[camel-cdr]

For anybody wondering the P670s are advertised to be around Cortex-A78 performance, their slides 5% slower on SpecINT2k6, but 50% less area) [0] and have 128 bit vector registers.

The X280s are more of an external accelerator with 512 bit wide vector registers and an integrated TPU using Sifives VCIX interface [1]. It enables custom vector coprocessors, in this case a TPU, to interface directly with the vector register file.

The X280s will very likely not run the same OS as the main P670s because of their different VLEN, which would't allow for moving processes between them.

[0] https://www.youtube.com/watch?v=C4DcZyB6i1E [1] https://www.youtube.com/watch?v=si5w7JQtSDk

[snvzz]

>5% slower on SpecINT2k6, but 50% less area

That's a huge difference in area.

In line with SiFive tradition, always huge area advantage against any given performance point's ARM design counterpart.

I am amazed they manage to still keep this sort of advantage up, while tackling higher and higher performance points.

[Klasiaster]

Maybe it's like with software, once you add bloat/baggage over the course of the years it's hard to get rid off when trying to optimize again.

[snvzz]

It is a good thing that at least the ISA managed to catch up with every important ARM or x86 feature yet avoid that baggage creep.

Microarchitectures might need a full redesign once in a while, but if it can keep its RISC-V software moat, that's alright. Worked for x86 so far.