Return to




I still want to see a die shrunk Excavator with RX 550 graphics as iGPU and 1GB eDRAM for a framebuffer as an embedded PC part to take on Atom processors.


Realistically, Vega and Zeppelin are better suited for that because you can downclock and undervolt them both to make them extremely power effective.

Excavator was terrible, and Polaris isn’t quite as effecient as Vega, and realistically doesn’t cost any less either. Vega is just a giant die.


I thought Zeppelin was a 16 core design? Or are you just talking Zen in general?

Didn’t know Vega was so good for power when undervolted. So maybe a cherry picked R3 2200G with eDRAM?


Zeppelin is a 4 core design. A 1700, for example, has 2 Zeppelin dies on one package. A 1950x has 4 Zeppelin dies on two packages, and similarly, the 32 core EPYC CPU has 8 Zeppelin dies on a 4 packages.

Packages being the things that you see after you delid the CPU.

I don’t even know that you need them to be cherry picked. Undervolting is easy, overclocking, not so much.


It would need to be fairly cool as were talking about 20W of heat at the high end. To compete on the low end your gonna need to be at around 5W of heat.

Could maybe get away with 1C/[email protected] if you can get the undervolting good enough


Yeah. AMD isn’t quite there on the low-end. AMD doesn’t quite have Intel’s power efficiency in mobile, but they’re definitely up there in the 3.5-4.0 range, and it scales wonderfully from 4 to 16 cores due to the MCM architecture.


Would sell like hotcakes if they could get a decent cpu and gpu combo at 10W of heat.


Why would AMD do that when they can just use cast off Ryzen dies much cheaper?

What you’re proposing is a bunch of R&D (to build the new die) and a seperate production line to produce dead superseded technology.

When they could just use dies they already made that they’d otherwise be throwing away…


Vega is pretty great at low power. The only real issues with vega are:

  • dependence on software for new feature support
  • clocks pushed too far

It scales down quite well. It just doesn’t clock as high as AMD might like (for higher end desktop parts) and at the high end is also somewhat bandwidth starved (yes, even with the HBM).

7nm Vega with 2x the HBM modules for more bandwidth, and higher clocks will fix that.


More speculation of 48 cores next year 64 cores in the wind.