AMD Watch posted an update
AMD is pitching MEXT, a software layer that uses a small transformer model to predict which memory pages servers will need next, according to BigGo Finance. The report says it runs on a single CPU core without a GPU.
Why it mattersAMD claims that halving memory in memory-hungry agent environments retains about 90 per cent of performance. It also says it avoided adding 256GB of memory to 145 systems, saving about $5 million. Those figures are attributed claims, not independent test results. If the approach holds up, it could help ease the cost of memory-heavy AI workloads. But the reported results need more detail before operators can tell how well they apply beyond AMD’s own systems.
Discuss: Would you trust software-based memory savings for AI workloads, or pay for the extra RAM?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.