how many gigabits per second does a CPU need

[The_Promised_End]

17 minutes ago, Cebolonha said:

You can have a really really long cooking recipe that uses only 1 or 2 ingredients. The dish takes hours and hours to get ready to eat. 

You can also have a really short recipe that uses 15 ingredients. Just mix them together and it's done in 5 minutes.

Recipe is the program, ingredient is the data your program uses.

got it thanks

[SpaceGhostC2C]

1 hour ago, The_Promised_End said:

lol what i ment is this we can use a gpu almost with out bottle necks at a speed of 40 gigabits per second could we do this with a cpu and how much more bandwidth would it need

Bandwidth to communicate with who?

The bandwidth you are referring to is the bandwidth of CPU-GPU communications. A CPU can live with almost 0 bandwidth of any kind by performing operations with data limited to its own cache.

[ac130kire]

**Disclaimer**(This assumes that you are talking about a modern cpu that you would find in a normal system)**Disclaimer** Well the problem is that the interface does not support another CPU nor will most CPU's support another CPU. Xeon (server grade processors) can support up to 3 other processors on the same motherboard, but consumer grade ones can only have one processor in the entire system.

 

Bandwidth really depends on what workload you are putting on the extra processor, for example when mining cryptocurrencies on a graphics card (basically another processor, but architecturally very different from a CPU), I barely saturate a PCIe 2.0 1x connector which has a max theoretical bandwidth of 500MB/s because the calculations happen on the GPU and the input stream and the output stream from the CPU to the GPU are not very data intensive. If I were to play games on that same graphics card, I would need much more bandwidth than that or I suffer a very significant loss of frames (ex. 240fps vs 30fps). As you can see different workloads require different amounts of bandwidth.

 

Certain very old computers form the late 1900's and maybe even some from the early 2000's used to be able to utilize something called a Math coprocessor, basically this chip aided the main CPU in calculating mathematical operations speeding up processing time for the whole system and this was attached using a connector kind of like PCIe, but it was probably very different in the way it worked, I don't know too much about this, I would have to do some research.

[Dash Lambda]

Another big problem here is that the only reason we can separate a GPU so much from the rest of the system is that it's essentially self-contained. Connecting an eGPU is more like hooking two full PCs together than it is putting the CPU on the end of a cable.

What this means for CPUs is that, in order to use an external one, it would basically have to be its own self-contained machine. Which, for the record, is absolutely something we do, or at least a form of it, with compute clusters and whatnot.