Richard Cownie (tich@pobox.com) on 8/4/10 wrote:
---------------------------

>But it seems to be almost a universal law of chip
>development that you can't survive long by selling high-end
>chips alone, against competition that has a low-end
>high-volume revenue stream. That's how we got the
>dominance of x86 over RISC. The high-end-only chips that
>survive are those that belong to systems companies who
>can make their money not just on the chips, but on selling
>complete systems, plus software and services (which isn't
>what the graphics-card market looks like, though quite a
>few companies have tried to make HPC look that way -
>and failed).

There is an important difference that seems to have been ignored so far.

In a typical CPU lineup, there is just one chip. The only difference is usually the clock speed, or at best caches.

Taping out multiple chips with different core counts/caches from a same architecture, is a relatively novel phenomena. I am pretty sure that in RISC days, the variety was much lower.

OTOH, in a gpu line up there are typically 4-5 chips in a generation. The lowest end ones typically receive the least R&D effort and are released at the end.

So from purely the R&D perspective, the NRE is relatively low on the <$100 card market.

Also, there is no ISA lock in for GPUs, an advantage which x86 had over RISCs. Further, loads of software is going to be written for GPUs for games for a very long time to come, so anything that is written for HPC is a bonus.

I guess what I am trying to say is that GPUs are not going to vanish from the HPC market even if nobody writes HPC codes for them. RISCs had nothing like that going for them.