AMD Ryzen Threadripper 1950X and 1920X

[HEXUS]

The mega-tasking era starts here.

Read more.

[lumireleon]

Has intel given us evidence that infinity fabric inside these chips has latency issues?

[LSG501]

Has intel given us evidence that infinity fabric inside these chips has latency issues?

Pretty sure AMD themselves have said there is a small latency increase when using infinity fabric, nothing to worry over with multithreaded software like cinebench etc but in gaming it can have a small impact (as shown in the game mode tables).

Having said that, it might be easy enough to overcome with some add in code by AMD and the game devs so we don't need to keep switching mode via restart. At the same time the majority of people buying threadripper will likely be looking at multithreaded performance over gaming I would guess but it's not like it's a slouch in gaming even when in 'creation' mode imo either.

[scaryjim]

Has intel given us evidence that infinity fabric inside these chips has latency issues?

What's that got to do with Intel? Why would Intel have evidence of what causes latency in a competitor's processor?

As it is, Hexus have provided evidence, if you'd bothered to read the review properly. Switching to Game Mode - which forces local-only memory access - drops the memory latency from 90ns to 65ns. The only possible reason for that difference is that the infinity fabric between the two dies is causing increased latency.


Anyway, with that out of the way:

Speculation time In the Intel Core-X reveal thread I suggested that 16 core Threadripper would probably be faster than 16 core Core X due to its relatively high base clock speed. Now we've got some numbers, we can have a bit of a play. I'm going to assume that cinebench runs all the processors at their base clock speeds, and scales perfectly with both frequency and core count. I'm sure that's not entirely accurate, but it should give us a ball park. Let's see how the 7960X (16C) and 7980XE (18C) might score, extrapolated from the 7900X (using Core ratio * base clock ratio * 7900X score):

7960X - 1.6 * (2.8/33) * 2187 = 2969 (oops, just behind TR 1950X)
7980XE - 1.8 * (2.6/3.3) * 2187 = 3102 (yay, a whole 69 points ahead of 1950X)

So there you have it ladies and gentlemen - pay twice the amount for your processor, get 2% more cinebench performance....

[QuorTek]

Dunno singlethread is possibly on the way out anyway ad developers now has a new aim to go towards for better multi threaded performance overall.

[scaryjim]

Dunno singlethread is possibly on the way out anyway ad developers now has a new aim to go towards for better multi threaded performance overall.

That's not strictly true - there will always be algorithms that simply can't be parallelised, and sometimes over-agressive threading can make code run slower (something I've fallen foul of myself a few times ). But developers are getting better at working out what works in parallel, and there are a lot of optimisation tools now that will do some of the work for you. But it's worth remembering that, for certain tasks, individual thread performance will still be key.

[LSG501]

What's that got to do with Intel? Why would Intel have evidence of what causes latency in a competitor's processor?

As it is, Hexus have provided evidence, if you'd bothered to read the review properly. Switching to Game Mode - which forces local-only memory access - drops the memory latency from 90ns to 65ns. The only possible reason for that difference is that the infinity fabric between the two dies is causing increased latency.


Anyway, with that out of the way:

Speculation time In the Intel Core-X reveal thread I suggested that 16 core Threadripper would probably be faster than 16 core Core X due to its relatively high base clock speed. Now we've got some numbers, we can have a bit of a play. I'm going to assume that cinebench runs all the processors at their base clock speeds, and scales perfectly with both frequency and core count. I'm sure that's not entirely accurate, but it should give us a ball park. Let's see how the 7960X (16C) and 7980XE (18C) might score, extrapolated from the 7900X (using Core ratio * base clock ratio * 7900X score):

7960X - 1.6 * (2.8/33) * 2187 = 2969 (oops, just behind TR 1950X)
7980XE - 1.8 * (2.6/3.3) * 2187 = 3102 (yay, a whole 69 points ahead of 1950X)

So there you have it ladies and gentlemen - pay twice the amount for your processor, get 2% more cinebench performance....

Supposedly the 16 core intel whatever it's called can hit 3300 on cinebench (in articles, like on engadget, with no visual proof)... personally I'm not buying it until I've seen some actual real world benchmarks/testing. Would I pay nearly twice the money for that small a gain in a very specific benchmark which as we all know is a benchmark and isn't 100% accurate of real world results...nope.

Having said that while cinebench is useful I don't fully trust it's end results because on the gpu side I've had faster results from older and weaker hardware.....

[scaryjim]

Supposedly the 16 core intel whatever it's called can hit 3300 on cinebench ...

If it can hit/maintain an all core boost of 3.1GHz that's possible. < 10% ahead with < 10% clock speed deficit ... suggests somewhere between 15% and 20% IPC advantage for Skylake-X. That's still quite significant, particularly when you take Ryzen's lower peak clocks into account. Wonder if AMD will target higher real-world IPC or faster clocks...?

[Ozaron]

That's not strictly true - there will always be algorithms that simply can't be parallelised, and sometimes over-agressive threading can make code run slower (something I've fallen foul of myself a few times ). But developers are getting better at working out what works in parallel, and there are a lot of optimisation tools now that will do some of the work for you. But it's worth remembering that, for certain tasks, individual thread performance will still be key.

Yeah, but doesn't this mean we can run more single threaded operations simultaneously without tapping out a core? Distributing them across 32 threads sounds somewhat appealing, not that I know anything at all about CPU management efficiency...

[lumireleon]

so infinity fabric is a none issue? Intel's GLUED processors chronicles..... what about this idea that intel pointed out that when a program needs let say 17 threads the last NO.17 thread will have to jump to the other CCX through infinity fabric thereby hurting performance badly.....~10-20% BADLY

[CAT-THE-FIFTH]

so infinity fabric is a none issue? Intel's GLUED processors chronicles..... what about this idea that intel pointed out that when a program needs let say 17 threads the last NO.17 thread will have to jump to the other CCX through infinity fabric thereby hurting performance badly.....~10-20% BADLY

Intel is full of crap. This is an example of it:

https://mobile.twitter.com/IanCutres...69342001508353

Intel released a new optimization manual. To get the best local latency, enable on-die NUMA/clustering. 4 NUMA nodes per CPU. Chapter 8.

That's the same NUMA clustering as EPYC.

Oh the irony.

[Bagpuss]

£1 = $1.30

$999 & $799 = £999 & £799

..why?

Even with 20% VAT, there is no excuse for it.

[scaryjim]

... doesn't this mean we can run more single threaded operations simultaneously without tapping out a core? ...

Only to the extent that any one single-thread operation doesn't tap out a core. If you have a resource intensive algorithm that cannot be parallelised, then it's going to max out a core and you can't make it go any faster just because you've got more cores. You can do other things at the same time, sure, but if what you want is to do that one thing as fast as possible, then you need the fastest possible single-thread performance. If that algorithm - or part thereof - is a blocking point in your software, throwing more cores at it won't make your software faster.

Interestingly, the inverse isn't necessarily true. OS schedulers can happily run multiple program threads on individual cores by simply slicing the time each thread gets on the core. So if you have one extremely fast core, it can potentially handle multi-threaded tasks better than a larger number of slower cores. It's all horses for courses.

That's getting more technical and theoretical than I really wanted to, though. My point here was simply to point out that there are some things that simply cannot be sped up by throwing more cores at them, which means that single threaded performance will never be "on the way out". It may become less important, but there will always be tasks for which it is critical.


EDIT:

£1 = $1.30

That's a mid-market rate - as a consumer you can't actually buy $1.30 for £1 at the minute. Add in export costs/additional shipping/localised marketing/etc. ... it's pretty damn close.

When we were nearer £1 = $1.50 and we were still getting direct £:$ reprices in the UK - that there was no excuse for. What we're getting now is probably about right...

[CAT-THE-FIFTH]

People do realise since TR has far more cores than the Intel equivalent,even if single core scores are a bit lower,you can simply run more instances of the a single threaded program leading to a bigger overall increase in throughput.

[scaryjim]

... you can simply run more instances of the a single threaded program leading to a bigger overall increase in throughput.

That's fine if you have several different things for that single threaded program to work on. Depends on your workflow. If you only have one thing to do and that one thing is single threaded, the extra cores won't help you.

Of course, the chances are, if you're considering threadripper, your workflow involves a lot of heavily threaded tasks. If you've got the kind of workflow that heavily depends on single-threaded performance then you should be looking elsewhere. I just think it's impprtant that people remember that you can't solve everything with more threads. If you could we'd all be using 16+ core Atoms already!

[CAT-THE-FIFTH]

That's fine if you have several different things for that single threaded program to work on. Depends on your workflow. If you only have one thing to do and that one thing is single threaded, the extra cores won't help you.

Of course, the chances are, if you're considering threadripper, your workflow involves a lot of heavily threaded tasks. If you've got the kind of workflow that heavily depends on single-threaded performance then you should be looking elsewhere. I just think it's impprtant that people remember that you can't solve everything with more threads. If you could we'd all be using 16+ core Atoms already!

For my mates bioinformatics stuff(he was running as a side project),having more cores on the Phenom II X6 made more sense than getting a Core i5 for example.

Some of it was multi-threaded,others were only single threaded programmes legacy programmes. After looking at some benchmarks for similar software,it made massive sense to get that Phenom II X6 as it was also cheaper.

The Core i5 was quicker per thread,but not quick enough to negate the advantage of 50% more cores,and having that extra throughput was very useful as the software got run for a few days(sometimes much longer).

Moreover,it was also useful if he wanted to actually use the computer for something else - even leaving one core not allocated on the Core i5 would have decreased the effective thread count by 25% instead of the 17% of the Phenom II X6.

Plus AMD having a lot of SATA3 ports also was useful too!

All I got to say that CPU has lasted a longtime and has been utterly hammered.

So he has decided his PC will be updated to a Ryzen 7 1700,and Intel won't be getting a look in TBH.

Edit!!

Also people repeating the Intel glued together rubbish should look at my other post.

Even their own optimisation guide tells people to treat their X299 CPUs as "4 NUMA nodes per CPU".