- LarryPParticipantOctober 5, 2014 at 6:38 amPost count: 76
I’ve been wondering: How well could the Mill do in DSP applications and markets?
I suspect that, for applications that do fit the vast majority of their work into big pipelined loops, the Mill won’t necessarily offer sufficient performance gain to compete aggressively.
However, there are many applications that need both lots of math/array work and also have a non-trivial workload that doesn’t run well on a DSP (e.g. handling asynchronous events, resource management….) A number of such applications are now handled using hybrid systems, either hybrid at the chip or at the board/box level. (Though what fraction of the market this is, I don’t have metrics.) In my experience, the CPU heterogeneity in such systems is a pain. I suspect that, since many applications have both DSP and non-DSP in their workloads, using a Mill CPU (possibly a multi-core one) might offer some real advantages.
I’d be very interested what DSP experts think about the Mill for DSP applications and markets.
- Ivan GodardKeymasterOctober 5, 2014 at 5:10 pmPost count: 518
The Mill will do very well for DSP work up to a limit of data parallelism. If the necessary SIMD is greater than 8 or so then a GPU-like architecture will do better, albeit at the cost of having to do GPU programming.
Of course, DSP covers many quite disparate loads. For example, if the load requires double or quad floating point (GPUs are notoriously bad at DP and don’t have quad at all) then a Mill will have getter performance than a GPU and lower power than an OOO CPU. In contrast, if the DSP is really not much more than a microcontroller (for example in audio work) then a Mill is overkill.
No simple single answer, sorry.
You must be logged in to reply to this topic.