AMD Zen 3 Rumored to Double Thread Count per Core with SMT4

Tsing

The FPS Review
Staff member
Joined
May 6, 2019
Messages
12,871
Points
113
AMD's next-generation CPU microarchitecture may come with huge improvements to multithreading. According to rumors from Hardwareluxx, Zen 3 will introduce a new level of simultaneous multithreading (SMT) that doubles execution thread count per core from two to four, enabling 64C/256T chips.

This would make the third generation of EPYC processors ("Milan") especially attractive to cloud, enterprise, and other high-performance computing sectors, which primarily benefit from SMT. AMD has already confirmed that the design for Zen 3 is complete.

AMD won’t be the first ever company to do this kind of thing. some iterations of IBM’s Power architecture support up to eight execution threads per core. However, if the rumors are to be believed, AMD will be the first ever to introduce an x86 microarchitecture capable of executing more than two threads per core.
 
stickx911 said:
Woo! They need to start kindling the same fire in the gpu department.
Click to expand...
They're working on it slowly, Lisa Su said she was focusing on CPU first, hopefully they start putting some of that towards GPUs. I think Navi is a good start (kind of like ryzen 1), but not yet putting lots of pressure on Nvidia. With the 5300 coming out and hopefully a 5500, we will be some replacements for the rx 5x0 series. I'm sure with the pending release of next gen consoles next year they've been focused on these; I'm sure some of that labor will come to PCs, hopefully in the shape of higher end cards.


And on topic:. I read this elsewhere and think it's a great move if they can just some real world results. Since AMD currently benefits more from SMT than Intel (due to architecture design), I'm hopeful that it will benefit, especially in the server market where you can have lots of unrelated tasks that use different parts of the CPU at the same time.

Don't know how useful this would be for every day use though.
 
As a consumer this makes the 8 core currently 16 thread Ryzen CPU even MORE appealing. If there truly is wiggle room in the instruction set to scale up to 8 core 32 thread and not murder my performance I would be all about it.

Especially if the OS were intelligent enough to allocate my physical cores to the most demanding workload (whatever processes need the most cpu) and the rest to everything else just touching the cpu on a blue moon comparatively that would be really awesome.

My worry is in this case will my primary thread CPU's start suffering because of compute path's dedicated to these other 'SMT' cores?

I don't know if many of you remember how Hyperthreading/SMT became a thing.

Pentiums were very fast CPU's during their time. But CPU manufacturers started to realize that having all of the specialized instruction sets meant they needed general compute as part of them. Eventually dual compute paths were the normal in the high end Pentium CPU's (before hyperthreading was realized) These dual compute threads eventually became hyperthreading or SMT. They literally ARE the same thing because you are utilizing multiple threads running concurrently on the same compute core.

Now if AMD's edge really is due to having 4 compute paths that run in parallel (Symmetrical Multi Threading) then they could realisticly split this off with little to no loss in overall performance of any 1 thread.

I wonder though what the impact will be with the shared Cache and such that each 'core' has access too. Will this make the CPU more vulnerable?

So to summarize.... The risks I see are:

1. SMT or Hyperthreading should only be used IF there truly are that many available compute paths on a core that can be ran at the same time.
2. Throttling might take place for the CPU cache depending on how fast they can access it.
3. Encryption sucks for speed so that might be a bottleneck on the AMD CPU's but it will also be more secure.

Thoughts?
 
I'm kind of surprised this isn't generating more chatter.
 
Grimlakin said:
I'm kind of surprised this isn't generating more chatter.
Click to expand...

All depends on if there is performance gain. I mean, you could right now say your computer is a bajillion threads, because the OS supports multi-threading. The hardware can't really simultaneously process all of those threads... but it can't really do that with SMT either, the gain from SMT isn't nearly as good as the gain from adding a physical core, it's only 10-40% (depending on the situation)

If this shows an actual productivity increase, then I'm all for it.

If it's just a bulletpoint so that AMD can say they have more threads, but we don't see any meaningful increase in productivity, then meh.
 
Brian_B said:
All depends on if there is performance gain. I mean, you could right now say your computer is a bajillion threads, because the OS supports multi-threading. The hardware can't really simultaneously process all of those threads... but it can't really do that with SMT either, the gain from SMT isn't nearly as good as the gain from adding a physical core, it's only 10-40% (depending on the situation)

If this shows an actual productivity increase, then I'm all for it.

If it's just a bulletpoint so that AMD can say they have more threads, but we don't see any meaningful increase in productivity, then meh.
Click to expand...


Brian you really should take a moment and read about how SMT works today. It's much more than 40% on heavily threaded workloads.
 
I think this will be more useful in server loads and/or hedt than normal at home work loads
 
Ready4Droid said:
I think this will be more useful in server loads and/or hedt than normal at home work loads
Click to expand...

I can tell you 100% without a doubt that this is true.

This will change if and when we get home based installed Hypervisor servers that drive all of the devices in the home.
 
Grimlakin said:
Brian you really should take a moment and read about how SMT works today. It's much more than 40% on heavily threaded workloads.
Click to expand...
www.techspot.com

How Screwed is Intel without Hyper-Threading?

Today we're exploring the impact disabling Hyper-Threading has on Intel processors. We've done this in the past and it's an interesting test, however there are new incentives...
www.techspot.com www.techspot.com

Not sure what you would consider "heavily threaded". This report just looks at Intel, and throws the kitchen sink at it with respect to workloads, and sees 25-35% deltas between SMT and non-SMT.

www.techspot.com

Testing Ryzen 9 with SMT On vs. SMT Off

When we compared AMD's Ryzen 9 3900X gaming head to head against the Core i9-9900K we found that the 12-core AMD processor was about 6% slower. We've...
www.techspot.com www.techspot.com

This one deals with Ryzen Zen 2 on the 12 core Ryzen 9 3900X... and found that results were all over the place. It just tests games, so it's not quite as comprehensive as the Intel review, but ... 1% faster improvement with SMT disabled overall.



AMD Ryzen 9 3900X SMT Linux Performance Benchmarks - Phoronix

www.phoronix.com www.phoronix.com

Here's one from Phornix that throws a lot of various workloads at a Ryzne 9 3900X. Overall change was on average 4% gain from SMT, but you can see one outlier in there that can take advantage of the threads (parboil). Everything else was <40% if it saw improvement at all.

Got any better data than that? I'm all ears, but I think I was being generous with 40%. It appears you can cherry pick very specific applications, but by and large SMT is not that big of a deal.
 
Brian_B said:
[/URL]

Not sure what you would consider "heavily threaded". This report just looks at Intel, and throws the kitchen sink at it with respect to workloads, and sees 25-35% deltas between SMT and non-SMT.

[/URL]

This one deals with Ryzen Zen 2 on the 12 core Ryzen 9 3900X... and found that results were all over the place. It just tests games, so it's not quite as comprehensive as the Intel review, but ... 1% faster improvement with SMT disabled overall.



[/URL]

Here's one from Phornix that throws a lot of various workloads at a Ryzne 9 3900X. Overall change was on average 4% gain from SMT, but you can see one outlier in there that can take advantage of the threads (parboil). Everything else was <40% if it saw improvement at all.

Got any better data than that? I'm all ears, but I think I was being generous with 40%. It appears you can cherry pick very specific applications, but by and large SMT is not that big of a deal.
Click to expand...


Well for me when I think SMT I think Hypervisor level. It makes a huge difference for that. Trust me I know very well the impact is insane when you loose half of your processing threads on a system that is actually hosting many VM's.

I suppose I should have been specific.
 
You must log in or register to reply here.
Become a Patron!
Back
Top