cpu Best single-threaded CPU performer?

[Rythmium]

I'm building a bit of an interesting computer right now. Its workload will be primarily single-threaded, but eventually that will change so it needs multithreaded capabilities for the future. However, I'm just focused on the single-threaded performance right now.

 

The Question: What is the CPU with the best single-thread performance available right now for under $500?

 

My thoughts on the answer:

 

1. My first inclination was to get a Pentium G3258, overclock it through the roof, and be done with it. But I need more cores for the future. So dual-core is out.

 

2. My next inclination was an i7-6700K. With a base clock of 4GHz and 8MB of L3 cache, while at the top of the pile of Intel's current-gen architecture, it's gotta have the best performance in every area, right?

 

3. That brings me to my third inclination. Since an i5-6600K has 4 cores and 4 threads, despite being clocked lower and having less cache than an i7, wouldn't it be more powerful per-thread than the i7, which has 8 threads?

 

I know there are lots of very knowledgeable people on this forum, so bring your knowledge! Benchmarks or theoretical technical insight are both greatly appreciated. I've been an avid watcher of LTT for the past couple years, but this is my first time on the forum. Thanks in advance!

[cortexcortex]

Check this:

 

http://cpu.userbenchmark.com/Compare/Intel-Core-i7-6700K-vs-Intel-Core-i5-6600K/3502vs3503

[HelplmChoking]

Probably 6700k, with a good overclock. Each new generation brings an IPC improvement, allowing the 4.0GHz to do more than the same clock speed on an older chip.  The i5 would likely be equal in single core performance to the i7, as they're basically the same, but the lower cache can make day to day tasks a little less snappy. The Pentium is way out, as even with a very high overclock it's very old architecture won't get as far as newer stuff. What sort of workload will this machine be under?

 

Also, welcome to the forum! Always great to have new people here

[TheRandomness]

1 minute ago, Rangaman42 said:

Probably 6700k, with a good overclock. Each new generation brings an IPC improvement, allowing the 4.0GHz to do more than the same clock speed on an older chip.  The i5 would likely be equal in single core performance to the i7, as they're basically the same, but the lower cache can make day to day tasks a little less snappy. The Pentium is way out, as even with a very high overclock it's very old architecture won't get as far as newer stuff. What sort of workload will this machine be under?

 

Also, welcome to the forum! Always great to have new people here

Haswell refresh isn't old at all..?

[JefferyD90]

7 minutes ago, Rythmium said:

I'm building a bit of an interesting computer right now. Its workload will be primarily single-threaded, but eventually that will change so it needs multithreaded capabilities for the future. However, I'm just focused on the single-threaded performance right now.

 

The Question: What is the CPU with the best single-thread performance available right now for under $500?

 

My thoughts on the answer:

 

1. My first inclination was to get a Pentium G3258, overclock it through the roof, and be done with it. But I need more cores for the future. So dual-core is out.

 

2. My next inclination was an i7-6700K. With a base clock of 4GHz and 8MB of L3 cache, while at the top of the pile of Intel's current-gen architecture, it's gotta have the best performance in every area, right?

 

3. That brings me to my third inclination. Since an i5-6600K has 4 cores and 4 threads, despite being clocked lower and having less cache than an i7, wouldn't it be more powerful per-thread than the i7, which has 8 threads?

 

I know there are lots of very knowledgeable people on this forum, so bring your knowledge! Benchmarks or theoretical technical insight are both greatly appreciated. I've been an avid watcher of LTT for the past couple years, but this is my first time on the forum. Thanks in advance!

The PlayStation actually has the best recorded ipc of any non specialized processor.  Upwards .93 IPC

[HelplmChoking]

Just now, JefferyD90 said:

The PlayStation actually has the best recorded ipc of any non specialized processor.  Upwards .93 IPC

I thought we were sitting around 1.1 on Skylake these days? Although that may have been a relative measurement comparing to AMD..

[JefferyD90]

2 minutes ago, Rangaman42 said:

I thought we were sitting around 1.1 on Skylake these days? Although that may have been a relative measurement comparing to AMD..

You cant possible have a IPC above 1.  That means the processor produces more work than its frequency allows for.  A perfect 1.0 IPC means that for every hertz on the core clock, you get 1 operation out of it.  You cant possibly get 2 operations out of one cycle.

 

I cant remember which chip it was (Haswell or Skylake) but last I seen Intel's latest chip was sitting somewhere around .64 IPC.  AMD's Piledriver was somewhere around .48 IPC.

 

And I was being a ass about my PlayStation comment, it was only clocked at something like 66mhz, so even a 1Ghz cpu with a .10 IPC would be able to out do the PlayStation's CPU.  But my point still stands, the PlayStation is still one of the modern hardware marvels.

 

Additionally, IPC doesn't mean much unless you mention which instruction set you're referring to.  And I'm 90% sure that the Intel and AMD CPUs are based on SSE3 (on the studies that I've seen).

[HelplmChoking]

3 minutes ago, TheRandomness said:

Haswell refresh isn't old at all..?

Two years old, and the Pentium has a bunch more flaws, like only having 2 threads, only 16 PCIe lanes, only up to 32GB RAM, no turbo boost etc. I know OP wants single threaded performance, but a dual core is not future-proof at all, and won't be able to perform in many modern situations.

[HelplmChoking]

2 minutes ago, JefferyD90 said:

You cant possible have a IPC above 1.  That means the processor produces more work than its frequency allows for.  A perfect 1.0 IPC means that for every hertz on the core clock, you get 1 operation out of it.  You cant possibly get 2 operations out of one cycle.

Oh I think I get it now, I'm pretty sure the article was comparing an 8320e to a 3770k, saying that if the 8320 had an IPC of 1, the 3770k would be 1.2-1.2 ish. 

 

Side note, where the heck is my profile picture?!

[JefferyD90]

3 minutes ago, Rangaman42 said:

Oh I think I get it now, I'm pretty sure the article was comparing an 8320e to a 3770k, saying that if the 8320 had an IPC of 1, the 3770k would be 1.2-1.2 ish. 

 

Side note, where the heck is my profile picture?!

edited my comment, double check it.

[TheRandomness]

3 minutes ago, Rangaman42 said:

Two years old, and the Pentium has a bunch more flaws, like only having 2 threads, only 16 PCIe lanes, only up to 32GB RAM, no turbo boost etc. I know OP wants single threaded performance, but a dual core is not future-proof at all, and won't be able to perform in many modern situations.

Well, the difference between the IPC isn't that great overall... http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

In some benches, Skylake chips (all were at 3GHz) were beaten by Haswell or Broadwell. Also, the differences are so small they won't really make a difference in real world performance unless you're doing sciency things c: All the other things you said about the Pentium were valid except the last. Who needs turbo boost when you can OC it to 5.2Ghz? (Friend got that on his chip, your mileage may vary.) 

[HelplmChoking]

Just now, TheRandomness said:

Well, the difference between the IPC isn't that great overall... http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

In some benches, Skylake chips (all were at 3GHz) were beaten by Haswell or Broadwell. Also, the differences are so small they won't really make a difference in real world performance unless you're doing sciency things c: All the other things you said about the Pentium were valid except the last. Who needs turbo boost when you can OC it to 5.2Ghz? (Friend got that on his chip, your mileage may vary.) 

Good point about turbo boost  Most people disable it when overclocking anyway. I only have mine on since I can only OC on the BLCK.

[Godlygamer23]

16 minutes ago, JefferyD90 said:

You cant possible have a IPC above 1.  That means the processor produces more work than its frequency allows for.  A perfect 1.0 IPC means that for every hertz on the core clock, you get 1 operation out of it.  You cant possibly get 2 operations out of one cycle.

DDR produces two data transfers per cycle. A cycle literally tells you nothing about the work the CPU is doing.

[JefferyD90]

3 minutes ago, Godlygamer23 said:

DDR produces two data transfers per cycle. A cycle literally tells you nothing about the work the CPU is doing.

Let me rephrase, "clock cycle".  DDR allows for double transfers per clock cycle, one going to the CPU and one going from the CPU.  (input > memory > CPU > memory > output)

[Godlygamer23]

4 minutes ago, JefferyD90 said:

Let me rephrase, "clock cycle".  DDR allows for double transfers per clock cycle, one going to the CPU and one going from the CPU.  (input > memory > CPU > memory > output)

And in the case of Sandy Bridge, it had an instruction decode rate of 4 instructions per clock per core.

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2

[agent_x007]

9 minutes ago, Godlygamer23 said:

And in the case of Sandy Bridge, it had an instruction decode rate of 4 instructions per clock per core.

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2

Add to that multi ALU/FPU build of modern CPU's (ie. "back end") : http://www.anandtech.com/show/6355/intels-haswell-architecture/8

[porina]

42 minutes ago, Rythmium said:

3. That brings me to my third inclination. Since an i5-6600K has 4 cores and 4 threads, despite being clocked lower and having less cache than an i7, wouldn't it be more powerful per-thread than the i7, which has 8 threads?

If you really care about a single thread, then the i5 could have an advantage over the i7 when HT is on. You could end up with less important stuff going on the same physical core as what you want to run, and it will slow that down. On the i5, it has to go to another core. Now, on the i7 you can turn off HT to stop this form happening, and still benefit from the bigger L3 cache size. For tasks I do where I know HT provides no benefit, turning it off does speed it up for multiple reasons. 

32 minutes ago, JefferyD90 said:

You cant possible have a IPC above 1.  That means the processor produces more work than its frequency allows for.  A perfect 1.0 IPC means that for every hertz on the core clock, you get 1 operation out of it.  You cant possibly get 2 operations out of one cycle.

It depends a bit on how literally you use the term IPC. A less literal meaning giving more real world performance indication could be more useful.

 

Consider Haswell or later Intel CPUs, with FMA instruction. That does what it says, a single instruction will do both a multiply and add to a register containing multiple chunks of data. What if you have an older processor like Sandy Bridge? No FMA there. You need to do the multiply and add steps separately for the same end result. For Prime95 like workloads which can make use of Intel FMA, Haswell will give you 50% more performance than Sandybridge, per core, per clock (where not otherwise limited). Similarly, Skylake is 14% faster than Haswell, per core, per clock. This might not be well captured if you use IPC too strictly. It would be more useful to have some kind of performance per clock metric, which I have incorrectly called IPC myself. This would also consider the case where a single instruction can operate on multiple data. In part this is why current AMD CPUs really suck at FPU tasks, as they simply can't operate on as much data in one go compared to an Intel, and their implementation of FMA takes more clocks and negates any potential benefit from having it.

[Izumi Reina]

1 hour ago, Rythmium said:

I'm building a bit of an interesting computer right now. Its workload will be primarily single-threaded, but eventually that will change so it needs multithreaded capabilities for the future. However, I'm just focused on the single-threaded performance right now.

 

The Question: What is the CPU with the best single-thread performance available right now for under $500?

 

My thoughts on the answer:

 

1. My first inclination was to get a Pentium G3258, overclock it through the roof, and be done with it. But I need more cores for the future. So dual-core is out.

 

2. My next inclination was an i7-6700K. With a base clock of 4GHz and 8MB of L3 cache, while at the top of the pile of Intel's current-gen architecture, it's gotta have the best performance in every area, right?

 

3. That brings me to my third inclination. Since an i5-6600K has 4 cores and 4 threads, despite being clocked lower and having less cache than an i7, wouldn't it be more powerful per-thread than the i7, which has 8 threads?

 

I know there are lots of very knowledgeable people on this forum, so bring your knowledge! Benchmarks or theoretical technical insight are both greatly appreciated. I've been an avid watcher of LTT for the past couple years, but this is my first time on the forum. Thanks in advance!

Have to go with the G3258 Anniversary if you win the silicon lottery and get a decent one overclocked to atleast 4.8Ghz itll crush even the most high end CPUs in single threaded performance, dont except anything special from multi-threaded though.

[Imakuni]

1 hour ago, Rythmium said:

1. My first inclination was to get a Pentium G3258, overclock it through the roof, and be done with it. But I need more cores for the future. So dual-core is out.

How much "in the future" is that? If you expect a year or more, you could gamble on a Skylake Pentium now and get Kaby Lake i5 / i7 upgrade later.

[JefferyD90]

1 hour ago, Godlygamer23 said:

And in the case of Sandy Bridge, it had an instruction decode rate of 4 instructions per clock per core.

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2

Now you're adding cores to the problem.  This was a single core question.  Maybe I should be more specific since you insist on being a turd about every little statement made.  IPC can not be greater than 1.0 per core per cycle per clock per transistor.

[Godlygamer23]

20 minutes ago, JefferyD90 said:

Now you're adding cores to the problem. 

Hm?

Spoiler

 

[Rythmium]

Wow, guys! This is really impressive! I expected to maybe have a couple responses by tomorrow, but in just a couple hours it looks like you guys have started up a great conversation! This is all really fascinating.

 

It appears the consensus is that the Skylake i5/i7 will have a single-thread advantage despite the G3258's huge clock speed thanks to higher generational IPC. That's good to know; at least going with 4 cores won't kill performance per thread.

1 hour ago, porina said:

If you really care about a single thread, then the i5 could have an advantage over the i7 when HT is on. You could end up with less important stuff going on the same physical core as what you want to run, and it will slow that down. On the i5, it has to go to another core. Now, on the i7 you can turn off HT to stop this form happening, and still benefit from the bigger L3 cache size. For tasks I do where I know HT provides no benefit, turning it off does speed it up for multiple reasons. 

This is actually a great idea. I agree that HT could potentially slow down each thread of an i7. Sounds like I should just go with the i7, and turn HT off until the day when I can actually benefit from more threads. It'll just be like an i5 with a higher clock speed and more cache.

 

 

Some people have asked what this computer will be used for. For now, besides being an everyday computer, it'll be handling some mathematical software that is only written to take advantage of 1 core at a time. Eventually it's possible (and likely) that I'll move over to video editing on this machine, which is why I need more than 2 cores for future use. It doesn't need to be a video editing beast, though. Just acceptable.

 

 

Thanks so much for all the responses! If you guys want to keep the conversation going, please do so—it's so interesting! And if anyone feels that they have more input, feel free! I'd love to hear new perspectives on this issue.