4 ms·

Are smartphone architectures really this amazing, or is Intel just dropping the ball so hard that they're getting lapped by mobile hardware?

danaris 6y ago
From what I've seen, it's a combination of Intel dropping the ball hard, and Apple aggressively pushing to make better and better chips for the iPhone and iPad—and, now, for their computers. This isn't something that you're going to see in "mobile hardware" more generally, I don't think, at least not unless Intel continues to be unable to find its balls for a few more years.
bgorman 6y ago
The smartphone architecture is smart and well executed, but it is nothing revolutionary. The key is definitely that Intel has lost its manufacturing lead, which is letting other chip companies release competitive CPU architectures. Moore's law ending has leveled the playing field for ASICs.
PaulHoule 6y ago
Intel has focused most on data center performance, where you are operating at scale and frequently using entirely different algorithms than people usually use on a PC or Phone. For instance https://en.wikipedia.org/wiki/Column-oriented_DBMS https://en.wikipedia.org/wiki/Column-oriented_DBMS dominates the competition for ad-hoc analytics under "enterprise" conditions. There the point is to maximize the use of memory bandwidth and never stall the pipelines, to turn a commercial data processing job into something that looks like HPC to the memory bus. (Actually columnar organization improves performance reliably for small data; video game programmers in the 1980s knew that it was better to put the x coordinates of all the spaceships together, then put the y coordinates together, then put the sprite id's together, ...) Most muggles though go with the row-orientation that comes with structs, classes, record types, and don't care. Pythoners know their language is slow but they know enough to use 'pandas' which is an object-oriented interface to an in-memory column store. Intel has sold it's performance on the basis of what a sophisticated organization could get out of complex programming. For instance, the AVX-512 instructions can nearly double performance for some workloads. If you are a cloud provider that will rack 1000s of identical machines that run the same software, it can be a big win. The average PC user runs software that has to run on a wide range of hardware and may be supported by lower-tier organizations. They will rightly privilege having a trouble-free product and low support costs vs giving every user the highest performance possible.)
- robertlagrant 6y ago
  Columnar is just a specialisation, like graph or document. It's not a general-purpose improvement on relational.
  - PaulHoule 6y ago
    From the viewpoint of relational it is an execution strategy. Many columnar databases return row wise results via odbc or some similar interface and they look like relational databases. The difficulties of doing columnar style queries for document and graph databases led at least one person I know to tell a certain three letter agency that they couldn't have the panopticon machine they wanted. For online transaction processing rows are good, but for OLAP it is not even competitive.
- tomaskafka 6y ago
  This - consumers basically get 'leftovers' of design made for datacenters, while Apple will probably optimize for consumer use cases.

danaris 6y ago

From what I've seen, it's a combination of Intel dropping the ball hard, and Apple aggressively pushing to make better and better chips for the iPhone and iPad—and, now, for their computers. This isn't something that you're going to see in "mobile hardware" more generally, I don't think, at least not unless Intel continues to be unable to find its balls for a few more years.

bgorman 6y ago

The smartphone architecture is smart and well executed, but it is nothing revolutionary. The key is definitely that Intel has lost its manufacturing lead, which is letting other chip companies release competitive CPU architectures. Moore's law ending has leveled the playing field for ASICs.

PaulHoule 6y ago

Intel has focused most on data center performance, where you are operating at scale and frequently using entirely different algorithms than people usually use on a PC or Phone. For instance https://en.wikipedia.org/wiki/Column-oriented_DBMS https://en.wikipedia.org/wiki/Column-oriented_DBMS dominates the competition for ad-hoc analytics under "enterprise" conditions. There the point is to maximize the use of memory bandwidth and never stall the pipelines, to turn a commercial data processing job into something that looks like HPC to the memory bus. (Actually columnar organization improves performance reliably for small data; video game programmers in the 1980s knew that it was better to put the x coordinates of all the spaceships together, then put the y coordinates together, then put the sprite id's together, ...) Most muggles though go with the row-orientation that comes with structs, classes, record types, and don't care. Pythoners know their language is slow but they know enough to use 'pandas' which is an object-oriented interface to an in-memory column store. Intel has sold it's performance on the basis of what a sophisticated organization could get out of complex programming. For instance, the AVX-512 instructions can nearly double performance for some workloads. If you are a cloud provider that will rack 1000s of identical machines that run the same software, it can be a big win. The average PC user runs software that has to run on a wide range of hardware and may be supported by lower-tier organizations. They will rightly privilege having a trouble-free product and low support costs vs giving every user the highest performance possible.)

robertlagrant 6y ago
Columnar is just a specialisation, like graph or document. It's not a general-purpose improvement on relational.
- PaulHoule 6y ago
  From the viewpoint of relational it is an execution strategy. Many columnar databases return row wise results via odbc or some similar interface and they look like relational databases. The difficulties of doing columnar style queries for document and graph databases led at least one person I know to tell a certain three letter agency that they couldn't have the panopticon machine they wanted. For online transaction processing rows are good, but for OLAP it is not even competitive.
tomaskafka 6y ago
This - consumers basically get 'leftovers' of design made for datacenters, while Apple will probably optimize for consumer use cases.