[Hardware] Why does this NVIDIA graphics card consume SIX TIMES less than the RTX 4090?

[#Steeven.™]

Recently, NVIDIA presented a variant of its RTX 4070 Ti, based on the AD104 chip, with a particularity, its consumption is much lower than the gaming version, being 70 W instead of 285 W. This model was presented in its latest GTC and it's about the RTX 4000 SFF, but the key question is: why such a difference in power between both graphics cards if they use the same GPU?

Many of you will have seen NVIDIA's professional graphics cards and one thing you will have observed is that they are much thinner and their design is much more elegant and beautiful than those for gaming. This is because they have much lower power consumption than models that are sold for gaming. However, we are not talking about the exact same hardware.

RTX 4000 SFF, a graphics card with a TDP of only 70 W
We are dealing with a graphics card that is not designed to play games, in fact, it has been designed to be in a workstation or a server running 24 hours a day and 7 days a week without interruption, so it cannot be allowed continuous exercises of speed climbs. Typically, several of these graphics cards are connected to the same system, doing tasks such as:

• Rendering of 3D scenes for movies and series.
• For inference, Deep Learning applications such as ChatGPT, Stable Diffusion, Dall-E and others.
• For video transcoding in streaming movies and series.

These graphics cards usually have an NVLink port, proprietary from NVIDIA, which allows them to synchronize their work, since many times not one, but several are used per system and the work is shared equally. What allows them to keep each one of them with little load and to be able to work uninterruptedly.

Why do these types of graphics cards exist?
We must start from the fact that the games are not designed to use more than one graphics card, on the other hand, many other applications are. For a gamer, having four graphics cards with a TDP of 70 W does not help him, since he will not be able to use more than the first one. But in other applications in terms of performance per watt yes and if you can add the power of all of them, well that is what you end up gaining in a workstation.

The only application that the RTX 4000 SFF does not have, apart from gaming, where it performs less, is in scientific computing and learning for artificial intelligence. In that case, NVIDIA's H100 chip is much better prepared, but its price is much higher and it is not necessary for certain applications.

The secret of its low consumption, its technical specifications

To understand the reason why RTX 4000 SFF consumes so little, we have to look at its technical characteristics, because it is where the key to everything is located.

• Like the fact of having a 160-bit GDDR6 configuration, instead of 192-bit.
• Its bandwidth is 320 GB/s, which means that it uses memory at 16 Gbps, something that can be achieved with GDDR6.
• The RTX 4070 Ti, which is based on the same chip, uses 502.4 GB/s GDDR6X memory.
• However, the amount of memory in the RTX 4000 SFF Ada is 20 GB, this is because it uses clamshell mode where two memory chips can be used, sharing the data bus, but increasing capacity.
• The memory supports error correction.
• It has a configuration of 6144 "CUDA" drives and therefore 48 full or SM cores. This is less than the 60 SM of the RTX 4070 Ti. So the rest of its specifications translate into the following:
• 6144 CUDA Cores.
• 192 Tensor Cores.
• 48 RT Cores.

Now, its power is 19.2 TFLOPS in 32-bit floating point, a figure that places it below the RTX 3070 and half that of an RTX 4070 Ti and that is that therein lies the secret of this graphics card. The RTX 4000 SFF consumes only 70 W due to the fact that its clock speed is much lower. Technically, we could consider it as a GPU for a laptop in the form of a graphics card and thanks to its low consumption it can be working uninterruptedly.

 

https://hardzone.es/noticias/tarjetas-graficas/rtx-4000-sff/