3 ms·

Ultimately if all you need to use the CPU for is After Effects CC then follow the results of that benchmark (or rather, follow the performance/$ for that specif

by d1zzy 9y ago

Ultimately if all you need to use the CPU for is After Effects CC then follow the results of that benchmark (or rather, follow the performance/$ for that specific benchmark). Otherwise, in general the AMD CPUs give more multi-core performance/$ than Intel, for example, in any multi-core x264 based test I've seen. And it makes sense why that is so, the AMD CPU has more cores and more cache for the same $. And that's only comparing CPU prices and performance, if we consider the motherboard prices then the needle moves further into AMD's favor. As to why After Effects CC seems to prefer Intel so much would indicate some specific optimizations Adobe performs, optimizations that seem to favor Intel for now, not general raw multi-core performance of the Intel CPU. For single core performance the situation is reversed, the higher clocked (but lower core count) Intel CPUs will obviously have much better single core performance which is important for games as even today they tend to be coded relying on a few threads.

dragontamer 9y ago
> As to why After Effects CC seems to prefer Intel so much would indicate some specific optimizations Adobe performs, optimizations that seem to favor Intel for now, not general raw multi-core performance of the Intel CPU. I bet its AVX-512. Which is "Intel Specific" but that's the entire point of AVX-512, a new Intel specific instruction set that has high performance benefits. Overall, AMD is betting on general purpose cores with at most 128-bits per operation. AMD supports the AVX 256-bit instruction set, but its "emulated", as the ALU pipelines are 128-bit (granted: AMD can gang together two ALUs in a 2x128-bit configuration to execute the 256-bit instructions) Intel supports 3x256-bits per operation on their CPUs, with the 512-bit AVX512 instructions using two execution ports at a time. So for any code that's written to be SIMD-heavy, Intel machines will likely be faster. Not only because of AVX512, but also because Intel's AVX-256 bit instructions are better supported compared to AMD (3x256-bit for Intel, vs 2x128-bit for AMD) ------------ With that being said: the benefits of SIMD approach Ahmdal's law just like anything else. In particular, AVX512 uses so much power that Intel cores are forced to downclock. So AMD can run more cores, at higher frequencies. But Intel's cores do more work per clock tick and per unit energy. It seems like AMD's 16-cores (each doing 2x128-bit) are only slightly slower than Intel's 10-cores (each doing 3x256-bits) in those SIMD-heavy tests.
- Dylan16807 9y ago
  > (3x256-bit for Intel, vs 2x128-bit for AMD) It's not quite that harsh for workloads that aren't entirely FMA. Zen has 2x128-bit adders plus 2x128-bit multipliers.
  - dragontamer 9y ago
    Fair enough. The full AMD pipeline contains 4x 128-bit execution ports. But this is a bit complicated, so I was ignoring that part of the core. IMO, the more important bottleneck is the load/store unit. Intel's system has 2x 256-bit loads + 1x 256-bit writes to the L1 cache. While AMD's system has 2x128-bit load+store units for its L1 cache. Thus my numbers above: 3x256-bit for Intel (2 reads + 1 write), while 2x128-bits for AMD (2x read-or-write).

by d1zzy 9y ago

dragontamer 9y ago

> As to why After Effects CC seems to prefer Intel so much would indicate some specific optimizations Adobe performs, optimizations that seem to favor Intel for now, not general raw multi-core performance of the Intel CPU. I bet its AVX-512. Which is "Intel Specific" but that's the entire point of AVX-512, a new Intel specific instruction set that has high performance benefits. Overall, AMD is betting on general purpose cores with at most 128-bits per operation. AMD supports the AVX 256-bit instruction set, but its "emulated", as the ALU pipelines are 128-bit (granted: AMD can gang together two ALUs in a 2x128-bit configuration to execute the 256-bit instructions) Intel supports 3x256-bits per operation on their CPUs, with the 512-bit AVX512 instructions using two execution ports at a time. So for any code that's written to be SIMD-heavy, Intel machines will likely be faster. Not only because of AVX512, but also because Intel's AVX-256 bit instructions are better supported compared to AMD (3x256-bit for Intel, vs 2x128-bit for AMD) ------------ With that being said: the benefits of SIMD approach Ahmdal's law just like anything else. In particular, AVX512 uses so much power that Intel cores are forced to downclock. So AMD can run more cores, at higher frequencies. But Intel's cores do more work per clock tick and per unit energy. It seems like AMD's 16-cores (each doing 2x128-bit) are only slightly slower than Intel's 10-cores (each doing 3x256-bits) in those SIMD-heavy tests.

Dylan16807 9y ago
> (3x256-bit for Intel, vs 2x128-bit for AMD) It's not quite that harsh for workloads that aren't entirely FMA. Zen has 2x128-bit adders plus 2x128-bit multipliers.
- dragontamer 9y ago
  Fair enough. The full AMD pipeline contains 4x 128-bit execution ports. But this is a bit complicated, so I was ignoring that part of the core. IMO, the more important bottleneck is the load/store unit. Intel's system has 2x 256-bit loads + 1x 256-bit writes to the L1 cache. While AMD's system has 2x128-bit load+store units for its L1 cache. Thus my numbers above: 3x256-bit for Intel (2 reads + 1 write), while 2x128-bits for AMD (2x read-or-write).