI has adjusted the arrangement of the SIMD cores. RV730 ships with half the number of stream processors per SIMD core, just 40 shaders in RV730 per SIMD core versus 80 shaders per core in RV770. RV730 boasts 8 SIMD cores total, resulting in a grand total of 320 stream processors (40 stream processors per SIMD core x 8 SIMD cores). Like RV770, each SIMD core in RV730 has four dedicated texture units bringing the total number of texture units to 32 inside RV730.
If you recall, texturing was a huge shortcoming in ATI’s previous generation architecture. RV670 supported just 16 texture units. So RV730 has two times the number of texture units as ATI’s previous generation high-end architecture, and four times the number of texture units as its direct predecessor, the Radeon 3650. Like RV770, RV730’s texture units have also been redesigned to deliver up to two times the texture cache bandwidth of RV670.
The ROP subsystem has been improved as well. RV730 sports 8 ROPs. That’s two times the number of ROPs as RV635. ATI has also tweaked the ROPs to provide double the AA fill-rate for 32-bit and 64-bit color as well as double the peak rate for depth/stencil operations. These improvements allow the RV730 to perform quite competitively with the Radeon 3850 despite its memory bandwidth disadvantage: 32GB/sec in Radeon 4670 versus 57.6GB/sec in the 3850.
Speaking of the memory interface, we should mention that like other mainstream card offerings it’s 128-bit wide, so no change there in comparison to RV635. However, ATI clocks the memory on the Radeon 4670 up to 1.0GHz (2.0GHz effective), yielding 32GB/sec of peak memory bandwidth. The Radeon 3650 topped out at 25.6GB/sec.
RV730 continues to rely on TSMC’s 55-nm manufacturing process, with the chip containing approximately 514 million transistors. Each RV730 chip sports a die size of just 146 square millimeters with a max TDP of 59W for the Radeon 4670, and 48W for the Radeon 4650.
[image]
| <% print_image("08"); %> |
RV730 SKUs
As we just mentioned, ATI will be offering two different GPUs based off RV730: the Radeon 4670 and the Radeon 4650. Digging deeper, ATI will be offering multiple 4670 and 4650 configurations, with memory sizes ranging from 512MB of GDDR3 graphics memory up to 1GB of DDR3 for the 4670, while 512MB and 1GB SKUs for the 4650 will also be provided. The following chart summarizes ATI’s 4600 series lineup:
| Radeon 4600 Comparison | |||
| Radeon 4670 512MB | Radeon 4670 1GB | Radeon 4650 512MB | |
| # of Stream Processors | 320 | 320 | 320 |
| Core Clock | 750MHz | 750MHz | 600MHz |
| Memory Interface | 128-bit | 128-bit | 128-bit |
| Memory Clock | 1000MHz | 900MHz | 500MHz |
| Memory Type | GDDR3 | DDR3 | DDR2 |
| Memory Bandwidth | 32GB/sec | 28.8GB/sec | 16GB/sec |
| Texture fill-rate | 24 Gigatexels/sec | 24 Gigatexels/sec | 19.2Gigatexels/sec |
| MSRP | $79 | NA | $69 |
| Max Board Power | 59W | 59W | 48W |
| |||
As you can see, ATI will be shipping the 4670 with GDDR3 and conventional DDR3 memory types. We’ve been told that latency between the two memory types should be the same, however AMD uses GDDR3 on the 512MB model because it supports 8-bit burst and also 1GHz DDR3 modules are more expensive. Instead ATI relies on 900MHz DDR3 for their 1GB Radeon 4670 boards.
| Radeon 4670 and GeForce 9600 GSO | Page:: ( 3 / 18 ) |
[image]
| <% print_image("09"); %> | <% print_image("10"); %> |
Board design
With a max board power draw of just 59W, the Radeon 4670 gets all its power from the PCIe interface, no external power connector is needed for operation. The board design of the card itself is quite simple, and the board’s PCB measures just a hair over 6.5”.
For cooling the RV730 GPU, ATI has integrated the same single-slot cooler used on the Radeon 3650. It’s an all-copper heatsink/fan unit with a ducted design that blows cool air over the GPU and out the left end of the duct. A copper plate is used to disperse heat off the GPU and memory modules on the top of the card, while the memory modules on the bottom of the card sit naked and uncooled.
Finally, a variable speed fan is responsible for supplying fresh cool air to the card. The fan runs quite quietly, although those of you concerned about board temps won’t be glad to hear that we observed board temps of 60 degrees Celsius at idle and 76 degrees at load. Fortunately the PCB itself doesn’t get nearly as hot as the 4850, so we don’t think the situation is quite as bad as the 4850 was, but it still wouldn’t hurt if ATI would crank up the fan’s RPMs a little more to keep the GPU cooler.
[image]
| <% print_image("11"); %> | <% print_image("12"); %> |
ATI continues to outfit the reference board design with two dual-link display outputs. HDMI and DisplayPort are also supported by the GPU as well, so conceivably we’ll see cards from select ATI board partners touting these display capabilities at some point in the near future. The Radeon 4600 series also supports integrated 8-channel digital audio natively over DisplayPort and HDMI. With GeForce GPUs a passthru cable must be connected from your graphics card to your sound card in order to pass audio over HDMI, so the Radeon 4600 card eliminates the hassle of connecting that cable.
[image]
| <% print_image("13"); %> | <% print_image("14"); %> |
GeForce 9600 GSO
To counter the Radeon 4670, NVIDIA is dusting off an old GPU we assumed was on its way to the boneyard for retirement: the GeForce 9600 GSO.
Based on NVIDIA’s popular G92 GPU, the GeForce 9600 GSO was originally prepped to replace the GeForce 8800 GS. Thanks to recent price cuts that just kicked in, prices on these cards start at just $75 after mail-in rebate on Newegg, with factory OC’ed cards selling for just $84. Just what is a GeForce 9600 GSO? It’s basically a GeForce 8800 GT/9800 GT with slower clocks, a narrower memory interface, and some of its shaders disabled. First let’s discuss the shader configuration.
Like the original GeForce 8800 GTS from two years ago, two of the GPU’s shader clusters are disabled on the GeForce 9600 GSO, turning off 32 shaders and leaving a total of 96 functional shaders.
Besides its 96-shader architecture, NVIDIA also disables one 64-bit memory controller on the GeForce 9600 GSO. As a result, this leaves the G92 GSO GPU with a 192-bit memory interface to its GDDR3 memory. The board is then outfitted with either 384MB or 768MB of memory.
Besides the narrower memory interface and reduced shader count, we also mentioned reduced clock speeds. In particular the GeForce 9600 GSO is clocked at 550MHz on the graphics core, while its stream processors run at 1375MHz. Finally, NVIDIA has also reduced the memory clock to 800MHz (1.6GHz effective), yielding 38.4GB/sec of peak memory bandwidth.
[image]
| <% print_image("16"); %> |
Fortunately this is where the cuts end for the GeForce 9600 GSO. NVIDIA uses the exact same reference board design and cooling originally used on the GeForce 8800 GT from last year for the 9600 GSO. While the original 8800 GT cooler was notorious for running hot, NVIDIA improved the card’s fan, making it slightly larger than the original cooler. This has a profound impact on temps and noise, and fortunately this is the exact same fan used for the 9600 GSO.
Officially the GeForce 9600 GSO is positioned between the GeForce 9500 GT, which is priced between $59-$69 and the 64-shader GeForce 9600 GT, which is now priced at $99.
With an MSRP of $79, the GeForce 9600 GSO is priced to take on the Radeon 4670 directly, and while it’s based on NVIDIA’s 65-nm G92 GPU today, NVIDIA has confirmed that like the GeForce 9800 GT, the 9600 GSO will soon be making the transition to 55-nm. 55-nm is in full production today, so expect to see 55-nm GeForce 9600 GSO boards based on NVIDIA’s G92b GPU hit the market in the weeks ahead.
| System Setup | Page:: ( 4 / 18 ) |
Intel Core 2 Duo E8600
Gigabyte X48-DQ6
4GB OCZ Platinum DDR2-1333
GeForce 9800 GT
GeForce 9600 GT
EVGA e-GeForce 9600 GSO
GeForce 9500 GT
ForceWare 177.92
ATI Radeon 4650 512MB
ATI Radeon 3850 512MB
ATI Radeon 4850 512MB
8.53-080805a-067874E-ATI Beta
300GB Western Digital Caviar SE
Windows Vista Ultimate 64-bit w/Service Pack 1
Benchmarks
Company of Heroes 1.71
World in Conflict
Elder Scrolls IV: Oblivion 1.2
Call of Duty 4 1.4
Half-Life 2 Episode Two
Lost Planet
Crysis 1.2
Quake Wars
BioShock
Devil May Cry Performance Benchmark
Game Settings Used
[image]
| <% print_image("17"); %> | <% print_image("18"); %> | <% print_image("19"); %> |
| <% print_image("20"); %> | <% print_image("21"); %> | <% print_image("22"); %> |
| <% print_image("23"); %> | <% print_image("24"); %> | <% print_image("25"); %> |
| <% print_image("26"); %> | <% print_image("27"); %> |
| <% print_image("28"); %> |
| <% print_image("29"); %> | <% print_image("30"); %> | <% print_image("31"); %> |
| <% print_image("32"); %> |
| World in Conflict Performance | Page:: ( 5 / 18 ) |
| World In Conflict 1280x1024x32 | ||
| Card | Min FPS | Max FPS |
| GeForce 9500 GT | 12 | 38 |
| GeForce 9600 GSO | 21 | 66 |
| GeForce 9600 GT | 29 | 88 |
| Radeon HD 3850 512MB | 23 | 69 |
| Radeon 4670 512MB | 19 | 60 |
| ||
| Company Of Heroes DX10 | Page:: ( 6 / 18 ) |
| Company of Heroes Performance 1280x1024x32 | ||
| Card | Max FPS | Min FPS |
| GeForce 9500 GT | 13.8 | 88.1 |
| GeForce 9600 GSO | 25.6 | 149.7 |
| GeForce 9600 GT | 26.8 | 184 |
| Radeon HD 3850 512MB | 28.7 | 135.6 |
| Radeon 4670 512MB | 23.4 | 141.8 |
| ||
| Quake Wars | Page:: ( 7 / 18 ) |
| HL2 Episode 2 | Page:: ( 8 / 18 ) |
| Lost Planet DX10 | Page:: ( 9 / 18 ) |
| Call Of Duty 4 | Page:: ( 10 / 18 ) |
| Call of Duty 4 Performance 1280x1024x32 | ||
| Card | Min FPS | Max FPS |
| GeForce 9500 GT | 21 | 62 |
| GeForce 9600 GSO | 36 | 104 |
| GeForce 9600 GT | 41 | 124 |
| Radeon HD 3850 512MB | 33 | 100 |
| Radeon 4670 512MB | 26 | 86 |
| ||
| Crysis DX10 | Page:: ( 11 / 18 ) |
Crysis High – Direct3D
| Crysis Performance 1280x1024x32 | ||
| Card | Min FPS | Max FPS |
| GeForce 9500 GT | 22 | 35.2 |
| GeForce 9600 GSO | 39.7 | 68.3 |
| GeForce 9600 GT | 39.4 | 74.9 |
| Radeon HD 3850 512MB | 37.8 | 75.9 |
| Radeon 4670 512MB | 37.4 | 67.5 |
| ||
| BioShock DX10 | Page:: ( 12 / 18 ) |
| BioShock Performance 1280x1024x32 | ||
| Card | Min FPS | Max FPS |
| GeForce 9500 GT | 20 | 41 |
| GeForce 9600 GSO | 36 | 81 |
| GeForce 9600 GT | 45 | 96 |
| Radeon HD 3850 512MB | 21 | 76 |
| Radeon 4670 512MB | 20 | 62 |
| ||
| Oblivion | Page:: ( 13 / 18 ) |
| Oblivion Performance 1280x1024x32 | ||
| Card | Min FPS | Max FPS |
| GeForce 9500 GT | 20 | 29 |
| GeForce 9600 GSO | 37 | 47 |
| GeForce 9600 GT | 42 | 59 |
| Radeon HD 3850 512MB | 38 | 51 |
| Radeon 4670 512MB | 37 | 48 |
| ||
| Devil May Cry 4 | Page:: ( 14 / 18 ) |
| CrossFire Performance | Page:: ( 15 / 18 ) |
| CrossFire Performance (cont’d) | Page:: ( 16 / 18 ) |
| Power, Overclocking, Noise | Page:: ( 17 / 18 ) |
We maxed out the sliders in Catalyst Control Center when OC’ing the 4670. 800MHz core/1050MHz memory was as far as ATI’s driver team let’s us go at this point.
| Conclusion | Page:: ( 18 / 18 ) |
With today’s introduction of the Radeon 4670, you not only get a very inexpensive GPU priced at just $79, you also get a card that’s capable of dishing out very good performance for its intended segment of the graphics market. In some cases the 4670 outran ATI’s Radeon 3850 in our testing! The card also scaled quite nicely when going from one to two cards for CrossFire (Crysis being the only exception).
This is all made possible thanks to its 320 shader architecture and improved texturing horsepower (among other things, ATI has quadrupled the number of texture units in RV730 when compared to RV635), not to mention the tweaks ATI has made to their ROP subsystem for the R7xx generation of GPUs.
All these enhancements allow the GPU to pack quite a bit of punch for a card that relies on a simple 128-bit memory interface.
DX10 games like Crysis, Company of Heroes, World in Conflict, and Devil May Cry 4 ran quite well on the 4670. In Crysis the card was just 10% slower than the Radeon HD 3850 at 1600x1200, delivering a playable 40 fps when using the game’s medium settings. In Devil May Cry 4 and Company of Heroes the 4670 narrowed that gap to just 6% at the same resolution with 4xAA. The Radeon 4670 actually outran the Radeon 3850 in our testing with HL2 Episode Two. The only test where the Radeon 3850 consistently pulled ahead of the 4670 by double digit margins was BioShock DX9 with 4xAA.
NVIDIA’s GeForce 9600 GSO is a formidable competitor to the Radeon 4670 however. In our tests the cards traded the performance crown several times, with games like Devil May Cry 4, Episode Two, and Quake Wars running faster on the Radeon 4670, while the 9600 GSO pulled ahead of the ATI card in Crysis, Company of Heroes, BioShock, Call of Duty 4, and Lost Planet (which uses the same game engine as Devil May Cry) all running faster on the 9600 GSO. World in Conflict and Oblivion were basically a draw.
Overall the GeForce 9600 GSO was a little bit faster than the Radeon 4670, but it wasn’t by much.
Of course, the Radeon 4670 is considerably cheaper for ATI to produce than GeForce 9600 GSO is for NVIDIA. NVIDIA’s basically selling their $150 9800 GT chips in $79 boards, while ATI’s architecture is purpose built for the sub-$100 market, sporting a much smaller die (146 square millimeters in RV730 versus G92’s 330 square millimeters) and a less complicated PCB design. The 4670 also has a power advantage over the GeForce 9600 GSO as well.
The only real tweak we’d like to see ATI implement is a slightly faster, or a slightly larger fan. The Radeon 4670 doesn’t run as hot as the Radeon 4850, but it sure would be nice if the card ran another 10 degrees cooler than the reference board design does today.
Which card would we pick if it were our money? That’s a really tough question that’s ultimately going to boil down to which games you play, and the application you plan on using the card for. In our opinion the Radeon 4670 would be the perfect choice for HTPC use thanks to its diminutive size, native audio over HDMI, and low power consumption. We’d probably pick up a card that integrates a better cooler than the ATI reference design though.
The GeForce 9600 GSO ever so slightly gets the overall edge in gaming, with PhysX support as icing on the cake. We still haven’t had an OMG PhysX moment, but with games like Brothers In Arms: Hell’s Highway, Borderlands, and Mirror’s Edge shipping between now and January 2009, we may finally see a compelling PhysX game soon. ATI will of course counter with DirectX 10.1 support. So far three titles have been confirmed: the RTS Battleforge from EA/Phenomic, SEGA's RTS Stormise, and NHN Games RPG Cloud 9.
NVIDIA had better not take ATI for granted anymore though. The Radeon 4000 series has put ATI back on the map; they’ve now got a very compelling GPU at multiple segments while NVIDIA is basically forced to slash prices on existing products in order to remain competitive. This is having an awful effect on NVIDIA’s margins and until they can respond with a new architecture of their own they’ll continue to suffer: while ATI has room to trim 4670 prices even further, NVIDIA really can’t afford to go any lower. They’re practically giving the GeForce 9600 GSO away at cost as it is.
This is the conundrum NVIDIA is in, and until they can work their GeForce GTX technology down to lower price points, they’ll have to continue to rely on previous generation G9x parts like the 9600 GSO. Interestingly enough, this is exactly why ATI’s abandoned the monolithic GPU. Perhaps NVIDIA may have to follow suit as well?