3 ms·

Those are tall words for describing a chip with... checks clipboard ...four discrete hardware-based SIMD accelerators.

PaulHoule 1y ago
Which one are we talking about? Apple is successful at getting the capabilities of their latest chips used in library code that applications use. For instance, back when Apple was running the legacy x86 architecture they packed in the Intel MKL libraries but Microsoft and Linux distributions didn't. MacOS comes with libraries that use the accelerators and applications often use those libraries so I've got some confidence that I'll actually use an accelerator that is on the M5 chip. In Wintel my expectation is that they will include some accelerator or other feature (say TSX or SGX) but it might be fused off or there won't be actual OS or library or application support or we might find out a year or two later that the feature doesn't actually work or leaks information. It's a nightmare we've been living for more than a decade. The fixed-width SIMD model that we're stuck with is a regression from the vector processors of the 1980s in that you have to rewrite your code with each generation of the instruction set. Couple that with Intel thinking AVX512 is just a "national labs and hyperscaler" feature but still adding it into consumer chips with it fused off -- too many people think this is business genius but it is not... if the customer isn't paying for the manufacturing of those fused off features the shareholder is. They would be better off not putting them in so ordinary customers get cheaper chips and telling the national labs they have to build a 2x bigger cluster.
- bigyabai 1y ago
  > Which one are we talking about? You tell me. AVX and SSE are at-most two discrete hardware blocks and more often one. Apple Silicon, in it's genius RISC glory, possesses: 1) A Neural Engine 2) ARM NEON as Accelerate 3) Generic GPU SIMD as MPS 4) MatMul units in M5+ hardware It's an impressive amount of fragmentation for a single vendor that supposedly knows what they're doing. Even Nvidia doesn't suffer from this level of identity crisis, despite burning the compute/raster candle at both ends. Apple's approach explains their NASDAQ performance but not much else. It's a bit baffling.

Which one are we talking about? Apple is successful at getting the capabilities of their latest chips used in library code that applications use. For instance, back when Apple was running the legacy x86 architecture they packed in the Intel MKL libraries but Microsoft and Linux distributions didn't. MacOS comes with libraries that use the accelerators and applications often use those libraries so I've got some confidence that I'll actually use an accelerator that is on the M5 chip. In Wintel my expectation is that they will include some accelerator or other feature (say TSX or SGX) but it might be fused off or there won't be actual OS or library or application support or we might find out a year or two later that the feature doesn't actually work or leaks information. It's a nightmare we've been living for more than a decade. The fixed-width SIMD model that we're stuck with is a regression from the vector processors of the 1980s in that you have to rewrite your code with each generation of the instruction set. Couple that with Intel thinking AVX512 is just a "national labs and hyperscaler" feature but still adding it into consumer chips with it fused off -- too many people think this is business genius but it is not... if the customer isn't paying for the manufacturing of those fused off features the shareholder is. They would be better off not putting them in so ordinary customers get cheaper chips and telling the national labs they have to build a 2x bigger cluster.

bigyabai 1y ago
> Which one are we talking about? You tell me. AVX and SSE are at-most two discrete hardware blocks and more often one. Apple Silicon, in it's genius RISC glory, possesses: 1) A Neural Engine 2) ARM NEON as Accelerate 3) Generic GPU SIMD as MPS 4) MatMul units in M5+ hardware It's an impressive amount of fragmentation for a single vendor that supposedly knows what they're doing. Even Nvidia doesn't suffer from this level of identity crisis, despite burning the compute/raster candle at both ends. Apple's approach explains their NASDAQ performance but not much else. It's a bit baffling.