How much IPC?

[FruitBasketSilex]

Everyone years IPC or instructions per clock thrown around, especially in a war between amd and Intel. Someone's always saying, oh the IPC of piledriver is nowhere near that of Haswell.

My question is what are the actual numbers? How many instructions does haswell, or even all architectures (in list form) have in a clock cycle?

Two? A million? 348,000?

[Dabombinable]

Everyone years IPC or instructions per clock thrown around, especially in a war between amd and Intel. Someone's always saying, oh the IPC of piledriver is nowhere near that of Haswell.

My question is what are the actual numbers? How many instructions does haswell, or even all architectures (in list form) have in a clock cycle?

Two? A million? 348,000?

Download Performance Test and look at the baselines (they need to be downloaded off the internet inside PT) and look for scores that compare the CPUs at the same clock speed. If the single threaded is score higher at the same clock speed, then you can see which CPU has the highest IPC. (An FX 8350 got owned by my Core 2 Duo E8500 at 4GHz in the single threaded test-I'll try to find the results)

[FruitBasketSilex]

Download Performance Test and look at the baselines (they need to be downloaded off the internet inside PT) and look for scores that compare the CPUs at the same clock speed. If the single threaded is score higher at the same clock speed, then you can see which CPU has the highest IPC. (An FX 8350 got owned by my Core 2 Duo E8500 at 4GHz in the single threaded test-I'll try to find the results)

That's not what I'm asking though. I know many things beat the 8350 because of its low IPC.

What I want to know is how many Instructions Per Clock that 8350 has, what has well has, what sandy bridge has, etc...

[Dabombinable]

That's not what I'm asking though. I know many things beat the 8350 because of its low IPC.

What I want to know is how many Instructions Per Clock that 8350 has, what has well has, what sandy bridge has, etc...

Ok, I'll look them up and give you the single threaded score-which is in millions of operations per second-with a higher number meaning a higher IPC. 

[tmlhalo]

The answer you're looking for isn't going to be that simple. IPC/CPI varies by instruction. There are tons of different types of instructions, varying instruction set support. Nor is that information of IPC/CPI given out by AMD and Intel.

[FruitBasketSilex]

Ok, I'll look them up and give you the single threaded score-which is in millions of operations per second-with a higher number meaning a higher IPC.

Is it the actual number of calculations or is it just a reference for how powerful it is, like a cinebench score?

[Dabombinable]

Is it the actual number of calculations or is it just a reference for how powerful it is, like a cinebench score?

Actual calculations, I'm aiming for CPUs at 4GHz, with the FX 8350 and some FX 6xxx/FX 4xxx/Phenom/Athlons for comparison.

[JustCallMeVlad]

I believe haswell has around 50% better ipc

[tmlhalo]

Is it the actual number of calculations or is it just a reference for how powerful it is, like a cinebench score?

Reference.

 

[FruitBasketSilex]

I believe haswell has around 50% better ipc

Still, not the question.

[FruitBasketSilex]

Why is this info so hard to get to?

[Dabombinable]

Why is this info so hard to get to?

Don't worry, I'm getting it now. 3 results per CPU, though its difficult finding them all at 4GHz.

[tmlhalo]

Why is this info so hard to get to?

Here is a list of instructions a CPU can do.

http://en.wikipedia.org/wiki/MIPS_instruction_set#MIPS_assembly_language

Loads, stores, addition. all take varying amounts of time to complete as they have differing IPC/CPI. You can get and idea an average IPC but what if the instructions consists of more loads? more Stores? 

There is the reason why we measure CPUs in GFLOPS / TFLOPS because floating point operations are usually the most time consuming calculations (worst case scenario) a processor can be doing instead of say integer. You can get averages from mixed types of instructions but in reality there is more than one IPC/CPI as there are instructions with varying complexity.

[FruitBasketSilex]

Here is a list of instructions a CPU can do.

http://en.wikipedia.org/wiki/MIPS_instruction_set#MIPS_assembly_language

Loads, stores, addition. all take varying amounts of time to complete as they have differing IPC/CPI. You can get and idea an average IPC but what if the instructions consists of more loads? more Stores?

There is the reason why we measure CPUs in GFLOPS / TFLOPS because floating point operations are usually the most time consuming calculations (worst case scenario) a processor can be doing instead of say integer. You can get averages from mixed types of instructions but in reality there is more than one IPC/CPI as there are instructions with varying complexity.

So maybe a better question is what is the gflop/tflop capability of a core of an 8350 and a core of a 4690k and many other CPUs.

[Dabombinable]

So maybe a better question is what is the gflop/tflop capability of a core of an 8350 and a core of a 4690k and many other CPUs.

Yep, and I was just going through the benchmarks-it took me a while to realize that V7 of Performance test no longer has its single threaded results, and version 6 does not want to install for me.

[tmlhalo]

So maybe a better question is what is the gflop/tflop capability of a core of an 8350 and a core of a 4690k and many other CPUs.

Here is the best link I could find at the moment. Second row is gflop rating.

http://www.overclockerscave.com/index.php/96-hot-hardware/172-amd-fx-8350-vishera-full-review-and-benchmarks

[QueenDemetria]

Everyone years IPC or instructions per clock thrown around, especially in a war between amd and Intel. Someone's always saying, oh the IPC of piledriver is nowhere near that of Haswell.

My question is what are the actual numbers? How many instructions does haswell, or even all architectures (in list form) have in a clock cycle?

Two? A million? 348,000?

I'm not sure it can be properly measured. For testing, the CPU's would have to be at the same exact speed, and not change at all during the test. Setting the speeds the same would be easy, but making sure they stay at that speed would be impossible(since look at you motherboard now, it will often vary by a few mhz if not more). Another issue is the operating system, since something like Windows would throw off the results for being too heavy, and lightweight OS's like Memtest(as an example) could vary in heavyness from other operating systems. Basically, you can't tell processor A and B apart by running benchmark X, but as far as "my processor has X ammount of IPC" no one would ever know. Maybe you could do the math and estimate the IPC with the instruction sets like SSE1,2,3,4, AES, VT-x, etc. and use architecture data sheets to find its efficiency, but even then that would be a theoretical result.

[szomeret]

It's hard because the IPC of each processor varies depending on workload.

Haswell arch is capable of decoding 5 instructions, issuing 4 fused uops and dispatching 8 uops each cycle. Each x86/x86-64 instruction can be broken into one or more uop. Bulldozer is also, hypothetically, capable of issuing 4 Micro-ops each cycle, but it cannot decode nor issue as many per cycle and has much worse cache timings, so that more cycles are wasted fetching from memory.

If you're curious, read David Kanter's articles from where these images originate: http://www.realworldtech.com/bulldozer/vs. http://www.realworldtech.com/haswell-cpu/ There is also Agner Fog's x86 Micro-architecture guide (#3).

[Deletive]

Sandra Whetstone: ( FPU GFLOPS) 8350 : 74 and 3570k : 95

Sandra Dhrystone: (ALU GIPS) 8350 : 114 and 3570k : 126

GIPS = Giga instructions per second

GFLOPS: Giga Floating-point Operations Per Second

Nothing professional but will give you a just of everything

http://www.overclockerscave.com/index.php/96-hot-hardware/172-amd-fx-8350-vishera-full-review-and-benchmarks

[FruitBasketSilex]

It's hard because the IPC of each processor varies depending on workload.

Haswell arch is capable of decoding 5 instructions, issuing 4 fused uops and dispatching 8 uops each cycle. Each x86/x86-64 instruction can be broken into one or more uop. Some uops are simply instructions for the 256-bit wide FPU. Bulldozer is also, hypothetically, capable of issuing 4 uops each cycle, but it cannot decode nor issue as many per cycle.

This shit gets too complex for my head at 1:35 am.

[FruitBasketSilex]

Sandra Whetstone: ( FPU GFLOPS) 8350 : 74 and 3570k : 95

Sandra Dhrystone: (ALU GIPS) 8350 : 114 and 3570k : 126

GIPS = Giga instructions per second

GFLOPS: Giga Floating-point Operations Per Second

Nothing professional but will give you a just of everything

http://www.overclockerscave.com/index.php/96-hot-hardware/172-amd-fx-8350-vishera-full-review-and-benchmarks

Too tired. Brain forgot what giga was.

[tmlhalo]

This shit gets too complex for my head at 1:35 am.

I warned you in my first post lol.

[tmlhalo]

Too tired. Brain forgot what giga was.

Giga is 10^9

[Deletive]

Too tired. Brain forgot what giga was.

In billions or

Giga is 10^9

[FruitBasketSilex]

I warned you in my first post lol.

I need more explanation. At best, draw it out really long and I can reread it tommorow before work. At worst. Explain it like I'm really tired. (Lol)