AMD possibly going to 4 threads per core

IdiotInCharge said:
About as hilarious as the lawsuit they just lost for marketing Bulldozer quad-cores as eight-core CPUs, when they were slower than Intel's quad-core CPUs?

I thought that was hilarious the entire time, after I got over AMD throwing away their good architecture and consigning themselves to irrelevance in the CPU space for a decade ;)
Click to expand...

Fair... although I wouldn't say Bulldozer was even that bad a design. It was actually a good design for the time... they got out fabbed by Intel no doubt. Intel had the better (ok much better) parts for awhile. AMDs FX chips where hardly terrible though

Still I'm not suggesting they market their 8 cores as 32 cores... no they need to market them exactly as what they are. Which if true would be 4 thread per core chips.

If the rumor is true I am looking forward to what Zen3 will bring to the table. AMD may have been thinking outside of the box and may push some new stuff. I don't believe SMT4 will bring a ton of real world performance... HOWEVER I could forsee some interesting implementations made possible by chiplet design. We have only really seen SMT4/8 from IBM so far and the gains there are very dependent on application. Still if AMD is able to use 7nm+ to take a 9 core chiplet as they have now and squeeze even more cache onboard in combination with an even longer prefetch history... its possible AMD might get some very nice performance gains with SMT4 even for things not traditionally effected by SMT. Games for instance mostly suck with SMT due to cache issues.... the prefetch units don't have enough history space for real time gaming loads and the split often hurts performance rather then helping. (in practice SMT is halving the amount of cache per thread, and we all know game loads are heavily influenced by cache) It is however very possible that with another doubling of cache space and 20-40% longer cache history on AMDs TAGE prefetch engine we might actually see nice FPS gains with SMT4. (I would say the advantage AMD has with SMT over Intel, and I know its not a huge advantage... is likely due to increased cache space and longer prefetch pipes as a result and little else)

Anyway I give this rumor a 50/50 of being true at best... and even if it is I would also say it could well be a 50/50 that this is a Epyc feature and the consumer Ryzens have the SMT4 unit disabled.
 
ChadD said:
Fair... although I wouldn't say Bulldozer was even that bad a design. It was actually a good design for the time... they got out fabbed by Intel no doubt. Intel had the better (ok much better) parts for awhile. AMDs FX chips where hardly terrible though.
Click to expand...

I'm on record stating that I like the idea, and I honestly still do- but not only did they get out-fabbed, which is to be expected with Intel, but their architecture while quite functional simply didn't compete. And then they mismarketed it.

ChadD said:
Still I'm not suggesting they market their 8 cores as 32 cores... no they need to market them exactly as what they are. Which if true would be 4 thread per core chips.
Click to expand...

Didn't mean to apply you were, you said 8C32T, and that's what I'm reading :).

ChadD said:
If the rumor is true I am looking forward to what Zen3 will bring to the table. AMD may have been thinking outside of the box and may push some new stuff.
Click to expand...

Their chiplet setup with Zen 2 caught me by surprise, quite honestly. I didn't expect it to do so well, and now I find myself only recommending Intel in edge cases. I do hope they continue to think outside the box here, and SMT4 isn't a bad start.

ChadD said:
I don't believe SMT4 will bring a ton of real world performance... HOWEVER I could forsee some interesting implementations made possible by chiplet design ... Still if AMD is able to use 7nm+ to take a 9 core chiplet as they have now and squeeze even more cache onboard in combination with an even longer prefetch history... its possible AMD might get some very nice performance gains with SMT4 even for things not traditionally effected by SMT. Games for instance mostly suck with SMT due to cache issues.... the prefetch units don't have enough history space for real time gaming loads and the split often hurts performance rather then helping. It is however very possible that with another doubling of cache space and 20-40% longer cache history on AMDs TAGE prefetch engine we might actually see nice FPS gains with SMT4.
Click to expand...

Agreed- usually cache is extremely expensive, but with chiplets it really becomes economical again, and if AMD is able to leverage that configuration to increase IPC -- well damn. Might have found myself a reason to upgrade!

ChadD said:
We have only really seen SMT4/8 from IBM so far and the gains there are very dependent on application.
Click to expand...

Splitting this out to say that at least we've seen it done in production, and if the cost of going from SMT2 to SMT4 (I guess this is our naming scheme?) is minimal, then why not, right?

ChadD said:
Anyway I give this rumor a 50/50 of being true at best... and even if it is I would also say it could well be a 50/50 that this is a Epyc feature and the consumer Ryzens have the SMT4 unit disabled.
Click to expand...

I don't see a reason that they couldn't adjust between SMT2/3/4 as needed for SKUs -- or even as needed on the fly based on processing needs.

One concern I have with increasing SMT levels is that it should result in higher CPU utilization, which will then result in higher power draw and heat generation, and that could become a limiting factor. Do you allow more threads to be assigned to a core, or do you increase core speed with fewer threads, when trying to balance TDP? And so on.
 
IdiotInCharge said:
Splitting this out to say that at least we've seen it done in production, and if the cost of going from SMT2 to SMT4 (I guess this is our naming scheme?) is minimal, then why not, right?

I don't see a reason that they couldn't adjust between SMT2/3/4 as needed for SKUs -- or even as needed on the fly based on processing needs.

One concern I have with increasing SMT levels is that it should result in higher CPU utilization, which will then result in higher power draw and heat generation, and that could become a limiting factor. Do you allow more threads to be assigned to a core, or do you increase core speed with fewer threads, when trying to balance TDP? And so on.
Click to expand...

Ya its possible they do like Intel and sell chips with SMT4 turned off... in the way Intel sells a bunch of non smt versions. If it's a really touchy bit silicon wise they could easily bin the best chips to Epyc and high end Ryzen and have SMT2 on the mid range and down. We'll see still a long shot rumor I think.

Oh in the IBM case as well talking about cache. Power 9 and IBMs SMT4/8 design uses dual L1 cache space (which intel moved to and so has AMD)... and a shortened pipeline (which reduces the amount of cache required) And well of course the 8 core intel Power 9s have 80mb of L3 Cache. I imagine for AMD to consider going SMT4 they are going to want to at least double the L3 cache on their 8 core chiplet to 64MB... they already have double the L1 and 4x the L2 cache vs IBMs power9. I'm not an engineer I would just think with the longer x86 pipelines SMT4 and beyond would really hit the cache hard. Guess it depends how much better 7nm+ is... I doubt the squeeze twice as much cache in within the same die space, but I guess if they could double it with only a smallish increase in die size it might make good sense.
 
You know. Axe my wishes for a hardware level caching optimizer.

I'd rather the arch stay the same and put all the effort and money into moving off silicone and into carbon. That alone without any changes to what the cpu is doing would offer a many fold increase in performance and efficiency far beyond whatever can be expected from adding cores.


But if more cores is all we are going to get smt4 will presumably utilize the super long pipelines the upcoming arch changes will bring, which is great. But I'd love to hear how we're suppose to cool a hyper utilized cpu putting out ~200 watts in the area we currently struggle removing 120 watts from.
 
M76 said:
It's called denial.
Click to expand...
Sorry but this BS that Ryzen is faster than Intel clock for clock in everything needs to stop. Those idiotic "not an apple fan" and "moore's law is dead" youtubers claim Ryzen is 13% faster than Intel clock for clock and that is horseshit especially in gaming. Hardwareunboxed did a video running both cpus at 4.0 and Intel was still a hair faster overall in games.
 
misterbobby said:
Sorry but this BS that Ryzen is faster than Intel clock for clock in everything needs to stop. Those idiotic "not an apple fan" and "moore's law is dead" youtubers claim Ryzen is 13% faster than Intel clock for clock and that is horseshit especially in gaming. Hardwareunboxed did a video running both cpus at 4.0 and Intel was still a hair faster overall in games.
Click to expand...
That sentence is gold.

"It's not faster in everyting ... at least it's not faster in gaming ... intel is a hair faster on aggregate in games"

Yeah, AMD is faster in everything that matters, Shadow of the Tomb raider keeps the overall score on intel's side. But in every good statistic outliers are eliminated.

And if we compare $ to $ it's not even worth comparing. These are terrible days to be an intel fanboy indeed.
 
M76 said:
That sentence is gold.

"It's not faster in everyting ... at least it's not faster in gaming ... intel is a hair faster on aggregate in games"

Yeah, AMD is faster in everything that matters, Shadow of the Tomb raider keeps the overall score on intel's side. But in every good statistic outliers are eliminated.

And if we compare $ to $ it's not even worth comparing. These are terrible days to be an intel fanboy indeed.
Click to expand...

You called him in denial and the comment he was replying to was mentioning that Ryzen was faster clock for clock at everything including gaming. It is NOT faster clock for clock in gaming and at the same 4.0 clocks Intel beat or in worst case matched Ryzen in the Hardware Unboxed article.

And I chose the 9900kf because I was not interested in tweaking and having to worry about BIOS updates left and right to address the teething issues. I recommend AMD setups quite a bit but people like you lack the capacity to understand there are pros and cons to everything and it has nothing to do with "fanboy".
 
Last edited:
IdiotInCharge said:
No, I want to compare IPC.



It wasn't clear that you knew, as your wording called that into question.
Click to expand...

https://linustechtips.com/main/topic/946917-intel-ht-vs-amd-smt-scaling/

Some impartial data... This was with a 1600, newer zen should be at least as good.

Good read on technical differences... Start @ paragraph 5 for details.
https://www.agner.org/optimize/blog/read.php?i=838

Edit:. Some more info, I think second image was the % difference SMT for zen and Skylake.
https://www.overclockers.com/forums/showthread.php/792250-Skylake-vs-Zen-vs-Zen-HT-SMT
 
Last edited:
PS. This is not an IPC comparison/rumor. Please attempt to keep this on SMT. I think we can all agree AMD and Intel are close in IPC overall and can go one way or the other depending on benchmark and work load. Let's leave it at that.
 
misterbobby said:
Sorry but this BS that Ryzen is faster than Intel clock for clock in everything needs to stop. Those idiotic "not an apple fan" and "moore's law is dead" youtubers claim Ryzen is 13% faster than Intel clock for clock and that is horseshit especially in gaming. Hardwareunboxed did a video running both cpus at 4.0 and Intel was still a hair faster overall in games.
Click to expand...

I don't think very many people disagreed on games.
But yeah everything else, Ryzen 3 IPC is definetly higher

But to try to stay on topic, I think SMT4 depending on how expensive in terms of die space compared to adding cores could be good. Anyone know how much die space SMT4/8 adds in Powers arch? Compared to their cores?
 
  • Like
Reactions: N4CR
like this
To make it work I expect that AMD will have to increase the number of execution units per core. This should also increase the IPC. I would expect this is where most of the die space increase will be.
 
Rockenrooster said:
Currently Ryzen 3000 has better IPC than anything Intel has. The percentage difference depends on the type of application. Games have a smaller difference, others have a much larger difference.
Click to expand...

IdiotInCharge said:
That'd be a nope.
Click to expand...

Legitreviews shows the 9900k at 4Ghz with a fairly certain lead on IPC

https://www.legitreviews.com/ipc-battle-amd-ryzen-9-3900x-vs-intel-core-i9-9900k_213719

and a slight lead in IPC for the the 9900k at 4.0Ghz at hothardware.
https://amp.hothardware.com/reviews/amd-ryzen-9-3900x-vs-core-i9-9900k-ipc-shootout?page=3
 
M76 said:
That sentence is gold.

"It's not faster in everyting ... at least it's not faster in gaming ... intel is a hair faster on aggregate in games"

Yeah, AMD is faster in everything that matters, Shadow of the Tomb raider keeps the overall score on intel's side. But in every good statistic outliers are eliminated.

And if we compare $ to $ it's not even worth comparing. These are terrible days to be an intel fanboy indeed.
Click to expand...

man 2016 must've been so nice for them
 
Archaea said:
Legitreviews shows the 9900k at 4Ghz with a fairly certain lead on IPC

https://www.legitreviews.com/ipc-battle-amd-ryzen-9-3900x-vs-intel-core-i9-9900k_213719

and a slight lead in IPC for the the 9900k at 4.0Ghz at hothardware.
https://amp.hothardware.com/reviews/amd-ryzen-9-3900x-vs-core-i9-9900k-ipc-shootout?page=3
Click to expand...

"Laughs at using BAPCo SYSmark"
Hothardware's review is much better.
Also Anandtech's review is pretty good.
I misspoke about gaming but in productivity apps AMD has the edge. Games= Intel has the edge. Vulerabilities= Intel has the edge lol.
Both are fast, both are expensive, but in reality ever since Ryzen came out things are changing real quick.
 
M76 said:
It's called denial.
Click to expand...

It's supported pretty clearly in benchmarks. AMD has put on a great show and released great products, but even three revisions of Zen in, now on a 'more advanced process' even, they've only just approached parity with Intel's five-year-old Skylake architecture. That's really only applaudable in light of the step backward they took with Bulldozer and Intel's 10nm stumble. Had Intel not stumbled, AMD might likely have been bought out by now, as Intel's 10nm cores have shown significant IPC increases over Skylake and Zen 2.

M76 said:
These are terrible days to be an intel fanboy indeed.
Click to expand...

There's really no need to troll. The only reason to post a sentence like this is to incite an emotional response -- it does not further discussion, and it further exposes yourself as not interested in actual discussion but rather posting incendiary comments meant to distract from the topic.

Rockenrooster said:
I misspoke about gaming but in productivity apps AMD has the edge.
Click to expand...

This is true in general. If a workload is mostly or entirely latency-insensitive CPU compute work, then it's going to run a bit better on an equally-cored, equally-clocked AMD part, and that part is going to be cheaper than Intel's.

Rockenrooster said:
Games= Intel has the edge.
Click to expand...

This is pretty resolutely true, but it also only really matters in the margins. Very hard to argue for Intel for gaming if that gaming target isn't max FPS and minimum frametimes, backed up by a high-refresh monitor and a top-end GPU. Otherwise, AMD brings the performance needed at a lower price.

Rockenrooster said:
Vulerabilities= Intel has the edge lol.
Click to expand...

For the moment- security experts have had years to pick apart Skylake, whereas Zen is relatively new, and the scholarly knowledge on vulnerabilities has also evolved over the time since Skylake's release, which obviously helped AMD design around newer known vulnerabilities than Intel could with Skylake. However, it should be noted that Intel's 10nm architecture, which also predates Zen, also accounts for the vulnerabilities present in Skylake. Obviously being late hasn't helped Intel in that department, though, as 10nm Intel parts can currently only be found in ultrabooks.

Rockenrooster said:
Both are fast, both are expensive, but in reality ever since Ryzen came out things are changing real quick.
Click to expand...

Ryzen has no doubt put pressure on Intel, but the pressure was already there- they were having to create new SKUs for 14nm revisions because the planned 10nm releases weren't happening. Kaby Lake and Coffee Lake weren't on the map, as Intel had expected to release eight-core 10nm Ice Lake arch CPUs on the desktop to succeed Skylake. They failed, and they ran with contingencies- that weren't all bad!- and they're paying for their failure in every market except mobile where AMD has not invested the R&D to actually compete.


And speaking of mobile: would anyone speculate that SMT4 might increase performance per watt, i.e., allow for performance to increase while keeping clockspeeds and voltages down with the goal of increasing battery life?
 
IdiotInCharge said:
And speaking of mobile: would anyone speculate that SMT4 might increase performance per watt, i.e., allow for performance to increase while keeping clockspeeds and voltages down with the goal of increasing battery life?
Click to expand...

I can see it now...
Bestbuy dude: "this Ultrabook has 8 cores!!!"
Me: "You mean 2 cores with SMT4?"
Bestbuy dude: "Wut?"

In all seriousness, I think it would help a quad core with 16 threads I would think will perform better and be more effiecient than a 4/8. Maybe even a 6/6, I would like to see a comparasion of that if it ever happens.
 
Lakados said:
For virtualizing I just need core counts and ram, speed is completely irrelivant 99% of the time. so a 32/128 would be a wet dream for me, or god forbid a 64/256...... Sploosh
Click to expand...


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?
 
IdiotInCharge said:
One concern I have with increasing SMT levels is that it should result in higher CPU utilization, which will then result in higher power draw and heat generation, and that could become a limiting factor. Do you allow more threads to be assigned to a core, or do you increase core speed with fewer threads, when trying to balance TDP? And so on.
Click to expand...

i could definitely see it being load controlled where the extra threads are treated as overflow.. so it'll prioritize 2 threads per core for higher clocks til something needs a thread then enables a 3rd or 4th so the cores slightly clock down due to extra load then parks the thread when it's not needed.. but the real question is how long would it take microsoft to unfuck their scheduler to actually make SMT4 work? AMD seems to always be held back due to that, we saw it with bulldozer and zen and we're seeing it again with zen 2 not prioritizing the correct cores. all be it we're talking single digit percentage performance differences but still.
 
4 way SMT/HT might have a place in a server or specialized workload/app, but this will not be helpful for desktop users.

one of the main reasons behind really high thread to core ratios is that some enterprise software has pricing that scales with the number of cores. Eight threads per core lets you get the most out of the cores you're paying for with very parallelized software. Consumer software rarely involves this level of rampant bullshit so it tends towards lower ratios.
Click to expand...

SMT's purpose is to put to work idle processing pipeline/capacity.

Haven't read the scaling articles yet, but 1 core is 1 core. If a single thread is keeping it busy, SMT/HT already buys you nothing.

And with 8 core cpu's being reasonably priced, for us, you will not want or need 4way SMT.

HyperThreading (HT) and Simultaneous Multi-Threading (SMT) are effective features for hiding latency through concurrency. In high-performance computing systems with low-power cores - e.g. IBM Blue Gene/Q and Intel Xeon Phi x100 coprocessors (aka Knights Corner) - 4-way threading hides instruction latency in addition to latencies associated with memory and I/O.

In server parts like Intel Xeon processors, IBM POWER processors, Sun Sparc processors, instruction issue latency is not a bottleneck and multiple threads per core is used to hide memory and I/O latency. A side effect of the latter latency hiding capability is that cloud operators can charge by the thread (sometimes called a virtual core) rather than the physical core.

However, adding arbitrary amounts of hardware threading to CPU cores is imprudent, since each thread requires its own architectural state (e.g. registers), and this state consumes transistors and power while increasing design complexity. In a processor designed for maximum single threaded performance per watt or cost, additional hardware thread per core may not make sense the way they do in server processors.
Click to expand...
 
ChadD said:
Fair... although I wouldn't say Bulldozer was even that bad a design. It was actually a good design for the time... they got out fabbed by Intel no doubt. Intel had the better (ok much better) parts for awhile. AMDs FX chips where hardly terrible though

Still I'm not suggesting they market their 8 cores as 32 cores... no they need to market them exactly as what they are. Which if true would be 4 thread per core chips.

If the rumor is true I am looking forward to what Zen3 will bring to the table. AMD may have been thinking outside of the box and may push some new stuff. I don't believe SMT4 will bring a ton of real world performance... HOWEVER I could forsee some interesting implementations made possible by chiplet design. We have only really seen SMT4/8 from IBM so far and the gains there are very dependent on application. Still if AMD is able to use 7nm+ to take a 9 core chiplet as they have now and squeeze even more cache onboard in combination with an even longer prefetch history... its possible AMD might get some very nice performance gains with SMT4 even for things not traditionally effected by SMT. Games for instance mostly suck with SMT due to cache issues.... the prefetch units don't have enough history space for real time gaming loads and the split often hurts performance rather then helping. (in practice SMT is halving the amount of cache per thread, and we all know game loads are heavily influenced by cache) It is however very possible that with another doubling of cache space and 20-40% longer cache history on AMDs TAGE prefetch engine we might actually see nice FPS gains with SMT4. (I would say the advantage AMD has with SMT over Intel, and I know its not a huge advantage... is likely due to increased cache space and longer prefetch pipes as a result and little else)

Anyway I give this rumor a 50/50 of being true at best... and even if it is I would also say it could well be a 50/50 that this is a Epyc feature and the consumer Ryzens have the SMT4 unit disabled.
Click to expand...

AMD didn't necessarily get outdone by Intel, remember Intel cheated by skimping on security so it's unclear, to me anyway, as to what degree AMD got outperformed in that era. I wonder if anyone has gone back to the Bulldozer era CPU's and re-run benchmarks with all of the Intel mitigations (including disabling hyperthreading) in various desktop/workstation/server workloads against Bulldozer?

Also I wonder what would happen if you ran an 8 core Bulldozer vs a fully mitigated Intel CPU (including disabling hyperthreading) of that era in a modern game that utilizes more than 4 cores?

Anyway this level of SMT that we're talking about seems like it belongs in heavily virtualized environments to better saturate CPU resources - not sure where else this would go?
 
  • Like
Reactions: ChadD
like this
remember Intel cheated by skimping on security
Click to expand...

I would argue that there is no evidence of that at all. Only speculation from AMD fans.

Meltdown was a bug however.
 
IdiotInCharge

SMT Yield results across a number of workloads by The Stilt:

Source: https://www.overclock.net/forum/10-amd-cpus/1728758-strictly-technical-matisse-not-really.html
The takeaway:
dD9y5H2.png
 
In general (from the data available that I've seen), AMD has better SMT utilization. This is dependent on work loads of course, but majority and average has AMD comfortably leading. Depending on what other technical changes they make, it could work out well or not at all. All speculation of course, and it may not even be anything, because they could almost say they already have SMT4 due to being able to perform 4 independent ops (2 integer + 2 float) per cycle vs Intel who can do 2 (1 integer + 1 float). Intel can do double the precision per operation though, so if it's double precision then AMD can only do 1 fpu op per cycle (which would make it difficult to claim SMT4). So worst case AMD should be on par with Intel, in best case it could potentially be twice as good. Since worst/best case are rarely reality (and a lot more things have effects), it falls in-between more often than not.
 
IdiotInCharge said:
There's really no need to troll. The only reason to post a sentence like this is to incite an emotional response -- it does not further discussion, and it further exposes yourself as not interested in actual discussion but rather posting incendiary comments meant to distract from the topic.
Click to expand...
People are already emotional that's why they are trying to downplay the leap that Zen3 is in the consumer space. I'm just glad I have no emotional attachment to any company so I can switch at any time without having to explain anything to anyone.
 
guys... we haven't seen performance parity between CPU vendors like this since... what Cyrix, AMD, and Intel were in competition? Ok AMD and Intel really. Most of you young bucks probably don't remember Cyrix at all.
 
pillagenburn said:
AMD didn't necessarily get outdone by Intel, remember Intel cheated by skimping on security so it's unclear, to me anyway, as to what degree AMD got outperformed in that era. I wonder if anyone has gone back to the Bulldozer era CPU's and re-run benchmarks with all of the Intel mitigations (including disabling hyperthreading) in various desktop/workstation/server workloads against Bulldozer?

Also I wonder what would happen if you ran an 8 core Bulldozer vs a fully mitigated Intel CPU (including disabling hyperthreading) of that era in a modern game that utilizes more than 4 cores?

Anyway this level of SMT that we're talking about seems like it belongs in heavily virtualized environments to better saturate CPU resources - not sure where else this would go?
Click to expand...

I agree that SMT is not always that great a performance gain. I think that mostly comes down to caching systems though. The main thing that holds the standard x86 SMT2 implementations is cache systems. AMD shows a slightly higher gain from SMT then Intel and I have no doubt that is due to the increase in cache space. On the crazy end the IBM SMT8 chips wtih 22 core 88 thread setups do show that if a chip is backed up by a lot of cache those extra threads do operate more as people might expect. IBM has 110 MB of L3 Cache on those chips. The P9 chips SMT threads do perform a lot closer to full cores vs the current x86 stuff.

I don't expect a AMD SMT4 implementation would be mind blowing unless it was backed up by 64+mb of memory. Mainly why I don't believe this rumor... 7nm+ will allow more cache but I am not sure about more then double again from 7 to 7+.
 
Last edited:
M76 said:
People are already emotional that's why they are trying to downplay the leap that Zen3 is in the consumer space.
Click to expand...

But it isn't...?

Enthusiasts, a bit, serious content creators, absolutely, but consumers? Consumers stopped benefiting at four cores. Consumers benefit from lower prices, and at best, higher single-core performance, and where consumers really benefit is lower power usage for longer mobile battery life.

M76 said:
I'm just glad I have no emotional attachment to any company so I can switch at any time without having to explain anything to anyone.
Click to expand...

Lol.

I've been doing this long enough to watch both major CPU companies succeed and fail, to be declared king and to be derided as not producing a product worth buying. I have proudly owned both, and others, back when those choices existed.

AMD has made inroads into the commercial space. That's where they're winning -- and a very large part of that is Intel's stumble at 10nm. Zen 2 isn't faster, but it is cheaper, and if AMD can keep up production, they'll be able to stay back in the game.

On the desktop space for enthusiasts, they're basically all I recommend outside of corner cases.

For everyone else on the desktop, it doesn't matter.

For mobile users, AMD is still in the same 'stay away' territory that the mobile Pentium IV was.

Overall, they're doing alright, and moving to > SMT2 looks like it might be interesting.
 
IdiotInCharge said:
But it isn't...?

Enthusiasts, a bit, serious content creators, absolutely, but consumers? Consumers stopped benefiting at four cores. Consumers benefit from lower prices, and at best, higher single-core performance, and where consumers really benefit is lower power usage for longer mobile battery life.
Click to expand...
Aren't enthusiasts consumers too? I thought everyone who is not a professional / enterprise client is a consumer/prosumer. The people who you refer too are the "normies" they didn't benefit from anything for the past ten years apart from as you say lower power, so they are kind of not the target audience for Ryzen 3000.


Lol.
Click to expand...
That's an emotional response.
I've been doing this long enough to watch both major CPU companies succeed and fail, to be declared king and to be derided as not producing a product worth buying. I have proudly owned both, and others, back when those choices existed.
Click to expand...
What is your point? I went from Intel to AMD, and back countless times.

AMD has made inroads into the commercial space. That's where they're winning -- and a very large part of that is Intel's stumble at 10nm. Zen 2 isn't faster, but it is cheaper, and if AMD can keep up production, they'll be able to stay back in the game.
Click to expand...
Not faster in games, faster in productivity and most content creation. I'd say that's a giant leap, a leap I was waiting for for the longest time, since the first 6 core cpus were released at wholly unattainable prices. Don't tell me, that making 8-12 core affordable is not a leap forward.


On the desktop space for enthusiasts, they're basically all I recommend outside of corner cases.

For everyone else on the desktop, it doesn't matter.
Click to expand...
For everyone else it stopped mattering a long time ago, as I've mentioned.
For mobile users, AMD is still in the same 'stay away' territory that the mobile Pentium IV was.
Click to expand...
I've made no comment for mobile. But I've had an AMD laptop and you get what you pay for.
 
M76 said:
Aren't enthusiasts consumers too?
Click to expand...

Yes, but you wouldn't confuse a shopper for a Camry with a shopper for a Charger Hellcat, despite both being four-door family sedans. One needs a car that does car things, the other has more specific requirements.

M76 said:
What is your point? I went from Intel to AMD, and back countless times.
Click to expand...

Same. And side-by-side.

M76 said:
Don't tell me, that making 8-12 core affordable is not a leap forward.
Click to expand...

More affordable, sure. If you needed the cores, they've been available on HEDT for quite some time. Shrinking that to a consumer socket is cool, but it's also a very small niche.

M76 said:
Not faster in games, faster in productivity and most content creation.
Click to expand...

But here's the thing: faster in games matters, because games are real-time input / output applications. Video rendering? So the kid's socker video takes a few more minutes? Or a few less?

For serious content creation there's a leap, but again, that leap isn't that great when the same performance has been available with HEDT for years -- and if you're spending thousands on a content-creation rig, the cost difference to get an HEDT has been around <10%.

M76 said:
since the first 6 core cpus were released at wholly unattainable prices.
Click to expand...

...they really weren't. They did come on HEDT platforms though, so there's that incremental platform cost, but when Intel released their six-core consumer CPU, it was priced in line with previous CPUs.


I'm not knocking Zen for what it is -- I recommend it! -- but I'm also not confused about what it isn't. It's a leap for serious content creators and the few others that can use that many disaggregated cores, but it's not 'faster' for most consumers or even most enthusiasts, and it's also not cheaper for the top end stuff. Zen's value lies in the middle, for consumers and enthusiasts.

M76 said:
I've made no comment for mobile. But I've had an AMD laptop and you get what you pay for.
Click to expand...

It's the largest consumer PC market, which is why I bring it up. Something like the XPS13 from Dell with equivalents from HP, Lenovo etc. are really the pinnacle of desktop computing. Beyond that you're adding hardware to do some form of 'work', be it gaming, content creation, development, etc.
 
The above comment does a great deal of understating large differences when it's in favor of ryzen while making a big deal out of things in a concerted effort to minimalize or marginalize the benefits ... Odd. Someone who was actually neutral on the issue wouldn't go out of their way to do that.

The only way the above is true is if money is no object in the server space and if you think "most consumers and most enthusiasts" are purchasing cpu's that cost > 500 bucks. Because that's how much you'd have to be spending for anything you said to make any sense when it comes to faster cpu's and zen2. Most consumers and most enthusiasts live in the budget that AMD markets to, and you'd have to be an idiot to think that intel has a viable answer to zen2 at this time.

If you're down with intel, that's fine. But you'll have to wait until they move to 10/7nm for an intel purchase to be based around practical reasons and not reasons soley based on subjective opinions.
 
Chiming in a little late, but I did work a lot on SMT when I was a cpu designer (it's been a while, but the concepts are the same).

Increased scaling with increased threads / SMT does not mean that SMT is a better implementation per se. We have to remember what SMT is: a tool to increase procunit (ALU and friends) utilization.

When you add more top end threads, and it increases performance, what does that mean? Well, the simplest explanation is that the cores were not being fully utilized before. That can be due to a large number of factors, but that core (pun) point remains true. There were idle resources, else you would still not increase throughput with more threads on the top end.

I have not studied current architectures to a significant degree to indicate what the real issues are for both architectures. It simply reflects that Intel's default thread to core ratio is more balanced in terms of execution needs. Intel may be much faster on the top end. Zen may be much faster on the bottom (and thus starved more).

It's a balancing act. If you get no increase in performance from SMT, that means your procunits are always fully saturated, and perhaps you should add more to handle per-thread ILP. On the other hand, you don't want to go nuts with procunits which are largely idle, and thus require a myriad threads feeding it to be competitive.

As always, please benchmark with the things which most closely approximate (or actually are) your expected workload.
 
Darth Ender said:
The above comment does a great deal of understating large differences when it's in favor of ryzen while making a big deal out of things in a concerted effort to minimalize or marginalize the benefits ... Odd.
Click to expand...

It's 'odd' to consider a product in relation to its uses?

Darth Ender said:
Someone who was actually neutral on the issue wouldn't go out of their way to do that.
Click to expand...

Someone who was 'actually neutral' wouldn't worry about what others thought and would just speak the truth.

Darth Ender said:
The only way the above is true is if money is no object in the server space and if you think "most consumers and most enthusiasts" are purchasing cpu's that cost > 500 bucks. Because that's how much you'd have to be spending for anything you said to make any sense when it comes to faster cpu's and zen2. Most consumers and most enthusiasts live in the budget that AMD markets to, and you'd have to be an idiot to think that intel has a viable answer to zen2 at this time.
Click to expand...

Please reread the discussion. There is significant historical reference that you're ignoring in this paragraph.

Darth Ender said:
If you're down with intel, that's fine.
Click to expand...

I agree? I'm down with performance, not with Intel or AMD or Nvidia or really anyone else. You can be down with Intel if you like.

Darth Ender said:
But you'll have to wait until they move to 10/7nm for an intel purchase to be based around practical reasons and not reasons soley based on subjective opinions.
Click to expand...

Intel's faster at gaming and that's the only thing that consumers and even enthusiasts do that actually makes use of more speed in a way that matters.

Now, Intel isn't faster enough at gaming for me to recommend them most of the time, so I don't, but that's neither here nor there.
 
Snowdog said:
That looks like one seriously cherry picked set of workloads. There are a lot of things were SMT isn't anywhere near that beneficial. Heck there are many games where SMT lowers performance.
Click to expand...

Obviously, considering the testing was being done to determine SMT yield as a point of architectural comparison.
 
PhaseNoise said:
When you add more top end threads, and it increases performance, what does that mean? Well, the simplest explanation is that the cores were not being fully utilized before. That can be due to a large number of factors, but that core (pun) point remains true. There were idle resources, else you would still not increase throughput with more threads on the top end.

I have not studied current architectures to a significant degree to indicate what the real issues are for both architectures. It simply reflects that Intel's default thread to core ratio is more balanced in terms of execution needs. Intel may be much faster on the top end. Zen may be much faster on the bottom (and thus starved more).

It's a balancing act. If you get no increase in performance from SMT, that means your procunits are always fully saturated, and perhaps you should add more to handle per-thread ILP. On the other hand, you don't want to go nuts with procunits which are largely idle, and thus require a myriad threads feeding it to be competitive.
Click to expand...

AMD's procunits have more resources.

What this may be saying is that for server loads they typically are underused for various reasons. As you said, could be due to a large number of factors.

This means is AMD is trying to squeeze more performance out of their cores for server workloads

For the destkop, generally speaking, this is irrelevant.
 
IdiotInCharge said:
and if you're spending thousands on a content-creation rig, the cost difference to get an HEDT has been around <10%.
Click to expand...
That's the most concentrated nonsense I've read in a long time. If the cost difference between a 12 core HEDT and a consumer build was 10% I'd have been all over that shit. Just a year or two ago the 12core intel CPU cost more than an entire consumer build.
But the difference is more than 10% even if you compare a 4 core consumer build to a 6 core HEDT intel build from the past 5-6 years. But the real world performance gains in productivity were minimal.
I've been a prosumer / hobby content creator for 20 years, doing video editing, 3d modeling, rendering, animations and so on. And I'm telling you Ryzen 3000 is a game changer for me, there was nothing affordable for a hobbyist until now. Maybe if I was living in the US, I could afford a $2000 CPU for my hobby, but not here in back ass nowhere.
 
You must log in or register to reply here.
Back
Top