Is there ever a case where a "hyperthreading core" is better than a "true core"?

[nerd866]

This is where I bump into the limits of my understanding of hyperthreading.

 

A classic analogy is that adding more cores is like adding more mouths for eating food. Adding hyperthreading is like adding more hands to help you get the food to the mouths.

 

Now, let's look at a 4-core machine in two scenarios:

* 4 true cores with no hyperthreading

* 3 true cores with 1 hyperthreaded core.

 

I know this doesn't exist on real CPUs but this is possible to create by setting up processor affinity, which is where the real-world application of this question comes in:

Is there ever a case where one could get better performance from an application by assigning it 3 true cores and one hyperthreaded "core" rather than assigning it 4 "true cores?"

 

If we go back to the food analogy, is there ever a case where it's better to have 3 mouths and 2 hands, rather than 4 mouths and one hand? When described this way, I can almost certainly say "yes" but I have a feeling this analogy breaks down somewhere, which is where the limits of my understanding come in.

 

is there ever a case where a hyperthreaded "core" is preferrable to assigning another true core to an operation?

[vorticalbox]

I am going with no. hyperthreading is not the same as a core. 2 cores + HT is not the same as 4 cores.

[Sakkura]

No. Because there's no such thing as a hyperthreaded core. There's just one physical core, which can address two threads. Neither of those threads is any more "real" than the other one.

[nerd866]

1 minute ago, vorticalbox said:

I am going with no. hyperthreading is not the same as a core. 2 cores + HT is not the same as 4 cores.

I agree that they're not the same thing. I feel like that's where the question gets tricky: Because they're not the same thing, that means they do different things.

 

Since they do different things, does that mean there could be a case where hyperthreading does a job better than a core?

[Floris_Heinen]

I assume your "3 true core + 1 HT core" CPU will effectively be a 3.5 core CPU.

 

I think the only thing that HT does is to allow 1 physical core to do 2 tasks simultaneously

[SLAYR]

1 minute ago, nerd866 said:

I agree that they're not the same thing. I feel like that's where the question gets tricky: Because they're not the same thing, that means they do different things.

 

Since they do different things, does that mean there could be a case where hyperthreading does a job better than a core?

A hyperthread performs only about 50-75% as good as an actual physical core. 

 

So no.

[nerd866]

2 minutes ago, Sakkura said:

No. Because there's no such thing as a hyperthreaded core. There's just one physical core, which can address two threads. Neither of those threads is any more "real" than the other one.

That's a good way to describe it. 

 

If this is the case, that one CPU can address two threads, does that mean it doesn't matter in terms of CPU affinity whether I assign a process to CPU0 or CPU1 if they're both referring to the same physical core?

[Djole123]

Of course there is a case where a CPU without hyperthreading can destroy a CPU that has hyperthreading, even if it's within the same brand.

 

Spoiler

My Atom N2600 gets destroyed by the Pentium G3258. The Atom has 2C/4T and the Pentium has 2C/2T

 

[Thony]

As mentioned, Physical cores can be split in two and then HT happens. 

 

So with 3+1 just means that its 3 physical cores and one of those cores is behaving like two cores.

[ThomasD]

3 minutes ago, Sakkura said:

No. Because there's no such thing as a hyperthreaded core. There's just one physical core, which can address two threads. Neither of those threads is any more "real" than the other one.

If efficient use of available computational power is the goal the this is the place where HT can be "better."

 

But, for most users, that is a very unimportant metric.

[vorticalbox]

3 minutes ago, Djole123 said:

Of course there is a case where a CPU without hyperthreading can destroy a CPU that has hyperthreading, even if it's within the same brand.

 

  Reveal hidden contents

My Atom N2600 gets destroyed by the Pentium G3258. The Atom has 2C/4T and the Pentium has 2C/2T

 

 

 

 

That is mostly due to differences  in architecture.   atom is 1.6 GHz vs 3.2 on the pent. The atom is 3.5 watts and the pent is 53 watts.

 

The Pentium is a much higher clock speed and much more power to work with.

[SLAYR]

6 minutes ago, Djole123 said:

Of course there is a case where a CPU without hyperthreading can destroy a CPU that has hyperthreading, even if it's within the same brand.

 

  Reveal hidden contents

My Atom N2600 gets destroyed by the Pentium G3258. The Atom has 2C/4T and the Pentium has 2C/2T

 

This is completely due to the atom being based on an old shit architecture, and being 1/2 of the clockspeed of the pentium. Saying hyperthreading doesn't help from that comparison is just completely biased.

 

http://cpu.userbenchmark.com/Compare/Intel-Atom-N2600-vs-Intel-Pentium-G3258/m704vs2434

[Simon771]

I know how HT works and I can't even write a good argument lol

 

HT just allows 1 core to do 2 things at the same time.

[porina]

A good case for study might be in distributed compute applications, where you don't necessarily care about how fast a single task is run, but you want as many total tasks run in a given time. Take an i7 or i3, and you can run them with or without HT. In my testing, the bonus throughput from HT varies from about 0% to 50%. That is, you could say at best a dual core without HT is roughly equivalent to 3 cores when HT is enabled. Worst case, there is no benefit at all or possibly some negative performance from extra overhead.

 

If anyone knows of any case where running HT will give more than 50% throughput compared to without HT, I'd be interested to hear about it.

[Taja]

A core with HT does not "transform" into 2 cores. People think a core with HT can do 2 things at the same time, but it is not really like that, the instructions are processed with less delay, thus making the core faster. But to process 2 instructions literaly at the same time, you need two cores.

 

Ps: im no expert, if I explained poorly, someone correct me please.

[rockon5622]

2 hours ago, Taja said:

A core with HT does not "transform" into 2 cores. People think a core with HT can do 2 things at the same time, but it is not really like that, the instructions are processed with less delay, thus making the core faster. But to process 2 instructions literaly at the same time, you need two cores.

 

Ps: im no expert, if I explained poorly, someone correct me please.

It can do two things at once. But not always, it can do two instructions at same time. As long as the instructions do not require the same modules. For example, if I do a multiply and a mem read with diff registers they should occur at same time. 

 

So,  yes 4 physical cores are better than 3. But the 3 core with one core with hyperthreading is more power friendly.

 

[Zodiark1593]

The benefits that can be gleaned from Hyperthreading are entirely dependent on the task at hand, and how well the task(s) is/are coded. By it's nature, Hyperthreading allows for utilization of resources that would otherwise go unused by duplicating select parts. Hyperthreading does not add raw compute as a result, but rather, allows the core to better utilize what it has. In addition, due to the components that are duplicated, latency in switching tasks is very minimal.

 

A 3 core Hyperthreaded processor has 3/4 the raw compute of an equally clocked, non hyperthreaded quad core. The former can address 6 tasks before it starts running into context switching penalties. The latter is more powerful, but isn't as likely to be fully utilized, and suffers from context switching penalties after just 4 tasks.

 

For all intents and purposes, the i5 is equally as powerful as the i7 in raw compute terms, though the lack of Hyperthreading makes utilization more difficult in the i5, thus lesser real-world performance is observed.

[Juggern0ttt]

2 physical cores, assuming they are of the same clock speed and architecture, will always perform better than a single hyper threaded core . 

 

A core is essentialy an assembly line for information. Try to visualize it as  a feeder dumps shit on a line, the line assembles it into useful parts, then a compiler puts out in order to send out. That's a physical core. 

 

A hyperthreaded core is the same but with two lines to organize the info, but the feeder can still only feed so much , and the compiler can only compile so much. Adding the extra line better utilizes and cuts back on down time in between organizing and compiling, but having an entire extra line and compiler can still do more work . 

 

Excuse my clunky visual aid. But there is only one feeder and one compiler per core whether HT or not.  they can only work so fast, essentially becoming a bottleneck in a HT core

Edited November 9, 2016 by Juggern0ttt
Added visual aid