Dropping the 3D cache on the 7900X — AI concession or no real loss?

  The 7900X's 64MB of L3 forfeits a slice of interactive sweetness, but its home turf is another class of workload: RAG chunking, data cleaning, embedding and batched quantized inference lean on memory and cores far more than cache, and 12 cores/24 threads at 5.6GHz with AVX-512 absorb them steadily, with the 4.7GHz base keeping short single-stream runs snappy.

  If your work is throughput-shaped, move the X3D savings into memory modules — the return is more direct than cache. Shoppers who live on interactive latency should put the 128MB part on their list instead. Desktop AI money should land on whichever resource runs out first, and your workload decides which one that is.