Nice work, Dylan

Expand full comment

You miss the elephant in the room afaict.

The top die (their key AI warrior) is double the size of their historical comfort zone (for which there are sound reasons laws of physics/power/heat/yields/economics...).

Why this desperation?

The answer is AMDs AMD's MCM GPUs on Fabric.

This is huge.

It is what zen chiplets did for CPU.

Even NVs software moat cannot withstand the raw muscle AMD can cost effectively apply to BIG tasks.

AMDs modules allow 2x epyc & 4x GPU to interconnect at 3TB/s via Infinity Cache. Lacking their own platform after the ARM fail, they know very well they face a big threat in increasingly Epyc dominated data centers.

Expand full comment

The total die size of AD102 is slightly smaller than GA102, albeit on a more expensive node.

Power for IO isn't free. Advanced packaging isn't a free lunch.

As for server, MI250X has some major deficiencies, MI300 will be a marvel of HW engineering, but you can't run much on it so...

Expand full comment

ur vague response re details ignores the key matter - amd can do ~gpu chiplets teamed with multiple epycs on a brutally fast & ~all embracing Fabric bus

i am skeptical of these deficiencies - amd did knowingly concede in design where they cant win for now (AI), & focused on areas like scientific, but "deficiencies" in Zen1 didnt stop them crushing a then totally dominant intel.

Expand full comment