Yes, I didn't mean to imply we wouldn't go far to squeeze Go further! At some point the downside of the complexity of dropping down from Go to intrinsics, bypas

Yes, I didn't mean to imply we wouldn't go far to squeeze Go further! At some point the downside of the complexity of dropping down from Go to intrinsics, bypassing the GC, etc. might overwhelm the benefits of staying with Go. We've spent a while writing simulations like the ones in the post to understand Go's GC's behaviour under lots of pressure, and it won't be a bottleneck for a long time. I also think that doing all the pre-requisite work to make SIMD remotely possibly will yield _far_ more performance than SIMD itself.