Signs of Raven Ridge - AMDs next APU

Ridelynn

https://browser.geekbench.com/v4/cpu/3989399

Starting to see some signs of Raven Ridge come out. That's the new Ryzen/Vega-based APU. I think it has a lot of potential, as I've always thought AMD APUs do. AMD APUs ~should~ be dominating laptops and budget/low-end gaming rigs right now, and maybe Raven Ridge will bring them around to seeing their potential.
 
That being said... It's just a canned benchmark, but this makes for an interesting discussion point. 

Cleffy

Threadripper APU would be epic. Can use a die the size of the one going into the XBox X with some on-chip memory. It's kind of what is expected. Half the die of Ryzen. So it shouldn't be as bad when running multiple threads across the infinity fabric.

Ridelynn

Cleffy said:

Threadripper APU would be epic. Can use a die the size of the one going into the XBox X with some on-chip memory. It's kind of what is expected. Half the die of Ryzen. So it shouldn't be as bad when running multiple threads across the infinity fabric.

That would be an interesting product - the TR package already has space for 2 GPU dies or HBM stacks to drop right in (the two "dummy" Ryzen dies under the heat spreader). The only issue would be power/thermal... TR already has a very high power and thermal density (which is astonishing, the package is the size of a credit card already), putting more under the lid makes it even more challenging to keep performance up and power/thermal under control.

Quizzical

Ridelynn said:

Cleffy said:

Threadripper APU would be epic. Can use a die the size of the one going into the XBox X with some on-chip memory. It's kind of what is expected. Half the die of Ryzen. So it shouldn't be as bad when running multiple threads across the infinity fabric.

That would be an interesting product - the TR package already has space for 2 GPU dies or HBM stacks to drop right in (the two "dummy" Ryzen dies under the heat spreader). The only issue would be power/thermal... TR already has a very high power and thermal density (which is astonishing, the package is the size of a credit card already), putting more under the lid makes it even more challenging to keep performance up and power/thermal under control.

Threadripper's thermal density isn't unusually high.  High end GPUs commonly have more heat than both threadripper dies added together coming off of a single die.  A Radeon RX Vega 64 or GeForce GTX Titan Xp probably does, for example, and that die is only about double the size of a ryzen die.

Ridelynn

Well, packages that have higher thermal (the GPUs you reference) also tend to come premated with a specifically engineered cooling solution - not multiple dies underneath a giant heat spreader and user-mounted heat sink/fan.

People are having issues cooling TR now without water - part of that is due to new package size and lack of options to match that. Part of that is the ATX form factor and the space you have available for physical geometries of a HSF. But part of that is just because 200W is a good deal of heat to remove, and that skyrockets if overclocked. There may be GPUs and CPUs (the 9590 comes to mind) that do more, but that doesn't mean TR isn't a lot. 

Quizzical

Ridelynn said:

Well, packages that have higher thermal (the GPUs you reference) also tend to come premated with a specifically engineered cooling solution - not multiple dies underneath a giant heat spreader and user-mounted heat sink/fan.

People are having issues cooling TR now without water - part of that is due to new package size and lack of options to match that. Part of that is the ATX form factor and the space you have available for physical geometries of a HSF. But part of that is just because 200W is a good deal of heat to remove, and that skyrockets if overclocked. There may be GPUs and CPUs (the 9590 comes to mind) that do more, but that doesn't mean TR isn't a lot. 

GPUs having multiple dies under the heatspreader is becoming a trend, with Fiji, Vega 10, and GP100 all having quite a few.  You're right about custom cooling solutions rather than user-mounted, however.  And a bunch of HBM dies isn't the same as a CPU die that cranks out 100 W.

The big problem with users cooling Threadripper is that the heatsinks that have existed in the past simply aren't built for something with heat spread that far apart.  You really have to get something specifically built for Threadripper (or Epyc, though servers are a different grade entirely) to have the appropriate areas covered.