MSI Drops First Hint of AMD Increasing AM4 CPU Core Counts

[NostaSeronx]

What the Threadripper line for then?
I thought these were MAINSTREAM chips.

Ryzen is Performance. Which is the cheaper form of Enthusiast. Which is categorized in its entirety under the banner of premium.
EPYC -> Not for consumers in a general sense. Uber Enthusiast.
Threadripper -> Enthusiast
Ryzen -> Performance

AMD's mainstream and essential/value product to this day is still the FX and A series. Ryzen and higher is aimed to get higher and higher ASPs. AMD would want increases in several dimensions to keep up the price ramp.

There is always need for more perf, or lower power. If it isn't improvable it isn't worth manufacturing.

Mores cores means more cores with higher frequency boosts. 100 percenting the entire SoC is bad, but >100 percenting a few parts of the SoC is good.

 

[bsp2020]

... I like a bunch of AI cores to go with the general purpose cores.

Until someone comes up with a decent AI cores, having multiple cores left over would help you run those AI tasks...

What AI tasks do you currently run that would benefit from AI cores?

 

[PeterScott]

Until someone comes up with a decent AI cores, having multiple cores left over would help you run those AI tasks...

What AI tasks do you currently run that would benefit from AI cores?

How about taking my whole post in context. I was talking about unforeseen options that might happen 5+ years in the future.

 

[itsmydamnation]

Well we can't figure out how to make stronger Lesser cores again...so lets just use more of them stuff.

Derp derp

I expect stronger, higher clocking and more cores from zen2 soc's.

 

[beginner99]

I'd much rather have 8 cores ,20% higher IPC, @ 4.8ghz.
That's the target audience.

Exactly. That it what AMD needs to challenge 8700k and the likes. If they offer me 16 "slow" cores for $500 and 8 for <$250 I get no benefit going with the top sku. If they offer 8 fast cores for $450 and 6 fast cores for $300 then I see an advantage in going with the top sku as 6vs8 will have some benefits to me. But 16 cores? no idea what I would need them for, except work stuff which I don't do on my home pc.

 

[f2bnp]

Lol, remember that just a year and a half ago a 4 core/8 thread CPU was the top end desktop CPU for 350$.

 

[aigomorla]

EPYC -> Not for consumers in a general sense. Uber Enthusiast.

I think you mean the term enterprise. To my knowledge you can not overclock a EPYC that's not a threadripper.

There is no Uber Enthusiast class, unless it means one who spends a ginormous amount of money retrofitting enterprise class for overclocking purpose.

The whole "Enthusiast" terminology takes root mostly from overclocking, which Enterprise class can not, which is why we got "overclocking enterprise" gear labeled as "Enthusiast", that was first introduced with SkullTrail from Intel back in 2007:
https://www.anandtech.com/show/2335/2

Then from there Intel found a new way to MILK us for a premium...
Charge Overclocking Enterprise Gear for more then Enterprise Price from the consumer and profit!
The following chips from then came out in 2 flavors... the LGA1366 (X58) which matured to the X299 today... and the LGA1156 which later evolved to LGA1155 (Z68) and is now currently today the Z270.

What the Threadripper line for then?

declaration of war to intel in the core wars.

In seriousness, that's AMD enthusiast line to contend against Intel's i9-79XX series.

 

[maddie]

der8auer says differently on EPYC overclocking.

 

[Headfoot]

core count for gaming is a chicken-and-the-egg problem. Until the high end gets more cores, so that the midrange can get more cores, so that the low end can get more cores, so that the # of cores across common install base goes up; no game developer will consider it.

In other words, now that everyone has dual cores you rarely see a game that doesn't use at least 2 cores anymore. Once midrange and up is 6 to 8 core across the board, and entry level is 4 core across the board you will see games use that by default.

Someone has to move first though.

Game console core count of course has an immense impact as well.

 

[Shivansps]

core count for gaming is a chicken-and-the-egg problem. Until the high end gets more cores, so that the midrange can get more cores, so that the low end can get more cores, so that the # of cores across common install base goes up; no game developer will consider it.

In other words, now that everyone has dual cores you rarely see a game that doesn't use at least 2 cores anymore. Once midrange and up is 6 to 8 core across the board, and entry level is 4 core across the board you will see games use that by default.

Someone has to move first though.

Game console core count of course has an immense impact as well.

Thats not enterely true, at least not anymore, that was from the 1C vs 2C and finally 2C vs 4C times, those times are long gone.

Howdays what they do is to create 4 threads: Main, Physics, Graphics, AI. And then each of then splits the work intro smaller "tasks" than can be assigned to any core by the OS.

But there is a limit, you cant expect games to use 16, 32, 64 cores... each task, each threads adds an overhead, there is a point were you start losing performance by trying to split the work like this.
Not to metion that performance per core will be always important, more cores tends to reduce frecuency.

Also this model can be also used on Dual cores, Quad cores and even single cores, but this is the reason of why 4 threads are needed for minimum playability on most games. And this is not going to change any time soon, not only 8 threads would need to be the lower end, the Main thread is a mayor problem when thinking about creating more mayor threads.

Now, there are a few exceptions to this, games that are running on old engines (Paradox for instance) that are cost prohibitive to upgrade, or they just dont care.

 

[PeterScott]

In other words, now that everyone has dual cores you rarely see a game that doesn't use at least 2 cores anymore. Once midrange and up is 6 to 8 core across the board, and entry level is 4 core across the board you will see games use that by default.

It doesn't really work like that. It isn't a case of programming for specific numbers of cores. It's a case of programming for multi-threaded where applicable.

Modern software will query the OS for the CPU count. When there is work that can be done in parallel. The will essentially spawn a number of threads to split that work depending on how many cores are deemed available.

Most games in the foreseeable future, are not going to see a huge performance boost from very high core counts, because most gaming problems are not highly parallel problems.

Let's look at an exception. A game that is mainly built around a highly parallel problem: Ashes of the Singularity. This is not a harbinger of things to come, nor programmers ahead of their time. It's a more unique problem space than most games.

AotS is basically about moving around thousands of independent units in parallel. It's as perfect a Parallel gaming problem as you will get.

3200 units/16 cores, so each core can work on moving 200 units.

It's the most parallel gaming workload on the market, and even it seems to stall out on high core counts. Why is that? Two things;

Bottlenecks and Amdahls Law. GPU will usually be a serious bottleneck but even when it isn't you have Amdahl's Law to contend with. Most games, IMO, will likely remain 50% Parallel max, meaning you will see some gains (but not 2X) moving from 4 cores to 8, but 8-16 will be insignificant.

I doubt even the most parallel games like AotS will get above 75% parallel overall, so gains will still be slight. 50-75% parallel are the bottom pink and blue lines on the graph:

Really, it's diminishing returns after quad cores, and isn't because programmer were coding for quad cores. It's the nature of mixed serieal/parallel problem spaces.

 

[aigomorla]

der8auer says differently on EPYC overclocking.

Show me a EPYC that is NOT threadripper that is overclocked on a EPYC board?

Im fairly sure you wont find it, because you dont overclock enterprise gear.

As i said there are extreme exceptions, like intel's cou*fake*gh 28 core 5ghz with sub ambient cooling that costs as much as the cpu itself, on a server board with a alpha bios specially coded by intel RnD.

That i would consider uber enthusiast gear.... because u would need to be sleeping with intel to get one.

 

[Headfoot]

It doesn't really work like that. It isn't a case of programming for specific numbers of cores. It's a case of programming for multi-threaded where applicable.

Modern software will query the OS for the CPU count. When there is work that can be done in parallel. The will essentially spawn a number of threads to split that work depending on how many cores are deemed available.

Most games in the foreseeable future, are not going to see a huge performance boost from very high core counts, because most gaming problems are not highly parallel problems.

Let's look at an exception. A game that is mainly built around a highly parallel problem: Ashes of the Singularity. This is not a harbinger of things to come, nor programmers ahead of their time. It's a more unique problem space than most games.

AotS is basically about moving around thousands of independent units in parallel. It's as perfect a Parallel gaming problem as you will get.

3200 units/16 cores, so each core can work on moving 200 units.

It's the most parallel gaming workload on the market, and even it seems to stall out on high core counts. Why is that? Two things;

Bottlenecks and Amdahls Law. GPU will usually be a serious bottleneck but even when it isn't you have Amdahl's Law to contend with. Most games, IMO, will likely remain 50% Parallel max, meaning you will see some gains (but not 2X) moving from 4 cores to 8, but 8-16 will be insignificant.

I doubt even the most parallel games like AotS will get above 75% parallel overall, so gains will still be slight. 50-75% parallel are the bottom pink and blue lines on the graph:

Really, it's diminishing returns after quad cores, and isn't because programmer were coding for quad cores. It's the nature of mixed serieal/parallel problem spaces.

For everyone eager to show off how much they know about multicore scaling - I understand Amdahl's law. Amdahl's law is not the rate limiting factor today - its still economics.

Economics still dominate. Just look at Ashes of the Singularity. Eventually multicore scaling will run out in games but there is absolutely zero chance that 2 cores is the limit. Bigger install base of bigger core counts will drive increase core count usage in games, no matter the theoretical view. We've already seen it happen with dual cores, saw it happen with xbo360 (3 core scaling), saw it happen with Xbone and PS4 (6 core scaling). So I'm gonna side with the real world evidence instead of the wikipedia copy paste on this one. In fact, I'll bet anyone here money that if the next gen of consoles comes out with more than 8 cores, average number of cores used across all games released will steadily improve after the release of that console.

 

[ryzenmaster]

Amdahl's law is a good reminder there's no infinite scalability. It does make assumptions though that do not always apply. I'll just leave this direct quote from its Wikipedia article:

Amdahl's law applies only to the cases where the problem size is fixed. In practice, as more computing resources become available, they tend to get used on larger problems (larger datasets), and the time spent in the parallelizable part often grows much faster than the inherently serial work. In this case, Gustafson's law gives a less pessimistic and more realistic assessment of the parallel performance.

 

[PeterScott]

Amdahl's law is a good reminder there's no infinite scalability. It does make assumptions though that do not always apply. I'll just leave this direct quote from its Wikipedia article:

Ultimately Amdahl's law applies to every specific program. Gustafson's law is no panacea, and is a poor fit for gaming.

It doesn't give you faster frame frame rates for more cores, it just argues that EVENTUALLY you will work on bigger data-sets so the parallel work will slow things down enough that the parallel portion of the code dominates again. It isn't about shrinking the serial bottleneck at all, and you still get diminishing returns.

 

[IEC]

People arguing that we aren't going much beyond 6 cores any time soon are just as short-sighted as all the folks in years past arguing that a dual core and a quad core i5 would be more than enough for literally everything for 10+ years.

They were wrong then, and they are wrong now.

As the pace of technological innovation and development continues to increase, it only stands to reason that 8+ cores will be mainstream within the next 5 years. Hell, 6+ cores are *already* common in phones, and consoles are pushing 8. Core counts will increase and the BIG.little arrangements in smartphones and/or 2/3/4-core turbo arrangements in desktop chips will address the need for lightly-threaded performance while the extra cores provide the horsepower needed for CPU-heavy tasks. Advanced power management techniques and 14 nm processes mean that 8+ core counts are already very attainable at very reasonable cost. Even more so once 7nm processes mature.

I have seen the future (read: not publicly/readily available, yet) in AR and VR devices, and although the technology is still in its infancy in relative terms, I predict that the market will approach or even exceed the smartphone market within the next 5-10 years. We're only scratching the surface in terms of what is possible.

 

[maddie]

Not claiming anything here about # cores needed for the future, but one thing to keep in mind about new and original breakthroughs, are that for the majority, they only seem possible in retrospect. Experts of the present are often confounded by this.

 

[PeterScott]

People arguing that we aren't going much beyond 6 cores any time soon are just as short-sighted as all the folks in years past arguing that a dual core and a quad core i5 would be more than enough for literally everything for 10+ years.

That isn't what people are arguing. More cores than 6 are coming.

People are arguing that diminishing returns are going to make those extra cores of negligible value outside of a few special cases.

 

[french toast]

Games are probably bottlenecked by the CPU are they not? I mean in regards to newer bigger and more complex world's and enviroments.
Many more draw calls, uber complex AI and behaviours, animation, massive numbers of non player characters each with complex independent reactions, expressions, speech, behaviour and schedule.
Same with animals and wildlife, better more complex physics, VR and AR.
In ten years time we will look back at these comments about multicore scaling in games and laugh at ourselves, because we don't yet have the complete picture of what is going to be possible once 8+ multicores (large) become mainstream in consoles and PCs and new software and techniques become available.

As soon as the consoles move to 8 big cores + some kind of AI acceleration, we won't know what is coming next, all the games on PC are held back in SCOPE and COMPLEXITY by the consoles, which in turn are bottlenecked by their meagre CPU resources.

Bring on MOAR cores I say.

 

[Cerb]

That isn't what people are arguing. More cores than 6 are coming.

People are arguing that diminishing returns are going to make those extra cores of negligible value outside of a few special cases.

Maybe, but faster cores just aren't happening, at least not on the silicon substrates we know. So, we can take the poorer than perfect scaling, or buy cheaper CPUs with fewer cores.

 

[maddie]

Maybe, but faster cores just aren't happening, at least not on the silicon substrates we know. So, we can take the poorer than perfect scaling, or buy cheaper CPUs with fewer cores.

And this is why the ability to disintegrate the SOC and then re-integrate it with an interposer will be important. Separate substrates can be mixed allowing possibilities which are simply impossible on a single one. A few extremely high clocked non-silicon cores can be a chiplet. Slower contemporary cores can stay on Silicon.

We can get the best of both worlds.

 

[Cerb]

And this is why the ability to disintegrate the SOC and then re-integrate it with an interposer will be important. Separate substrates can be mixed allowing possibilities which are simply impossible on a single one. A few extremely high clocked non-silicon cores can be a chiplet. Slower contemporary cores can stay on Silicon.

We can get the best of both worlds.

If we drop silicon as we know it, it would likely be better to just move on completely, not mix and match on a single package.

 

[ChiefBigFeather]

Most games in the foreseeable future, are not going to see a huge performance boost from very high core counts, because most gaming problems are not highly parallel problems.

No, they are mostly cache optimization problems. Most CPU bound games are not CPU bound because of single core throughput, like say pi fast is. They are limited because they try to do things that like to produce a lot of cache misses. This is why Amdahls Law can be misleading when looking at scaling on multi core systems. It isn't really clear if the extra cores or the extra cache is causing the speedup. On the software side multi threading might not benefit the underlying mathematical problem but it might lead to unexpected gains because the extra threads fetch data causing less idle time across all cores.