from Hacker News

Arm's Neoverse V2, in AWS's Graviton 4

by gautamcgoel on 7/22/24, 10:47 PM with 5 comments

  • by adrian_b on 7/23/24, 5:53 AM

    It is interesting that a Neoverse V2 core has the same area as a Zen 4c core (the low-frequency compact variant of Zen 4).

    It is true that at this equal area the Neoverse V2 core includes an additional 1 MB of L2 cache memory, but the greater L2 cache memory is not enough to make it reach the performance of Zen 4c for the applications that are not limited by the memory bandwidth (where Graviton 4 may win).

    While Neoverse V2 has a lower, but nonetheless acceptable, performance in comparison with the old Zen 4, it is likely that it also has a lower power consumption, therefore lower operating costs for Amazon, but the value is not disclosed by Amazon. In any case, because the Graviton 4 instances are offered at a lower price, they may be preferable for many applications.

    However the new Zen 5 will be in a different performance league. Arm has also announced the successor of Neoverse V2, i.e. Neoverse V3, which is presumably derived from Cortex X4. That will be a faster core, but the differences between Neoverse V2 and Neoverse V3 are much smaller than those between Zen 4 and Zen 5, so the advance of Zen 5 vs. Neoverse V3 will be greater, in everything except possibly the power consumption.

  • by janice1999 on 7/22/24, 11:58 PM

    AMD was originally planning to have an ARM CPU that would be compatible with its AM4 socket. I believe their first customer was Amazon but Amazon dropped them for an in-house design that became Graviton due to performance reasons. Given AMD was not in a good financial state at the time it made sense to concentrate on Zen. It would probably be a very different world if AMD had managed to debut a 'Zen' inspired ARM CPU alongside its first batch of Zen processors.

  • by tedunangst on 7/23/24, 1:24 AM

    Amazing times that there are two CPU architectures from five vendors that mostly work.

  • by fulafel on 7/23/24, 9:01 AM

    The memory latency graph is pretty dramatic, also for the other systems vs the older Westmere Xen. Intel was doing something right in 2010.