3 ms·
I wish people would quit asking this question. This is just another case of the Mhz myth. Who cares what the clock speed is if modern CPUs can execute instructi
by DiabloD3 15y ago
I wish people would quit asking this question. This is just another case of the Mhz myth. Who cares what the clock speed is if modern CPUs can execute instructions 2-4x faster per clock than they did 30 years ago.
Go look at the Bulldozer design, 2 hardware decoder/scheduler engines, 4 integer ALUs, 2 fp ALUs, all per core...
A shared L3 cache per socket that is owned by the memory controller and is socket-local to other sockets (ie, all memory controllers conspire to cache system memory efficiently and synchronously know what is cached across all sockets)...
And the memory controllers also accept memory requests from ANY core on the Hypertransport bus, no matter which socket, and multi-socket boards commonly have one memory bank per socket, thus 4 sockets of dual channel DDR3-1600 would indeed give 820 gbit/sec of bandwidth that can be accessed (almost) in full by any individual core[1]...
The ALUs have execution queues, and any thread (currently 2 per core) can schedule instructions on it to maximize ALU packing...
And you can now buy Bulldozers for Socket G34 that have 16 threads/8 cores per socket, and G34 boards usually have 4 sockets.
So again, who cares what the mhz is?
[1]: A 4 socket setup has 4 or 6 Hypertransport links, on AM3+ and G34+ sockets this would be HTX 3.1 16bit wide links running at 3.2ghz, or 204 gbit/sec per link.
On a 4 socket ring, that would be 204 gbit/sec off each neighbor's memory bank, plus another 204 gbit/sec (also the speed of dual channel DDR3-1600) from the local memory bank, thus leading to 612 gbit/sec that could be theoretically saturated by a single core.
On a 4 socket full crossbar, it would be the full 820 gbit/sec.
- dchest 15y agoWho cares what the clock speed is if modern CPUs can execute instructions 2-4x faster per clock than they did 30 years ago. If you could increase clock speed on these modern processors twofold, would their performance improve?
- jlouis 15y agoNot necessarily. The fact is that you could probably do just exactly that by making the pipeline deeper. But making the pipeline deeper has a price: your control logic becomes much more complex and emptying the pipeline is more expensive. The question is then: does the price you pay become larger than the gain. As this is not happening I have a good guess :) Historically, the Pentium 4 is an example of a Mhz-oriented design. The pipeline was made ridiculously deep. This meant that when you had heavy floating point intensive work, then the Pentium 4 was very very fast. Essentially this was because you could fill up the pipeline in the CPU and you have relatively few jumps in that code. It was just a processing job. On the other hand, the P4 suffered for almost any other task. Even to the point where a more beefy Pentium 3 (Tualatin) could beat it. Intel chose a path which was based on the P3 design for their Pentium-M CPU and that is the design which guided Core, Core 2 Duo, Core i7 and so on. So it is possible, but it is not a given it would lead to faster CPUs. Mhz is but one of the knobs you can turn. The problem we are facing is that all the parallelism we can squeeze out of a sequential program has been done. Modern CPUs dynamically parallelizes what can be run concurrently. But to gain more, now, you need to write programs in a style where they expose more parallelism. Welcome to parallel programming :)
- DiabloD3 15y agoYes, if you increase memory bandwidth twofold and decreased memory latency twofold and increased the CPU bus's speed twofold. If you don't, you'll be hacked off at the kneecaps by lack of memory bandwidth.
- wmf 15y agoWho cares what the clock speed is if modern CPUs can execute instructions 2-4x faster per clock than they did 30 years ago. I think we're more concerned with 3 years ago; if clock doesn't increase and IPC increases very slowly, that means new processors aren't faster enough to be worth buying. Go look at the Bulldozer design Indeed; look at the fact that it's not faster than its predecessor for many workloads.
- DiabloD3 15y agoBut the workloads when it is faster? I'm very impressed with AMD's offering. Have code thats largely complex branch and loop conditions that bust the branch predictor's balls and isn't cache friendly and is largely integers? Bulldozer eats it for breakfast and asks for more.