Cores: Physical V. Logical

>>61282591
Depends on if it's Intel or AMD, because AMD's SMT is 25-30% more efficient.

>>61282591
Threads just organise the data being fed into the cores to maximise efficiency,so it really depends on the workload.

If you want an idea of if you would benefit from more cores or more threads you need to analysis every piece of software you plan to run on a case-by-case basis.

>>61282591
Virtual cores have about 25% the performance of a physical core.

>>61282591

it depends strongly on the design of the full core. If you are talking about a fat, wide, branch predicting, out of order core you probably want the extra full core.

If you have something like an early atom that was an in-order design that required two instruction streams to fetch from in order to keep the inter-instruction dependencies low, well you probably want the HT/SMT.

Note that as you specified the IPC would be the same heavy -> heavy with smt and lite -> lite with smt.

But because you are asking probably wrt to fat intel and amd desktop products, I'd probably go with the full core. Amd smt gets something like a 10% boost, and intel gets something like a 20% boost, and these are kind of best case. Giving each instruction stream its own cache is always good.

>>61282591
5 years ago I would have told you to just get the 4-core since nothing really used those extra threads except for a few snowflake programs. Today is it somewhat different going from a Pentium to an i3 now can yield much better performance even in some general stuff, even games which never gave a shit about extra threads a few years ago now want those extra threads.
I would still say go for the 4-core for a general use system because the vast majority of stuff will not benefit from extra threads

>>61283078
AMD's SMT is slightly better than Intel's.

>>61283122

I think you are looking at the benchmarks that have smt-on vs smt-off, forgetting that AMD has 8 extra smt threads to 4 extra smt threads for Intel per cpu. On a per thread basis, am pretty sure intel wins, and that is what we were talking about.

Yes, yes, yes, Ryzen is a fantastic cpu.

>>61282709
That literally just means that Zen's branch prediction, caches or prefetch are garbage compared to Intel.
Which actually makes me very excited for Zen+, as even with these obvious flaws their IPC rivals Intel's. They're hilariously fucked if AMD manages to iron these out in Zen+

Quad-core without HT/SMT.
Executing thread twice as often but only on half core (because there's another thread running at the same time on the same core) is not "nice" to the thread if you think about it.

speaking of which, kinda related:
https://www.semiwiki.com/forum/content/6879-exclusive-globalfoundries-discloses-7nm-process-detail.html

>>61283443
Just amazing, from 10000nm to 7nm in 46 years. 2300 Transistors to 10 billion.

>>61283303
depending on who you listen to (e.g., Agner), it could potentially be a slightly bottlenecked instruction fetch etc. early in the frontend.

my personal feeling is that Intel's much faster datapaths (64B up to Frylake-E, 128B thereafter) mandated by wider AVX can prevent a lot of potential integer pipe stalls too, at the cost of sucking down quite a bit of power.

we won't see till HotChips at the earliest how IF really handles coherency snoop forwarding and filtering, but there's also a good chance that AMD made choices favoring power consumption over bandwidth and latency here as well. I was rather shocked at first that Epyc's GMI links are relatively slow and that the IF crossbars are still clocked at ~1.2 GHz, but it does show where AMD is trying to go with their first gen Zen products.

EDIT: I am not in the market for a new CPU. My 6600K is working wonders.

Anyways, I think the threads v clocks war going to be an incredibly interesting development over the next few years. I can see a future where AMD's SMT and thread count, even with their lower IPC might prove advantageous as applications become more and more thread based.

Are we heading to a future where threads > clocks? If so, Intel needs to buckle up. Their current i9 lineup is a joke.

IMO, if Intel wanted to really smash Ryzen the smart move would have been to increase core/thread count on their existing i3/5/7 models, perhaps the i5 becomes the i3 (or a 3C/6T) while the i5 becomes a 6C/6T and the i7 remains an i5 with HP and better binning.

The thread wars are here to stay, and I think the new X299 was a huge misstep for the company.

>>61282873
Pretty much this. HT/SMT can even have adversal effect because of cache sharing for certain tasks. Really depends on the software.