Fermi’s has had a very a rough path to realization. From the many delays, the paper launch and its heat guzzling characteristics, Fermi had the potential to turn into another GeForce 5 series for Nvidia. The flagship card based on the GF-100 is arguably the fastest single GPU graphics card around. But while getting there, it eats a ton of electricity and generates a heap of heat. Not to mention that it does cost a little fortune to buy as well. Nvidia did try and remedy the situation with another GF-100 based card, the GTX-465, but again fate was not on its side. The card was slower than its rival (the ATi 5850) and was priced about the same.

FACTS

THE SON OF GF-100; THE GF-104

Nvidia had the performance crown with the GTX-480. But the real money is not at the extreme high end. The margins in that segment are great, but sales are not. Most of the money lies in the mainstream segment. And Nvidia had nothing to show for it. ATi’s (RIP) 58xx series had the entire sub US$ 200 market to itself. This was salt to Nvidia’s already sore wounds –it was nearly a year behind ATi in terms of releasing a DirectX 11 compliant card and it was literally handing out money to ATi by not covering all the market segments.

The savior for Nvidia comes in the shape of Fermi’s second outing the GF-104. This is simply not a cut down version of the GF-100. There are architectural improvements that make this card perform better than the GTX-465, despite featuring stream processors. The GF-104 is a redesigned Fermi, rather than a cut down GF-100.

THE NUMBERS GAME

Nvidia changed the way it names its execution units within the GPU. Here is a quick recap of how Nvidia names things now.

Each GPU is made up of a certain number of Graphics Processing Cluster (GPCs). The GTX 480, for example, has 4 of these GPCs. Each GPC has 4 Streaming Multiprocessors (SMs), and its own raster engine. Before diving into the GF-104, we’ll start off with the original Fermi –the GF-100

THE GF-100
Each SM is made up of a 32 “CUDA” cores (what used to be called stream processors), 4 Special Function Units (SFUs), 4 texture units and a polymorph engine. The polymorph engine is responsible for tessellation function of the Fermi GPU. The GF-100 thus will have 512 of these CUDA cores (maths follows). These CUDA cores are the unified shaders.

32 cores per SM x 4 Stream Multi processors per GPC x 4 Graphics Processing Clusters = 512 CUDA Cores

4 texture units per SM x 4 Stream Multi processors per GPC x 4 Graphics Processing Clusters = 64 Texture Units

If you look at the table below, you will see that not even the GTX 480 has all 512 CUDA cores. Nvidia had to disable one of the SMs to improve yields.

THE GF-104

For the GF-104 Nvidia changed some of the numbers. There are 2 GPCs.

Each GPC is made up of 4 SMs. It is the architecture of the SMs that has changed with the GF-104. Rather than 32 CUDA cores and 4 SFUs & texture units, the GF-104 has 48 CUDA cores and 8 SFUs and texture units.

48 cores per SM x 4 Stream Multi processors per GPC x 2 Graphics Processing Clusters = 384 CUDA Cores

8 Texture Units per SM x 4 Stream Multi processors per GPC x 2 Graphics Processing Clusters = 64 Texture Units

The number of SFUs and texture units has been doubled, while an extra 16 CUDA cores have been added per SM. Again, one SM has been disabled (to improve yields) just like Nvidia did with the GF-100.

The final tally is shown in the table below. (The red figures show the clocks & price for the cyclone oc’d edition)

There are two variants of the GTX 460, one with 1GB of onboard memory and the other with 768 MB of memory.

THE NON-REFERENCE 460: MSI CYCLONE

MSI DESIGN MODIFICATIONS

The MSI Cyclone is a non-reference design GTX 460. The cooler, some components on the PCB and GPU clocks are non-reference.

1. COOLER

The card uses a large dual slot GPU cooler. The “Cyclone” in the card’s name refers to the cooler which uses two 6mm radial heat pipes to transfer heat to Aluminum fins. There is a large 90mm fan at the center which helps dissipate heat. The cooler is quoted to be about 16% quieter than the standard reference cooler under full load.

The memory modules are passively cooled by the air venting out from the fins.

Though this is a dual slot solution with rear venting holes, the hot is not exhausted to the out side.

2. PCB

The PCB utilizes solid state chokes and special “Tantanlum capacitors”. The former prevents buzzing noise from bog standard chokes, while the latter improves current characteristics being provided to the GPU. MSI calls these “Military Class” components.

3. CLOCKS
The core clock is 50 MHz over reference, and thus the shader clock is higher by 100 MHz. The memory clock remains the same at 900 MHz. This could be due to the nature by which the memory is cooled on the card (no heat sinks/ active cooling).

4. VOLTAGE ADJUSTMENT
In comparison to the stock card, the cyclone allows adjustment of GPU voltage settings via bundled software to improve over clocking.

The card requires two 6 pin PCI-e connectors to power it up.

PACKING

The video card comes in a large(ish) flip open cardboard box. The top of the flip has a picture of the cyclone cooler. Inside of the flip pictorially describes MSI’s customization of the GTX 460.

There is a small transparent window which shows the cooler on the video card itself.

Within the box, the card comes in a protective foam shroud covered with a transparent plastic lid.

Below the foam shroud is another compartment which contains the two user’s guides, connector dongles and the bundled software disk (see below).

The connector dongles help change the DVI to either VGA or HDMI (1.3a complaint) ports. Another dongle converts the onboard mini-HDMI port to a normal HDMI port.

The molex-to-PCIe power converter cables lie next to the card in the.

A CLOSER LOOK AT THE CARD

The first thing that stands out on the card is the cyclone cooler, mounted on a black PCB. It helps define this as a MSI product.

All the connectors are protected by a plastic cover. This includes the PCIe connector as well.

The card is 9.5” long and should fit in almost all mid tower cases.

BUNDLED SOFTWARE

The card comes with two customized programs. The first is the MSI “Afterburner” which is MSI specific version of Rivatuner. The second is called “Kombustor” Which is a derivative of Furmark application.

TEST SETUP

METRO 2033

This is another game with very impressive visuals as well as support for DirectX 11 specific tessellation function. The cyclone is in the lead here, which shows that Nvidia did a lot of work on DirectX 11 functions of the card. It edges out both the 5830 and the 5850.

CRYSIS

Despite being more than a couple of years old the CryEngine 2 powered game can generate very impressive visuals. Here the cyclone is neck in neck with the 5830, with the 5850 in lead.

STALKER –CLEAR SKY

The cyclone does well here, coming very close to a much higher priced ATi 5850.

MAFIA 2

The cyclone does a commendable job of pulling ahead of its direct rival. Phys-X (dedicated) and APEX rendering (for realistic cloth effects) were disabled as all cards did not support this feature.

TEMPERATURES

Furmark was used to check temperature of the GPU under load. The ambient temperature was 26C.

The fan speed increased automatically to about 80% under load.

CONCLUSIONS

This is perhaps what Fermi should have been from the start: A well performing GPU that remains within “sane” thermal profile. At its recommended price it clearly out performs its nearest rival (the ATi 5830) and in games that support DirectX 11 it even bests a card that is superior to it (ATi 5850).

MSI have not only improved on the GPU clock, added an after market cooler, but also improved upon the components used on the PCB. The 460, in any form is an excellent card, the Cyclone makes it only better.

PROS

• Performs very well
• DirectX 11 performance much better than its competitors
• After market cooling and components
• Factory overclock

CONS

• None