AMD Hyperthreading?

[lx_srs]

You are both wrong. There are two types of logic unit per core (both for Intel and AMD and whomever else makes x86 processors) integer and floating point units. These work on different kinds of math.

 

1 Intel 'core' = 1 integer and 1 floating point unit

1 AMD 'module' = 1 integer and 2 floating point units

 

This means that for integer math, an FX-8350 is a quad core. For floating point math, it's an 8-core.

Intel is always symmetrical and WYSIWYG.

Thx m8.

[Sauron]

No, and why do you care? having more physical cores is better than hyperthreading anyways.

 

depends on how fast those cores are...

[boboman342]

The FX CPUs are somewhat similar, they are actually modules with two FPUs each. For example the 8 core FX is actually 4 compute modules, with 8 FPUs that are considered "cores"

Isn't it the reverse? Don't 2 cores share 1 FPU?

[harrynowl]

Isn't it the reverse? Don't 2 cores share 1 FPU?

Yes. See my post above for more info.

[Faa]

 

 

 

 

You are both wrong. There are two types of logic unit per core (both for Intel and AMD and whomever else makes x86 processors) integer and floating point units. These work on different kinds of math.

 

1 Intel 'core' = 1 integer and 1 floating point unit

1 AMD 'module' = 1 integer and 2 floating point units

 

This means that for integer math, an FX-8350 is a quad core. For floating point math, it's an 8-core.

Intel is always symmetrical and WYSIWYG.

 

SandyBridge has 3 ALU's & 6x FPU's with 3 of them at 256bit.

Haswell has 4 ALU's & 6x FPU's with all of them at 256 bit for AVX2

8350 has 4 ALU's and 4x 128bit FP's.

Everything that was marked with 256bit are the FP's, 4 remaining ones are ALU's.

[boboman342]

Yes. See my post above for more info.

Sorry, I am not so technically inclined as to the actual "physics" behind how the tech works.

[GangstaRas]

amd doesnt do hyperthreading but the concept of it is similar when they did the fx4000, 6000 and 8000 series. the modules were basically the usual complex cores that make other cpu cores of this world. but what amd did was made split these into simpler cores so it is 4, 6 and 8 physical cores respectively, but they are simpler cores that can only do less specified things and are tailored so that they do the most common process of general cpu computing which is integer instruction work which makes these integer cores. so when a process that came around that didnt involve much integer processing, you would see the phenoms getting better scores than these. intel cores are complex cores but hyperthreading is an efficiency type of approach of the physical cores available. so for every wasted gigahertz cycle (id assume im not sure) the cpu is idling with wasted resources so you do clever scheduling of processes to effectively keep the core working all the time more effectively, logical cores were born.

 

only relation they have, some resources  that are suppose to be solely per cpu are shared amongst the cores whether physical (amd) or logical (intel) and that's where it ends

[Razor512]

The AMD core module can be considered as a high performing style of hyperthreading. windows now has proper support for them, and will in most cases use 1 thread per module before doubling up on the modules.

 

A number of traditional CPU core components are shared and thus if 1 threads attempt to use the same core module for the same or similar task , then both cores will struggle for resources.

 

It is more than just cache being shared http://www.extremetech.com/computing/138394-amds-fx-8350-analyzed-does-piledriver-deliver-where-bulldozer-fell-short/2

 

 

 

 

Overall, the reason why I I feel it is a high performance version of hyperthreading, is because it shares the same flaws (but to a lesser extent) as hyperthreading.

 

With current intel CPU's, e.g., with a core i5, a compute intensive load on 1 core will not impact the performance of the core next to it as they are not sharing components core components needed to process the data.

 

when you use hyperthreading and a 2 compute threads fall onto the same core, then they fight over resources.

 

hyperthreading can give up to around 30% additional performance, as some components are added to support multiple threads which also minimizes downtime of the core when waiting for the scheduler in the OS to issue tasks for the CPU to do. The core module will add around 70-80% additional performance depending on the task.

 

 

Furthermore, even with updates to windows to improve support, for applications that cannot use 8 threads, you get better overall performance by turning off 1 core in each module rather than leaving them all on, as it prevents windows from accidentally putting more than 1 thread on a single core.

[computergeek227]

No hyperthreading is available in AMD processors but a physical core is better than a logical 'core'

[79wjd]

No hyperthreading is available in AMD processors but a physical core is better than a logical 'core'

Not when their 'physical' cores are basically logical cores.

[harrynowl]

Not when their 'physical' cores are basically logical cores.

Well a physical core from AMD is faster than a single hyperthread, it's just a quad core intel performs about the same as an 8 core AMD at the same clock. So adding 30% more performance with an extra 4 threads obviously leaves AMD behind.

[Faa]

The AMD core module can be considered as a high performing style of hyperthreading. windows now has proper support for them, and will in most cases use 1 thread per module before doubling up on the modules.

Hyperthreading is about re-using the same execution resources where as AMD is just duplicating their resources so the 2nd thread can run on the duplicated resources. What AMD is doing can be achieved by just making a 2nd core (no the module itself is the core)

Haswell has 4 ALU's per core, Bulldozer 4 ALU's. If you disable Hyperthreading on Haswell you still have 4 ALU's you can use, if you disable "Hyperthreading" (proper word is CMT) on AMD you have 2 ALU's that can be used.

 

 

No hyperthreading is available in AMD processors but a physical core is better than a logical 'core'

There's only one core with Hyperthreading enabled. Lets quit that virtual/logical core bs. All hyperthreading does it to allow to execute a 2nd thread at the same time which you would have needed a 2nd core if you don't have HT. There's where people pulled that virtual/logical core from.

Yeah a single core is capable of processing multiple threads, just had to keep it simple.

 

 

Well a physical core from AMD is faster than a single hyperthread, it's just a quad core intel performs about the same as an 8 core AMD at the same clock. So adding 30% more performance with an extra 4 threads obviously leaves AMD behind.

Uh? You just got two doors to access a core. If you run singlethreaded mode cinebench on core#0 only it will perform the same as with core#1. You'll get the 30% HT boost if you use core#0 & core#1 when you take the multithreaded mode.

A single module is faster than a single Intel core but slower than a single intel core with Hyperthreading enabled.

[harrynowl]

Uh? You just got two doors to access a core. If you run singlethreaded mode cinebench on core#0 only it will perform the same as with core#1. You'll get the 30% HT boost if you use core#0 & core#1 when you take the multithreaded mode.

A single module is faster than a single Intel core but slower than a single intel core with Hyperthreading enabled.

I know that it isn't really measurable in terms of single thread, but I was kind of saying in a peak performance mode a single "hyperthread" (or hyperthreaded cores) would have a lower performance than a real core(s), but as intels single threaded advantage is so good an intel 8 thread is much faster than an amd 8 thread despite having half the physical cores (or execution units, whatever people want to call them).

 

As I didn't really understand the logic behind this I tried to make that comparison:

 

Not when their 'physical' cores are basically logical cores.

[Faa]

I know that it isn't really measurable in terms of single thread, but I was kind of saying in a peak performance mode a single "hyperthread" (or hyperthreaded cores) would have a lower performance than a real core(s), but as intels single threaded advantage is so good an intel 8 thread is much faster than an amd 8 thread despite having half the physical cores (or execution units, whatever people want to call them).

 

As I didn't really understand the logic behind this I tried to make that comparison:

Well lets say core#0 is a physical core and #core1 is a fake/ht core which is 30% as fast as the physical core. The AMD will be faster than the fake core, so I get what you meant.

The computer guy said a physical core is better than a logical core, the djdwos guy says not when their physical cores (he means the ALU clusters) are logical cores. So I assume he meant the two logical cores are better than the physical core (which Intel apparently only is).

[harrynowl]

Well lets say core#0 is a physical core and #core1 is a fake/ht core which is 30% as fast as the physical core. The AMD will be faster than the fake core, so I get what you meant.

The computer guy said a physical core is better than a logical core, the djdwos guy says not when their physical cores (he means the ALU clusters) are logical cores. So I assume he meant the two logical cores are better than the physical core (which Intel apparently only is).

I'm not entirely sure, maybe I interpreted something wrong somewhere

[RubyRoks]

Do amd cpus have hyper threading? If so, which ones?

Technically yes. The octalcore cpus are quad cores with logical processors for each core. i believe it's a similar story for the Hex core chips as well

[Razor512]

While the design is clearly different, from an end result standpoint, the AMD core module is closer to hyper threading than it is to true multiple cores, as it shares the same flaws.

 

With a CPU like the core i5, you pretty much get perfect scaling when going from 1-4 threads.

 

with the AMD core module, if you only use 1 core per module, then you also get pretty much perfect scaling, but when you start using the second core on each module, then you end up with both threads on the module running slower (higher overall performance but more like 70% scaling instead of 100% scaling).

 

It is clearly better than actual hyper threading as it scales better, but the increased level of resource sharing. causes poor scaling. If AMD would quit it with the core module and give each core their own resources like with the phenom II series, then they will likely get an extra 25-30% performance boost without any additional optimization.

 

Technically is not hyper threading, but it has the same flaws as hyper threading. This means that you are more likely to get performance drops due to workloads on adjacent threads.

 

Keep in mind that most PC games are not fully multithreaded. Instead they assign different tasks to different cores, e.g., 1 core may handle AI and another core may prepare frames for the GPU to render, and another may handle physics, and so on

The problem with this multithreading, is that if one of the threads saturate, e.g., the core handling the AI is simply not fast enough to handle the work load, then you end up with all of the other threads waiting on that single AI thread.

 

The issue with the core module is, suppose that core 1 is just able to handle all of the AI with 100% no additional resources available to handle any more work, and then the game makes core 2 handle the physics, then what will happen is when ever that physics thread needs to do a lot of work, then it will eat into the AI performance because core 1 may lose up to 30% of its performance.

 

hyper threading has this same issue, but it is not as underhanded as intel never claims to have more cores. the core i7 4770k is still marketed as a quad core CPU.

 

There is also a lot more OS level optimization to deal with hyper threading in terms of not using the logical unless it absolutely needs to.

 

With the AMD core modules, it is harder to avoid those modules, and even with an up to date version of windows, a 4 module 4 core config (1 core per module disabled), consistently benchmarks higher than a 4 module 8 core config when only 2-4 threads are being used. You still regularly end up with 2 threads on a single module when there is an additional core module available.

 

The CPU is still fast, but its design holds it back from achieving its true performance.

 

Furthermore, in moving to the core module, they actually lowered IPC, AMD FX needs a 400-500MHz clock speed increase per core over the older Phemon II cores in order to match the single threaded performance.