Can we extract information from TDP values?

[porina]

Right, that title should do it. I've been a bit tiered of seeing the same misunderstanding of how TDP is defined so I thought I'd have a go at extracting some different information out of it, combined with other published specifications.

 

To recap, on the Intel side at least, TDP is the cooler rating you need for the CPU to operate at least at its base clock. I'm sure there's more to the definition, such as ambient test temperatures and such, but let's focus on what we do have easy access to. We know the TDP, we know the base clock, we know the core count. We can use those three values to determine a kind of efficiency rating. It is recommended but not enforced that the CPU limits to the TDP power in the long term. Since we are looking at the base clock, this is the case for the calculation.

 

So we have a CPU limited to a known power (=TDP) and will result in a minimum base clock. It may be higher, but this is the minimum specified. We have a certain number of cores. To make comparison across models easier, we'll look at how one core behaves. If you give one core 1 W of power, what would its base clock be?

 

Above is the result for selected models of Comet Lake. For each of the 900/800/700 families, I show the T (35W) sku (left 3), normal (65W) sku (middle 3) and k (125W) skus (right 3). The 10850k is by itself at 125W, and is considered a lower binned 10900k. I decided to plot the clock efficiency against base clock. Those more experienced, especially in overclocking, will know that generally speaking you need more voltage as the clock increases. Thus the efficiency drops at higher clocks, and that is exactly what we see in this chart! Other features can be seen as expected, that T is more efficient than normal, is more efficient than k, before you start overclocking it and make it worse still. It can also be seen more cores = more efficient, for a given power limit. The 10850k is as expected a little worse than the 10900k.

 

There are also some lower skus, which I chose not to add. They would simply be worse than the same core count and TDP ones shown.

 

What about AMD you ask? I'm moving into dangerous territory here. Firstly, I've been unable to get a definition for AMD's base clock. I assume it works much the same as Intel's, but I don't know that for sure. Next, we know there are the cores, and separately the IOD. The TDP would be for the whole package. Could or should I separate out the IOD from the CCD? I think not, for the specific reason it is not meaningful to do so. It counts towards the socket power usage, and is unavoidable in use. Each core will take a share of the IODs power, so it isn't strictly only pure core power, but it is in a practical sense. So what do we get? Applying Zen 3 to the same chart we get the following:

 

Going from bottom to top, it is the 5800X, 5600X, 5900X, 5950X. Yes, the 5800X could be argued to be worse than the 5600X. This is in part due to the lower TDP of the 5600X. For the Intel CPUs, for each family (same coloured dots) you can imagine a line drawn through the points. That's where Intel can choose to put a product, and you can extrapolate a bit beyond each with caution. In clock/W efficiency terms, Zen 3 8 core is slightly ahead of Intel's 10 core, which itself is better than the lower bin 10850k and 8/6 core parts. The higher core count parts do run away though, although this is in part due to simply having more cores, and in part from process related efficiencies.

 

It should be stated that this only looks at clock in isolation, and doesn't consider IPC or other performance considerations. So don't read too much into it. Also remember these are minimums, they could be higher for a given sample. It would probably be an interesting exercise to go repeat this with other models, but I think I've spend enough time on it for now.

[Guest]

4 hours ago, porina said:

 Firstly, I've been unable to get a definition for AMD's base clock. 

 

Cpu-z save txt, open, scroll to p-states. The highest p-state will be your base clock and advertised TDP.

 

 

[porina]

7 hours ago, ShrimpBrime said:

Cpu-z save txt, open, scroll to p-states. The highest p-state will be your base clock and advertised TDP.

Thanks, but not what I'm looking for which is a worded definition of how AMD treats TDP value. Does it work much the same way as Intel which is my assumption?

[Eigenvektor]

13 minutes ago, porina said:

Thanks, but not what I'm looking for which is a worded definition of how AMD treats TDP value. Does it work much the same way as Intel which is my assumption?

This might help some:

https://www.gamersnexus.net/guides/3525-amd-ryzen-tdp-explained-deep-dive-cooler-manufacturer-opinions

[Guest]

3 hours ago, porina said:

Thanks, but not what I'm looking for which is a worded definition of how AMD treats TDP value. Does it work much the same way as Intel which is my assumption?

It's the same as Intel, and has been for a long time, pretty much since "Boost" or "turbo" clocks became a thing.

 

But as far as I'm aware, the TDP is directly max P-state power consumption.

Under the boost clocks, but over low power states, they call it an average and this figure helps designate the cooling needed for the processor.