3 ms·

Fair enough, but the 5960X is still a better comparison, surely?

by phaemon 12y ago

bhouston 12y ago
I was trying to make the point that per core speed hasn't increased. Both the 980X and the 5960K are the top of the line 6 core consumer grade CPUs. This ties in with my point that the consumer grade CPUs from Skylake are all 2 and 4 core CPUs, which isn't that different from low-end consumer grade CPUs from 2010/2011, because per core performance hasn't increased significantly. My solution was for Intel to increase core count for consumer grade CPUs. Because you can get real performance increases this way, as you proved by comparing a 6 core CPU with an 8 core CPU.
- Sanddancer 12y ago
  At the high end, per-core performance hasn't increased significantly. However, at the 2-4 core level, it has. For example, the i3-4360t is 72% faster than the i3 550, even though the 4360t uses half the power of the 550. Plus, you need to keep in mind other things that will affect perceptions of performance, like the pci-e lanes being on-die with the most recent generations of chips. While the high end has plateaued to some extent with recent generations, Intel's pumped a significant amount of work making their other lines considerably more performant.
  - vardump 12y ago
    > At the high end, per-core performance hasn't increased significantly. When comparing aforementioned CPUs (Intel Core i7 980X and Intel Core i7 5930K), I'd call 4->16 FLOPS per clock and 25.6 GB/s -> 68 GB/s memory bandwidth significant. I think integer rate has also at least quadrupled. Of course, because of high latencies, when running pointer chasing or branchy code, there's little improvement. The software just can't take advantage of hardware capabilities. What kind of software are you considering when making that statement?
- vardump 12y ago
  More cores buys you only so much unless memory bandwidth and preferably number of memory subsystems is also increased (NUMA sockets). It's often easy to consume all L3/memory bandwidth with just one core. A single Haswell core can load over 150 GB of data per second, 64 bytes per cycle.

Fair enough, but the 5960X is still a better comparison, surely?

by phaemon 12y ago

Fair enough, but the 5960X is still a better comparison, surely?

I was trying to make the point that per core speed hasn't increased. Both the 980X and the 5960K are the top of the line 6 core consumer grade CPUs. This ties in with my point that the consumer grade CPUs from Skylake are all 2 and 4 core CPUs, which isn't that different from low-end consumer grade CPUs from 2010/2011, because per core performance hasn't increased significantly. My solution was for Intel to increase core count for consumer grade CPUs. Because you can get real performance increases this way, as you proved by comparing a 6 core CPU with an 8 core CPU.

Sanddancer 12y ago
At the high end, per-core performance hasn't increased significantly. However, at the 2-4 core level, it has. For example, the i3-4360t is 72% faster than the i3 550, even though the 4360t uses half the power of the 550. Plus, you need to keep in mind other things that will affect perceptions of performance, like the pci-e lanes being on-die with the most recent generations of chips. While the high end has plateaued to some extent with recent generations, Intel's pumped a significant amount of work making their other lines considerably more performant.
- vardump 12y ago
  > At the high end, per-core performance hasn't increased significantly. When comparing aforementioned CPUs (Intel Core i7 980X and Intel Core i7 5930K), I'd call 4->16 FLOPS per clock and 25.6 GB/s -> 68 GB/s memory bandwidth significant. I think integer rate has also at least quadrupled. Of course, because of high latencies, when running pointer chasing or branchy code, there's little improvement. The software just can't take advantage of hardware capabilities. What kind of software are you considering when making that statement?
vardump 12y ago
More cores buys you only so much unless memory bandwidth and preferably number of memory subsystems is also increased (NUMA sockets). It's often easy to consume all L3/memory bandwidth with just one core. A single Haswell core can load over 150 GB of data per second, 64 bytes per cycle.