Which would be faster theoretically?

[Maximilian]

Ok, this is another one of those questions that dosent really serve a purpose but it would be interesting to know.

Which would be faster out of these two CPU's:

a 5ghz 486 - say its been given all the instructions that the athlon has, SSE etc
an athlon 64 3200+ 2.2ghz venice

So one uses a crappy by todays standards architecture but has a brute clock speed advantage, and the other is modern but with a much lesser clockspeed. Heh that scenario sounds familier eh? But this isnt a P4 its a 486.

I would like to think that with a 2.8ghz clock speed advantage the 486 would win but no idea tbh. What do you guys think?

 

[Furen]

The K8, of course. All I have to say is superscalar, out-of-order design... and enough memory bandwidth to keep it fed...

 

[F1shF4t]

Originally posted by: Soviet
Ok, this is another one of those questions that dosent really serve a purpose but it would be interesting to know.

Which would be faster out of these two CPU's:

a 5ghz 486 - say its been given all the instructions that the athlon has, SSE etc
an athlon 64 3200+ 2.2ghz venice

So one uses a crappy by todays standards architecture but has a brute clock speed advantage, and the other is modern but with a much lesser clockspeed. Heh that scenario sounds familier eh? But this isnt a P4 its a 486.

I would like to think that with a 2.8ghz clock speed advantage the 486 would win but no idea tbh. What do you guys think?

Theoreticaly a p4 should be able to blow the athlon 64 out of the water, see all the synthetic benchmarks it displays an ideal situation, problem is when the situation is not ideal the p4 gets wasted. A design is only as good as its weakest link in that case branchy code and long pipeline, lots of mispredicts and pipeline stalls.

Now if u only count the execution core of the cpu if the clock speed can counter the ineficiency and the lower through put per clock cycle of the 486, than it could be quicker than the athlon 64.

Now if u take into account the old cache system and other things, it does not matter what speed the 486 will be as the system will not be able to keep the core buisy.

Remeber the celeron northwoods, why a celeron northwood 2.4ghz would be easily beaten by a pentium 4 2.4ghz 400mhz fsb, cause the cashe is not large enough to keep the cpu buisy when it comes to complex programs like games, so its much slower. So at that point any increase in clockspeed after a point will not increase its performance same thing applies to the p4 but that will be at higher clock speed.

so answer is yes and no, depends on the rest of the system.

 

[Maximilian]

Hmm interesting. Thanks for that

 

[Furen]

The 486 will NEVER be faster than a K8.

Let's assume you look at only the execution core, and that you have no branches and only execute instructions that both CPUs can do in a single clock (let's say, for example, an ALU add) and that there WILL be instructions to throw at the execution units every clock. The 486 will be able to execute 1 per clock, while the K8 will be able to execute 3 per clock (3-issue, superscalar design) so assuming the maximum theorethical peak rates of both, the 486 will hit 5GI/s while the K8 will hit 6.6GI/s. If you have more realistic code with branches, condition, etc (and code that you can micro-op fuse) then the difference becomes even larger because of Out of Order execution, branch prediction, micro-op fusion, etc. If you look at the CPU as a whole then it's not even close, the 486 will be insanely bottlenecked by its caches (let's assume that you have 1.25GHz bus with enough ram bandwidth at its disposal to saturate it, you'll still be hitting main memory a lot).

 

[zephyrprime]

The K8 would blow the 486 out of the water.
A 486 takes 12-42 cycles to do an integer multiply. A K8 takes 4 cycles. And the 486 cannot accept 1 multiply every cycle like the K8 can. I think it could only take 1 every 12 cycles tops. The same sort of thing applies to every instruction that takes more than 1 cycle to execute on the 486 since it only has 1 pipeline stage devoted to execution on its 4 stage pipeline. (Simpler instructions take less time of course.) So you see, the clockspeed advantage of the 486 is more than negated by it high instruction latencies.

Then there's the FPU. Atrocious. Add takes 8-20 cycles on a 486; 1 cycle on a K8. Multiply: 16/486, 4/K8.

Then there are the cache issues, lack of superscaler, and lack of OOE. Also, the 486 only had a 32bit FSB. The one advantage the 486 has is that its pipeline is so short that its L1 cache latency is just 1 cycle and lacking a branch predictor doesn't hurt is so much.

The nice thing about the 486 is that it's so small. Only 1.2M transistors. All the extra features we add to processors are hugely expensive in transistor terms. Features like OOE and branch prediction are very inefficient relative to cost. That's why there's currently a back to basics move as evidenced by processors like the Niagra from Sun.