Introduction
It’s been an awfully busy week if you follow PC graphics closely. On Monday AMD and NVIDIA debuted a pair of teraflop workstation graphics cards, while NVIDIA also introduced new high-end cards for the performance and enthusiast segments: the GeForce GTX 260 and GTX 280. Oh, and don’t forget AMD’s Cinema 2.0 event which Alex attended. AMD then followed that up with new performance drivers which have already made some of our 3870 X2 performance results from Monday obsolete!
Now we’re here today to report on two more GPUs: the Radeon HD 4850 and the GeForce 9800 GTX+!
AMD’s Radeon HD 4850
AMD’s Radeon HD 4850 is built around AMD’s new 800 stream processor RV770 Pro GPU. The GPU is clocked at 625MHz and ships with 512MB of GDDR3 memory running at 1.0GHz (2.0GHz effective):
| Radeon GPU Comparison |
| Radeon HD 4870 512MB | Radeon HD 4850 512MB | Radeon HD 3870 512MB |
| Transistors | 965 Million | 965 Million | 666 Million |
| Manufacturing Process | 55-nm | 55-nm | 55-nm |
| Stream Processors | 800 | 800 | 320 |
| Texture Units | 40 | 40 | 16 |
| ROPs | 16 | 16 | 16 |
| Core Clock Speed | 750MHz | 625MHz | 775MHz |
| Memory Clock Speed | 3.6GHz Effective GDDR5 | 2.0GHz Effective GDDR3 | 2.25GHz Effective GDDR4 |
| Math Processing Rate | 1.2 Teraflops | 1.0 Teraflops | 0.497 Terflops |
| DirectX Support | 10.1 | 10.1 | 10.1 |
| Tessellation Unit | Yes | Yes | Yes |
| UVD | Yes | Yes | Yes |
| PowerPlay | Yes | Yes | Yes |
| MSRP | $299 | $199 | |
 |
The Radeon HD 4850 is a single-slot board, with a cooler similar to AMD’s previous Radeon HD 3850. As you can see on our VisionTek card, the reference board design shares many traits with 3850. You can’t miss the copper heatsink/fan unit AMD employs, the cooler is even large enough to cool the board’s VRM circuitry. The fan is a variable speed unit that adjusts RPMs based on temperature.
During operation the fan runs quietly, although in our short amount of testing we’ve already noticed that the board’s PCB gets extremely hot under load. This doesn’t seem to affect GPU temps though, as Catalyst Control Center records idle temp of 55 degrees Celsius, while load temps topped out at 83 degrees. Because of the high PCB temps, you probably will want to place a fan near the card just to prevent creating a hotspot within your case.
Moving on to the backplate of the card, AMD employs two dual-link DVIs. HDMI adapters are then included with the card for HTPC users. The card requires a single 6-pin PCIe power connector for power.
NVIDIA’s answer to the Radeon 4800 series: GeForce 9800 GTX+
Not wanting to be outdone by AMD, NVIDIA is countering the Radeon HD 4800 series with a new GeForce 9800 GTX+ GPU, as well as price cuts to the rest of the GeForce 8/9 family.
As its name implies, the GeForce 9800 GTX+ is an improved version of the original GeForce 9800 GTX. The GPU itself is based on TSMC’s smaller 55-nm manufacturing process, and the 9800 GTX+ runs at higher clock speeds than the GeForce 9800 GTX: 738MHz on the graphics core versus the original 9800 GTX’s 675MHz. Meanwhile, the stream processors run at 1836MHz, an improvement of 148MHz over the 9800 GTX. The memory clock speed remains unchanged at 1.1GHz.
The board design and cooling of the GeForce 9800 GTX+ is identical to the original GeForce 9800 GTX. If you place the 9800 GTX+ side-by-side with the 9800 GTX, you wouldn’t spot any differences. The original 9800 GTX cooler does a good job of cooling the GPU without generating much noise, so this is a bonus in our opinion.
So what’s the price point on the new GTX+? As of today, NVIDIA’s GeForce lineup now looks like this:
GeForce GTX 280: $649
GeForce GTX 260: $399
GeForce 9800 GTX+: $229
GeForce 9800 GTX: $199
GeForce 8800 GT 512MB: $169
GeForce 9600 GT 512MB: $149
NVIDIA has prepped a new 177.39 driver to coincide with the 9800 GTX+ launch which includes PhysX and CUDA support, but to bring you this quick take article as quickly as possible we’re using NVIDIA’s current WHQL-certified 175.19 drivers for testing. The 175.19 driver properly detects the 9800 GTX+, although obviously lacks PhysX and CUDA support. We’ll re-run our tests with 177.39 ASAP.