David Kanter (dkanter.delete@this.realworldtech.com) on February 8, 2015 1:39 am wrote:
> Not on any vector workloads. The A15 had some terrible behavior when doing SIMD.

I've never seen a vector workload where Krait outperforms A15. Krait's raw FP perf/cycle capabilities are a little worse than A15 (although they improved it between 200 and 300):

http://www.anandtech.com/show/6971/exploring-the-floating-point-performance-of-modern-arm-processors

Yes you can make up for this with clockspeed but you're not going to find an awful lot of code running on mobile platforms that care about nothing but peak FP throughput.

As for integer throughput, A15 is more like 3x64-bit/cycle than 2x, despite claims (based on tests I've done), and I run very integer SIMD heavy code that performs better on it than Krait.

> What do you think the glass jaws of Krait are?

High latency (6 cycles) to L1 cache when 4KB direct mapped misses, which under some workloads will be a lot. Small (16KB) L1 cache. Very high latency L2 cache, partly because of asynchronous clocking. Bizarre and crippling decode stalls limitations for simple code. See here:

http://www.7-cpu.com/cpu/Krait.html

> I don't think so. I believe Qualcomm is only using ARM
> cores for their APs due to time to market constraints.

Just how many years do you think it would have taken them to update Krait to 64 bits?