Finally, NVIDIA announced mid-end solutions based on its latest GeForce GTX 400 architecture that we examined as far back as in the beginning of this year (the most powerful product based on this architecture was rolled out later, in March). Now it's time to review the more affordable GPU and graphics cards that come with it.
Obviously, NVIDIA is very interested in affordable solutions based on the new architecture. Top-class products are beneficial, but don't sell as good as mid-end products do. For some time, GeForce GTX 260 was enough for most users. But in 2010, the need for a new architecture became obvious. The existing GPUs only supported DirectX 10, while the main rival had already been offering DirectX 11 solutions across all price ranges.
The reason for the sluggish roll out of both GF100 and GF104 were the same old problems with making such powerful GPUs according to the TSMC's 40nm process technology. The long-time mastering of the process technology was launched long ago. First, ATI and NVIDIA had mobile GPUs made, then ATI made its RV740, and NVIDIA got its GT21x solutions.
ATI was quicker in making new products based on the 40nm process technology. This can be said about the top-class RV870 as well as later GPUs for mid-end and low-end price ranges. One reason is that ATI GPUs were somewhat simpler (in terms of transistors per chip). Besides, NVIDIA's GF100/GF104 is a completely new architecture, not a reworked old one.
But even now the release of GeForce GTX 460 is timely enough. And the solution itself promises to be quite a success. At least according to NVIDIA's surveys, a typical user changes the graphics card (or the entire system) once in three years, spending about $200-250. According to Steam Hardware Survey stats, 31% of users have graphics cards of the aforementioned price range. While 55% of users have cheaper and less than 15% of users have more expensive graphics cards.
NVIDIA believes that GeForce 8800 GT was exactly the same solution in the fall of 2007. It costed the same, it was also aimed at upper mid-end and offered excellent DX10 performance at affordable price. GeForce GTX 460 may become as successful for the newer DirectX 11 API. But these are just assumptions. We'll see how it goes.
The theoretical part won' be long, because GF104 (and GeForce GTX 460) are quite similar to GF100 (GeForce GTX 470 and GeForce GTX 480), all differences being quantitative. Besides, most information on the Fermi architecture was disclosed as far back as in January. Just make sure you've read the reviews of the GF100 architecture and GeForce GTX 470/480 graphics cards.
You might also want to read articles on NVIDIA's previous architectures and solutions:
For now we'll assume that you're already familiar with all NVIDIA GPU architectures and proceed to the specifications of the GF104 GPU and products based on it.
GeForce GTX 460 series specifications
- Codenamed GF104
- 40nm process technology
- 1.95 billion transistors
- Unified architecture with an array of processors for stream processing of vertices, pixels, etc.
- Hardware support for the DirectX 11 API, including the Shader Model 5.0, geometry and compute shaders, as well as tesselation
- 256-bit memory bus, 4 independent 64-bit controllers, support for GDDR5 memory
- 675 MHz core clock rate
- Doubled ALU clock rate of 1350 MHz
- 8 Stream Multiprocessors, including 384 scalar ALUs for floating-point computing (support integer, FP, FP32, FP64 according to the IEEE 754-2008 standard)
- 64 texture addressing and filtering units supporting FP16, FP32 precision in textures, as well as support for trilinear and anisotropic filtering for all texture formats
- 4 wide ROPs (32 pixels) supporting antialiasing up to 32x, also with FP16, FP32 frame buffers; each unit features an array of configurable ALUs and handles Z generation and comparison, MSAA, blending
- Recording results of up to 8 frame buffers simultaneously (MRT)
- Integrated support for RAMDAC, 2 x Dual Link DVI, HDMI, DisplayPort
GeForce GTX 460 reference card specifications
- 675 MHz core clock rate
- 1350 MHz universal processors clock rate
- 336 stream processors
- 56 TMUs, 24/32 blending units
- 3600 (900x4) MHz effective memory clock rate
- GDDR5 memory, 192/256-bit memory bus
- 768/1024 MB memory
- 86.4/115.2 GB/s memory bandwidth
- 16.2/21.6 Gpixel/s theoretical peak fillrate
- 37.8 Gtexel/s theoretical texture fetch
- 2 x Dual Link DVI-I, 1 x HDMI mini, resolutions up to 2560x1600
- Dual SLI connector
- PCI Express 2.0
- Supports HDCP, HDMI, DisplayPort
- Consumes up to 150/160 W via two 6-pin connectors
- Dual-slot design
- MSRP $199/$229
Like its top-class predecessor, the new mid-end GPU GF104 is also made according to the 40nm process technology. That's the only way to make such a powerful solution with almost 2 billion transistors and keep the cost of ready graphics card at about $200.
NVIDIA hasn't changed the naming policy, meaning that GeForce GTX 460 should be on the same level as the previous-generation GeForce GTX 260. As for the 3xx series, those solutions are based on older GPUs and intended for OEM builders.
The GF104 series currently includes only GeForce GTX 460 that should replace everything from GeForce GTX 260 to GeForce GTX 285. But as you can see in the specs, there will be two different models named GeForce GTX 460.
Having different amount of memory, the models will have the same GPU/memory clocks rates and the same number of ALUs and TMUs. However, GeForce GTX 460 768MB will have 192-bit bus, 24 ROPs and 384KB L2 cache, while GeForce GTX 460 1GB will have 256-bit bus, 32 ROPs and 512KB L2 cache. Obviously, the two will have different memory bandwidths.
All these differences are related to the number of active 64-bit memory channels. GeForce GTX 460 768MB has 3 of those, while GeForce GTX 460 1GB has 4. Thus, the number of memory chips on PCB is also different. By the way, power consumption is also a bit different: 150W (GTX 768MB) and 160W (GTX 1GB). It's not quite clear why NVIDIA decided to go with two such models of the same name.
In all other aspects, the solutions are identical. Both have 336 stream processors (of 384 physical ALUs present in the GPU) and 56 TMUs (of 64 physical TMUs present in the GPU). Yet again we can see that there are no solutions with the fully working GPU. Both GeForce GTX 460 variants are cut down to a certain extent. The performance difference will hardly be more than 10-15%, except for situations where 768MB of memory is not enough. But that's still too much a difference for two cards of matching name.
As the yield of GF104 GPUs increases, and GeForce GTX 470 sales go on, NVIDIA may release at least one more graphics card based on this GPU. It may be a kind of replacement for GeForce GTX 465 (or even GeForce GTX 470) that will use all of the 384 ALUs and 64 TMUs and will have higher GPU/memory clock rates. If the GPU will work at 750MHz and memory at 1200 (4800) MHz, that solution will be able to compete with GeForce GTX 470. All the more so NVIDIA claims that GF104 has excellent overclocking potential.
