|
|
|
|
|
by chancho
5310 days ago
|
|
> what happens in HPC tends to filter down to servers Is this conventional wisdom? How does a petaflop race affect app servers and databases? It seems like most traditional server workloads could get by without a single FPU. The only thing they have in common is IO. Are there many data centers using Infiniband? (Maybe there are I don't know.) The Cell architecture is an evolutionary dead end. SPARC is no more of a threat to X86 now than before. GPUs may be the next big thing for HPC but its got a long way to go to get out of its niche in the server market. (That niche being... face detection for photo sharing sites? Black-Scholes? Help me out here.) I mean, I agree with your overall point, but I think it's more likely that ARM will steal all the data center work before anything from the HPC world does. They are too focused on LINPACK. |
|
I'm not quite sure it's valid to write off SPARC as an architectural dead end when the current fastest computer in the world uses it, and the next crop of US competitors for that crown are all based on the Cell/BlueGene lineage. GPUs are also more broadly applicable than you might think. Besides video and audio processing, they can be used for many crypto-related tasks (witness their popularity for Bitcoin mining), various kinds of math relevant to data storage (e.g erasure codes or hashes for dedup), and so on. Many of their architectural features are also being copied by more general-purpose processors as core counts increase, as well.
Yes, high-end HPC is too obsessed with LINPACK. Nonetheless, it remains a good place to look when trying to predict the future of commodity servers. Even if ARM does displace x86 instead, many features besides the ISA are likely to come from HPC. Perhaps more relevantly, either outcome is still very bad for Intel.