IIRC when AVX-512 first launched on consumer CPUs, it had serious thermal issues: even though AVX-512 let you do more work per cycle, it generated so much heat that the processors downclocked and you lost most of the performance benefit. Assuming this was actually a real issue and not just some clickbait I saw, had it been mitigated in newer Intel CPUs? Or is that part of the reason they're deemphasizing it?
I would imagine that it's more because the utilization of AVX-512 is low. I'm sure there are lots of applications that make use of it, but many of the instructions are bizarre and are maybe useful to just a handful of people.
Weigh that against the cost of design and manufacturing, it's not hard to imagine they're finding out it's not worth it.
I'm guessing this was due to Intel's node issues, as Intel consumer CPUs with high core count were also having thermal issues. AVX-512 can theoretically double the throughput over AVX2. That's like doubling the core count.
And with new instructions added in AVX-512, even more workloads can be vectorized. That means even bigger theoretical performance bump.
Intel’s product segmentation is just horrible. “The same die is used on server and consumer chips, so let’s just fuse off this huge chunk of silicon (that could be used for something else) because money.”
It’s no wonder AMD ate their lunch with Ryzen. Intel focused of stagnation and segmentation while AMD did the opposite.
4 comments
[ 2.9 ms ] story [ 21.5 ms ] threadWeigh that against the cost of design and manufacturing, it's not hard to imagine they're finding out it's not worth it.
This is just speculation though.
And with new instructions added in AVX-512, even more workloads can be vectorized. That means even bigger theoretical performance bump.
https://images.anandtech.com/graphs/graph16495/122770.png
It’s no wonder AMD ate their lunch with Ryzen. Intel focused of stagnation and segmentation while AMD did the opposite.