Compute Units in GPU

[Cysiu]

Hello, i was thinking about , what are the Compute Units in GPU

Example : RTX 3080 have 68 Compute Units

Did they have any special meaning in gaming or taking part in graphic card efficiency?

( And I'll appreciate some theory )

[FloRolf]

The Compute Unit (or SM) is just a group of Cuda Cores. In the 3080's case, each SM contains 128 Cuda Cores. 68*128 =8704. 

That's how they do the 3090 for example, that has all 82 SMs enabled, which measn 10496 Cuda Cores. 

 

EDIT: The full GA102 even has 84 SMs. So both the 3090 and 3080 are just cut down versions of the same GPU. 

[Freakwise]

its basically what its said it, was a clusters of cuda core

so rtx 3080 have 8704 cuda core and 68 compute units

so you devided total core count with total compute units

 

and you got 128 so 1 compute units have 128 core and iirc vega igpu have 64 core per compute units

so yeah, basically other way to say spesification

[Jurrunio]

More compute units = more parallel computing = higher performance. Also brings more power draw, more memory and PCIe bandwidth usage etc. Efficiency depends on architecture, process node, frequency and voltages. Number of compute units is not a factor.

[Eigenvektor]

53 minutes ago, Cysiu said:

Did they have any special meaning in gaming or taking part in graphic card efficiency?

It's just a way of logically grouping things together on a GPU. Your GPU has a certain number of Cuda cores or Stream Processors (SP). Every n of these SPs form a logical unit, a Compute Unit (CU) or Streaming Multiprocessor (SM). Each unit may then be combined with or share access to other components such as Texture Mapping Units (TMUs) or Render Ouput Units (ROPs).

 

For us users, these technical specs are somewhat meaningless. A card of the same generation with more CUs is certainly going to be faster compared to a card with less CUs. But you can't compare them across generations and/or vendors . The number of SPs/ROPs/TMUs per CU may change, their performance and efficiency may change and so on.

 

At the end of the day you'll want to look at benchmarks to see how the performance of these cards differs, because double the core count doesn't always translate to double the performance.