3 ms·

You're probably joking, but the "we need CPUs that are optimized for our high-level languages"-argument has been raised again and again in all seriousness in th

by adwn 11y ago

You're probably joking, but the "we need CPUs that are optimized for our high-level languages"-argument has been raised again and again in all seriousness in the past, without anyone ever showing a convincing way to speed up Javascript, Python, etc. in hardware. Fact is, chasing pointers through small objects or looking up methods in hash tables is very cache-unfriendly, and memory latency is the bottleneck for many computational tasks today. Instead, we should work on improved languages that provide better compromises between performance and convenience. Rust, for example, is a good step in this direction.

marcosdumay 11y ago
Well, whatever language you want to run, processors that are good for cache unfriendly tasks would be huge for logical AI, data mining, some kinds of image processing, and possibly a lot of other tasks. As dies get bigger, there's no law of nature saying that CPU power must be concentrated far from memory.
- adwn 11y ago
  > As dies get bigger, there's no law of nature saying that CPU power must be concentrated far from memory. Actually, there is. The main reason why the L1 cache is so small, is that a larger cache would take more cycles to access. The L2 cache is allowed to have a higher latency, so the electrical signals have more time to propagate through a physically larger area. 3D stacked dies alleviate this problem, and the trend goes towards putting several dies side-by-side in a single package (a chip), enabling lower latencies, higher throughput, and less power consumption.
  - marcosdumay 11y ago
    Yes, if you concentrate all your CPU power on a single region, it'll be far from most of the memory. There's natural law that mandates the CPU power to be concentrated.
simonster 11y ago
There's http://www.azulsystems.com/products/vega/processor http://www.azulsystems.com/products/vega/processor, which is a CPU with special GC instructions, created by the same company as one of the most sophisticated garbage collectors. Apparently they thought they could speed things up with specialized hardware. Not sure what the performance looks like in practice, but it is interesting that it exists...
- ghaff 11y ago
  Azul had/has some interesting technology. (They're still around but just do software now.) I had some conversations on this topic broadly at a conference I was speaking at a month or so ago which had more hardware/software overlap than is typical. IMO, we're unlikely to come up with anything else in the near-term that delivered the kind of scaling CMOS process technology did. As a result, it's at least possible that we'll see specialized hardware make more sense in a world where "Moore's Law" won't let general purpose x86 chips just hit the same performance level in a couple of years.

by adwn 11y ago

marcosdumay 11y ago

Well, whatever language you want to run, processors that are good for cache unfriendly tasks would be huge for logical AI, data mining, some kinds of image processing, and possibly a lot of other tasks. As dies get bigger, there's no law of nature saying that CPU power must be concentrated far from memory.

adwn 11y ago
> As dies get bigger, there's no law of nature saying that CPU power must be concentrated far from memory. Actually, there is. The main reason why the L1 cache is so small, is that a larger cache would take more cycles to access. The L2 cache is allowed to have a higher latency, so the electrical signals have more time to propagate through a physically larger area. 3D stacked dies alleviate this problem, and the trend goes towards putting several dies side-by-side in a single package (a chip), enabling lower latencies, higher throughput, and less power consumption.
- marcosdumay 11y ago
  Yes, if you concentrate all your CPU power on a single region, it'll be far from most of the memory. There's natural law that mandates the CPU power to be concentrated.

simonster 11y ago

There's http://www.azulsystems.com/products/vega/processor http://www.azulsystems.com/products/vega/processor, which is a CPU with special GC instructions, created by the same company as one of the most sophisticated garbage collectors. Apparently they thought they could speed things up with specialized hardware. Not sure what the performance looks like in practice, but it is interesting that it exists...

ghaff 11y ago
Azul had/has some interesting technology. (They're still around but just do software now.) I had some conversations on this topic broadly at a conference I was speaking at a month or so ago which had more hardware/software overlap than is typical. IMO, we're unlikely to come up with anything else in the near-term that delivered the kind of scaling CMOS process technology did. As a result, it's at least possible that we'll see specialized hardware make more sense in a world where "Moore's Law" won't let general purpose x86 chips just hit the same performance level in a couple of years.