Hacker News Reader

Comment by leguminous | Hacker News Reader

The differences between the 5600 and 5500 are interesting. The $199 5600 has 6 cores, 32MB of L3, no GPU, and uses the chiplet packaging. The $159 5500 has 6 cores, 16MB of L3, a disabled/fused off GPU, and is one larger, monolithic die.

I thought the chiplet packaging was supposed to result in lower costs due to better yields on smaller dies? I guess the extra cache takes up more than enough die space to make up for that?

View on HN

The 5500 is less performant, therefore lower cost. It only exists to sell salvaged dies, and won't likely appear in large quantities (similar to how the 3300 / 3100 were impossible to find).

Everything I've been reading has been that 7nm has been having excellent yields and chiplets do bring their own costs. Lisa Su (AMD CEO) indicated in an interview a while back that there is a price floor at which chiplets make sense. Another thing to consider is that most of the chiplets used in the consumer parts could be considered rejects from Epyc. (i.e. they need to make a lot of relative duds to get the small number of 'golden samples' they use in their server chips)

None of their chiplet designs currently have GPU. All of their APUs are currently monolithic. The 5500 is just a neutered 5600g. Probably mainly trying make some money on dies with bad GPUs.

The yield issue is going to be much more of a factor when a process node is young and defect rates are high. The big savings is in engineering hours where you can re-use a design across multiple products.

5500 is APU based, and it has only PCIe-3... you can consider it a failed laptop sale.

the 5500 is a monolithic die, it's based on the monolithic APU lineup rather than the chiplet enthusiast lineup.

This also implies other limitations like PCIe 3.0 and half the cache of the equivalent enthusiast lineup, because that's how the laptop chips were designed. The APU lineup has always performed a bit worse than the equivalent desktop lineup as a result.

Reply on news.ycombinator.com