Mini-news: AMD Zen 2 architecture said to have 16% higher IPC than the original 1st-gen Zen CPUs

[ravenshrike]

45 minutes ago, Okjoek said:

What? I don't understand the question.

 

RTX joke regarding battlefield 5 and ray tracing.

[Stefan Payne]

6 hours ago, pas008 said:

this is where amd can hold mainstream for awhile

 

and on intel surviving dont forget intel can do mesh emib(might not be able to get same power consumption though lol) plus they have many other areas besides cpus they have been dabbling everywhere

think it would take along time for them to die

For that you need to have a design first.

They don't

So its only the usual "Intel will be back!!!111" claim without proof.

 

So why do you do that, when there is nothing to indicate your claim??

Even if they do that, that won't cange the design of the CPU, the inefficiency, the high power consumption. 

 

And the advantage AMD has with 7nm manufacturing over Intel with 14nm...

 

But we might see that next week, if the guys having 9900K are right...

[Stefan Payne]

43 minutes ago, Taf the Ghost said:

Cannonlake, the doomed 10nm Skylake shrink, does have a single AVX512 unit per core. So it was supposed to come to desktop. We'll see it with Icelake in 2020, but the AVX512 unit actually isn't the full instruction set just yet. That appears to be either Tigerlake or the one after that. Another reason AMD isn't also going to be adding it too soon.

Also, as you can see with Skylake-X, it increases the Power Consumption dramatically...

Since we are already power limited, its not really a good idea right now.

 

We need other solutions for our problems, not AVX512...

[Okjoek]

43 minutes ago, ravenshrike said:

RTX joke regarding battlefield 5 and ray tracing.

Oh that one went over my head, sorry. I only know the RTX on/off meme.

[leadeater]

9 minutes ago, Stefan Payne said:

Also, as you can see with Skylake-X, it increases the Power Consumption dramatically...

Since we are already power limited, its not really a good idea right now.

 

We need other solutions for our problems, not AVX512...

For a very large performance boost, AVX and all their iterations are more power efficient (perf/W) than the last and also than not using it. Intel is already putting power controls in place, AVX offset to lower the power usage when using AVX.

 

AVX is the solution to the problem, the problem that existed and was created to solve. Any ASIC like instruction set is far superior than a non accelerated path, there isn't an alternative because that would involve finding a completely new mathematical law to base it off which would also have to be better than what we have now.

[Stefan Payne]

Never said anything against AVX.

AVX is not AVX512.

[leadeater]

6 minutes ago, Stefan Payne said:

Never said anything against AVX.

AVX is not AVX512.

They are all the same thing, AVX-512 might have some new instruction sets as part of it but the core of it is the same. AVX-512 is not different to AVX1 or AVX2, not computationally. The difference is the register size, how many floating point numbers fit in to the register. The more that fit in the register the more operations you can do per clock and that is where the performance increase comes from.

 

Quote

Intel AVX-512 enables twice the number of floating point operations per second (FLOPS) per clock cycle compared to its predecessor, Intel AVX2. A single register under Intel AVX-512 can hold up to eight double-precision or 16 single-precision floating-point numbers. Or in other words, Intel AVX-512 enables processing of twice the number of data elements that Intel AVX or Intel AVX2 can process with a single instruction, and four times that of Streaming SIMD Extensions (SSE) (see Figure 1).

https://www.prowesscorp.com/what-is-intel-avx-512-and-why-does-it-matter/

[ravenshrike]

18 minutes ago, leadeater said:

They are all the same thing, AVX-512 might have some new instruction sets as part of it but the core of it is the same. AVX-512 is not different to AVX1 or AVX2, not computationally. The difference is the register size, how many floating point numbers fit in to the register. The more that fit in the register the more operations you can do per clock and that is where the performance increase comes from.

 

https://www.prowesscorp.com/what-is-intel-avx-512-and-why-does-it-matter/

With the advent of dedicated Vector cards how many workloads would benefit from vector calcs but not be large enough or dedicated enough to make switching to them a bad option?

 

https://www.google.com/amp/s/www.nextplatform.com/2017/11/22/deep-dive-necs-aurora-vector-engine/amp/

 

[leadeater]

2 hours ago, ravenshrike said:

With the advent of dedicated Vector cards how many workloads would benefit from vector calcs but not be large enough or dedicated enough to make switching to them a bad option?

 

https://www.google.com/amp/s/www.nextplatform.com/2017/11/22/deep-dive-necs-aurora-vector-engine/amp/

 

Well those sorts of things only make sense for the big compute people, the math itself is very common so having the capability within the CPU is still good. Whether developers utilize it or not.... well we know how that's been going already heh.

 

FPGAs and compute accelerators are already a thing, it's not just GPUs that get put in servers. Xeon Phi PCIe cards for example, Intel even has Skylake-SP processors with FPGAs on package now which make using them much more accessible then ever before.

 

The big hold up in the HPC space right now is memory bandwidth and general I/O bandwidth, CPUs are just too fast for how quickly we can actually move the data around. A lot of temporary measures have come around to try an mitigate those issues, like NVDIMM and Optane DIMM, and while actually very effective still has draw backs. You still need to copy the data set in to those faster storage layers before you can run the compute task.

 

You can have all the CPU horse power you like but if you can't use it then it goes to waste. Total job run time length is a really important factor when looking at the cluster efficiency, the brake down in phases and times lets you know where you need to improve to actually get more performance therefore more jobs processed from the job queue. One of the clusters we have has a job queue with an ETA over 500 days, and that's just for the existing jobs not new ones that might get added.

[Canada EH]

17 hours ago, Humbug said:

Expect some great 7nm APUs

Interesting

What will the price point be?

a) sand in open cut

b) sleeping on rocks

or c) not too bad

[asus killer]

edit: some italian website already tested zen 2 in games. That seems a little far-fetched 

 

 

 

 

[vanished]

That sounds like a pretty big jump by today's standards, but then they still have a fair ways to go to match Intel so it's not unthinkable they could get this much if they pulled it off.  I mean, just look at FX -> Zen.  I also tend to be particularly cautious of AMD rumours since they tend to get far more overhyped with no/fake evidence which only ends up letting people down when the real thing comes out, even if the real thing is great.

[GoldenLag]

4 hours ago, Taf the Ghost said:

APUs with "easy" to attach HBM would probably be for Apple.

could be interesting for Tiny 7nm vega + 7nm Ryzen for laptops. its probably one of the few things i really want from AMD at this point. 

[asus killer]

3 minutes ago, GoldenLag said:

its probably one of the few things i really want from AMD at this point. 

what about a good gpu 

[MMKing]

14 minutes ago, Ryan_Vickers said:

That sounds like a pretty big jump by today's standards, but then they still have a fair ways to go to match Intel so it's not unthinkable they could get this much if they pulled it off.  I mean, just look at FX -> Zen.  I also tend to be particularly cautious of AMD rumours since they tend to get far more overhyped with no/fake evidence which only ends up letting people down when the real thing comes out, even if the real thing is great.

Fair ways to go matching Intels IPC? Or fair way to go matching Intel's single thread performance?

 

I would hardly call a less than 3% difference a fair way to go, if you're referring to IPC. A 16% increase would put them 12.9% ahead of Intel's current 8700K.

 

Spoiler

 

[vanished]

Just now, MMKing said:

Fair ways to go matching Intels IPC? Or fair way to go matching Intel's single thread performance?

 

I would hardly call a less than 3% difference a fair way to go, if you're referring to IPC. A 16% increase would put them 12.9% ahead of Intel's current 8700K.

 

  Reveal hidden contents

 

Both, last I checked.  Granted I haven't looked closely at anything 8th gen or newer from Intel, or even Ryzen second gen, but from what I heard it wasn't enough to make up the gap that existed at the time.

[GoldenLag]

29 minutes ago, asus killer said:

some italian website already tested zen 2 in games. That seems a little far-fetched 

 

 

 

 

their spaghetti is entangled in any saucy news. so no surprise really

[GoldenLag]

6 minutes ago, MMKing said:

Fair ways to go matching Intels IPC? Or fair way to go matching Intel's single thread performance?

 

I would hardly call a less than 3% difference a fair way to go, if you're referring to IPC. A 16% increase would put them 12.9% ahead of Intel's current 8700K.

 

  Reveal hidden contents

 

that is a single benchmark. IPC varies between different workloads.

 

the IPC impovement between Zen and Zen + is quite nice considering is was essentially a die-shrink with minor improvements to cache. 

[Phentos]

4 minutes ago, GoldenLag said:

their spaghetti is entangled in any saucy news. so no surprise really

They've got their hands full with testing RTX effects on spicy meatballs slick with oil.

 

^{pretty high atm, don't judge me loudly.}

[GoldenLag]

Just now, Phentos said:

They've got their hands full with testing RTX effects on spicy meatballs slick with oil.

no wonder why the french arent leaking anything. people do not like them sticking their baguette into their business

[asus killer]

13 minutes ago, GoldenLag said:

their spaghetti is entangled in any saucy news. so no surprise really

the way they drive, probably fell of the back of a truck while doing a tight corner at 100kms/h

[Taf the Ghost]

49 minutes ago, GoldenLag said:

could be interesting for Tiny 7nm vega + 7nm Ryzen for laptops. its probably one of the few things i really want from AMD at this point. 

AMD will be using Navi derivatives by the time the 7nm APUs launch. There's a 2nd gen Raven Ridge coming in a few months which we don't know much about, since the APUs come during the off-cycle for AMD.

[porina]

6 hours ago, leadeater said:

I've actually seen a few tests where AVX-512 was slower than AVX2 on the Skylake-SP CPU tested, came down to being a very memory bandwidth demanding workload so only took a few cores doing AVX-512 to completely use all the bandwidth there was leaving a lot of cores unable to do much. DDR5 really needs to come quickly now.

I would have thought it would be no worse, in that you're just hitting a bottleneck sooner, unless there are more subtle things going on. DDR5 could be nice, but I'd settle for more ram channels in the short term.

5 hours ago, Taf the Ghost said:

AVX512 unit actually isn't the full instruction set just yet. That appears to be either Tigerlake or the one after that. Another reason AMD isn't also going to be adding it too soon.

As a non-programmer, the variations of AVX-512 that Intel have produced sounds like a nightmare. As a glass half full/half empty/twice as big as it needs to be type question, is there an end game to AVX-512 or are Intel just expanding it when they feel like?

4 hours ago, Stefan Payne said:

Also, as you can see with Skylake-X, it increases the Power Consumption dramatically...

Since we are already power limited, its not really a good idea right now.

I don't think we're power limited. We're ram bandwidth limited. High or low power has to be viewed in the context of how much work you're doing. AVX-512 does a LOT, if you can feed it.

4 hours ago, Stefan Payne said:

AVX is not AVX512.

Best way to imagine AVX-512 is that one unit of it has about the same potential as AVX2. Depending on the processor, each core may have one or two AVX-512 units. Skylake-X has two units per core, so AVX-512 on that has about double the potential throughput per clock of AVX2.

[leadeater]

18 minutes ago, porina said:

I would have thought it would be no worse, in that you're just hitting a bottleneck sooner, unless there are more subtle things going on. DDR5 could be nice, but I'd settle for more ram channels in the short term.

I'd have to try and find it again, I think it has to do with the larger swings in core loading and memory loading compared to more cores with a lower demand more evenly spreading the load. That and I think AVX-512 might have a lower boost offset? That's the annoying thing with actual workloads vs theoretical, real life can just trip you up for unforeseen reasons.

[porina]

6 minutes ago, leadeater said:

That and I think AVX-512 might have a lower boost offset? That's the annoying thing with actual workloads vs theoretical, real life can just be trip you up for unforeseen reasons.

Yes, clocks when running AVX-512 are typically lower than AVX, which is lower than none of the above. In that context, mixed code type running at the same time is not the best optimisation. You'd want to try and bunch up all the AVX work together.

 

Also, I think I saw that if you can't hit the AVX (in general) hard enough for long enough, you don't get its peak performance as it needs a bunch of clock cycles to fully wake up. I think it was in Agner Fog's CPU architecture guide. Even as a non-programmer, I find this stuff interesting.