Skylake i3 vs i5: how does hyperthreading compare?

[Mr Jack]

The i3 runs four threads on two cores whilst the i5 runs four threads on four cores. There's also the larger L3 cache to consider (6Mb vs 4Mb) but this seems to be the big difference between the two chips and there's a sizeable gulf in the price to go along with it.

Does anyone know how the real world specs compare for multi-threaded work? Does hyperthreading stand up to the performance of having four genuine cores? Given that you can get higher clock speed at a lower price, does it make more sense to prefer a 3.9Ghz i3 over a lower speed i5.

Any thoughts?

 

[wazza300]

true cores are always slightly better that two/two virtual on an i3

 

[DragonQ]

It totally depends on the application. For applications that fully utilise 4 cores, HyperThreading tends to add +50% performance on average (extremely rough number) - the best number I've seen is 70-75% (for video encoding I think). Two extra real cores obviously adds +100% performance.

 

[IT Troll]

It totally depends on the application. For applications that fully utilise 4 cores, HyperThreading tends to add +50% performance on average (extremely rough number) - the best number I've seen is 70-75% (for video encoding I think). Two extra real cores obviously adds +100% performance.

Those figures are way too high for hyperthreading. I've never seen a real life benchmark which shows more than a 15% increase. More often it is down to single figures.

Hyperthreading doesn't magically give you more processing power. It just optimises what you have already got so it can be used more efficiently.

 

[joeyjojo]

Those figures are way too high for hyperthreading. I've never seen a real life benchmark which shows more than a 15% increase. More often it is down to single figures.

Hyperthreading doesn't magically give you more processing power. It just optimises what you have already got so it can be used more efficiently.

This, Mr Jack.

Compare the i3 4330 (3.5 GHz, 2C/4T) to the i5 4670K (3.5 GHz, 4C/4T): http://www.anandtech.com/bench/product/837?vs=1192

In single-threaded (ST) Cinebench the i5 is 8% faster than the i3 due to the increased cache you mentioned.

In multi-threaded (MT) Cinebench the i3 score increases by 2.35 times, the extra 35% thanks to the Hyperthreading. (This corroborates with i5 vs i7 numbers - an i7 4770K scores 37% higher than an i5 4670K, 36% when controlling for cache.) As IT Troll hints this is very unusually high, Cinebench must have a lot of blocking behaviour that favours thread switching.

Also in MT Cinebench the i5 is 67% faster than the i3. Relative to ST numbers we'd expect it to be 4 x faster from the core count, +8% cache effect, compared to the i3's 2x from core count +35% from hyperthreading, i.e. 4.08/2.35=74% faster, so 67% seems reasonable.*

* The underperformance may be inherent in parallelising Cinebench - from ST to MT the i5 only increases by 365%, so we'd expect the i5 to beat the i3 by 59%. 67% is half way between 59% and 74%

All this is very Cinebench specific. Compare the i5 and i7 I mentioned earlier and you see 10-20% gains are more common. And for others reading this - games benefit very little from hyperthreading at realistic resolutions.

 

[Mr Jack]

Cool, thanks for the information.

That leaves the second part of my question: given the price differential, is it better to go for a low end i5 or the top of the pile i3 if I can't stretch to a higher i5. For example, a 3.9GHz i3 costs £20 less than a 2.7GHz i5.

It looks like the 3.9Ghz is going to play games better due to significantly higher single thread performance but the i5 will score better on all round benchmarks due to the big jump from 4 virtual cores to 4 real cores. Does that sound about right?

 

[CAT-THE-FIFTH]

This, Mr Jack.

Compare the i3 4330 (3.5 GHz, 2C/4T) to the i5 4670K (3.5 GHz, 4C/4T): http://www.anandtech.com/bench/product/837?vs=1192

In single-threaded (ST) Cinebench the i5 is 8% faster than the i3 due to the increased cache you mentioned.

In multi-threaded (MT) Cinebench the i3 score increases by 2.35 times, the extra 35% thanks to the Hyperthreading. (This corroborates with i5 vs i7 numbers - an i7 4770K scores 37% higher than an i5 4670K, 36% when controlling for cache.) As IT Troll hints this is very unusually high, Cinebench must have a lot of blocking behaviour that favours thread switching.

Also in MT Cinebench the i5 is 67% faster than the i3. Relative to ST numbers we'd expect it to be 4 x faster from the core count, +8% cache effect, compared to the i3's 2x from core count +35% from hyperthreading, i.e. 4.08/2.35=74% faster, so 67% seems reasonable.*

* The underperformance may be inherent in parallelising Cinebench - from ST to MT the i5 only increases by 365%, so we'd expect the i5 to beat the i3 by 59%. 67% is half way between 59% and 74%

All this is very Cinebench specific. Compare the i5 and i7 I mentioned earlier and you see 10-20% gains are more common. And for others reading this - games benefit very little from hyperthreading at realistic resolutions.

The Core i5 4670K has Turbo upto 3.8GHZ which would explain the Cinebench difference.

Plus compared to a dual core without HT,you can see differences in frametimes too,which means the Pentiums can have much less smoother gameplay and this is going to get worse as time progresses.

 

[humbug]

This, Mr Jack.

Compare the i3 4330 (3.5 GHz, 2C/4T) to the i5 4670K (3.5 GHz, 4C/4T): http://www.anandtech.com/bench/product/837?vs=1192

In single-threaded (ST) Cinebench the i5 is 8% faster than the i3 due to the increased cache you mentioned.

In multi-threaded (MT) Cinebench the i3 score increases by 2.35 times, the extra 35% thanks to the Hyperthreading. (This corroborates with i5 vs i7 numbers - an i7 4770K scores 37% higher than an i5 4670K, 36% when controlling for cache.) As IT Troll hints this is very unusually high, Cinebench must have a lot of blocking behaviour that favours thread switching.

Also in MT Cinebench the i5 is 67% faster than the i3. Relative to ST numbers we'd expect it to be 4 x faster from the core count, +8% cache effect, compared to the i3's 2x from core count +35% from hyperthreading, i.e. 4.08/2.35=74% faster, so 67% seems reasonable.*

* The underperformance may be inherent in parallelising Cinebench - from ST to MT the i5 only increases by 365%, so we'd expect the i5 to beat the i3 by 59%. 67% is half way between 59% and 74%

All this is very Cinebench specific. Compare the i5 and i7 I mentioned earlier and you see 10-20% gains are more common. And for others reading this - games benefit very little from hyperthreading at realistic resolutions.

Multicores do not scale 100%

The best way to see what you get for HT is to use the same number of cores at the same clock rate with and without HT.

http://www.anandtech.com/bench/product/837?vs=1198

Haswell i5 @ 3.5Ghz: 4c - 4t = 6.21
Haswell i7 @ 3.5Ghz: 4c - 8t = 6.69 (+8%)

Its nowhere near 30 odd % let alone 50%

 

[Mr Jack]

http://www.anandtech.com/bench/product/837?vs=1198

Haswell i5 @ 3.5Ghz: 4c - 4t = 6.21
Haswell i7 @ 3.5Ghz: 4c - 8t = 6.69 (+8%)

I think you have made a mistake: the link - which seems to be where the numbers you give comes from - compares two i5 processors not an i5 and an i7.

 

[humbug]

I think you have made a mistake: the link - which seems to be where the numbers you give comes from - compares two i5 processors not an i5 and an i7.

Yeah, never-mind, the numbering system Intel use threw me.

 

[JasonM]

Hyper-threading add's performance anything from about -3% (yes minus) to +100% per core. Typically however it's about +25%. Much depends if the software compiler was hyper threaded aware, it's only old software that had issues with hyper-threading.

I have some multi-threaded Neural Network software I've written myself in Visual Studio 2013. I don't see any slow down using either 4 or 8 threads on the i7 4770's, to me it just feels like 8 cores, but Visual Studio 2013 compiler is aware of the features inside a 4th generation i7. If I was using say Visual Studio 2005 performance could be less.

 

[Rroff]

TBH the main advantages of hyper-threading (when its working properly) is that it keeps things smooth under heavy multiple application/threaded loads - you might even see a small performance decrease but "usually" things will be more responsive than without it.

 

[IT Troll]

Hyper-threading add's performance anything from about -3% (yes minus) to +100% per core. Typically however it's about +25%.

What in reality shows a 100% performance increase per core with hyperthreading? Whilst it might be possible to code something which artificially favours hyperthreading this bares no resemblance to real world use.

 

[JasonM]

What in reality shows a 100% performance increase per core with hyperthreading? Whilst it might be possible to code something which artificially favours hyperthreading this bares no resemblance to real world use.

A few years ago a review site did an extensive benchmark test on HT. Out of about 30 tests there was one piece of software where the performance was doubled, I recall the software being quite old. The average increase was about 25% however.

There is another review where someone did a comprehensive test of HT on SQL Server. There were some conditions where performance was x4, but this was probably due HT reducing SQL query wait times.