Ryzen 3000 hype and review oddities

Morkai

Limp Gawd
Joined
May 10, 2011
Messages
426
I, like many others, am in the situation that I'm sitting on a slightly aged (but well performing 24/7 overclocked) intel quadcore 6700k ,and considering upgrading.

I haven't owned an AMD cpu since the Athlon 1800+ and ABIT were royalty, and I was cautiously set on getting a 3000-series ryzen this time.

For the very sad reason that my dog was sick and old and needed looking after, I had all the time in the world to sit with her and read all the major reviews in detail, and i noticed some interesting things.



Of course, disclaimer, all this could change with bios upgrades, windows patches etc, but as it looks now;



If you just scratch the surface of reviews, there so so much hype and it'll look like AMD caught up, whipped intel, out-cored them, all this with better power efficiency, price etc etc. The fact that Intel is still slightly ahead in games is acknowledged, but considered unimportant because intel is just 5% ahead or so (according to LTT video average, for example). It might be a bit more than 5%, and certainly more overclocked as the 3000-series hardly overclocks at all.

Many reviews did not include overclocked gaming performance.



This fact alone is pretty big. Many people only do low-intensive tasks or work on their computers, and the only demanding task they do is play games.

5% (+ optional overclock) for the main task while not having to worry about the existing and still not fixed issues with chiplet/cluster latency is pretty solid (LTT vid at this point in time describes inconsistencies probably caused by this ) . This will probably be optimized in windows scheduler and/or in newer games, but it is not today.

The new xbox and playstation will also use this architecture, so it will help prioritize solving the issues but it might not happen fast. (And it can never be completely solved since its a physical design, only software workarounds can be created).



Next, lets look at power efficiency. The pre-release hype was big on this, but lets look at benchmarks.

First of all, many review sites use blender. and use a fixed load to process, such as "rendering the gamers nexus logo" etc.

I noticed that in blender, the ryzen 3000 simply outperformed the intel lineup.

https://www.sweclockers.com/test/27760-amd-ryzen-9-3900x-och-7-3700x-matisse/17#content

Here ryzen3000 outperforms the intel lineup by a fair amount. But for a fixed workload, efficiency is just slightly better:

https://www.sweclockers.com/test/27760-amd-ryzen-9-3900x-och-7-3700x-matisse/27#content

The GN test have it as 87 for the 3700x stock vs 91 for the 9900k stock. between several benchmarks like this, it seems that in blender, a best case ryzen3000 scenario, the power efficiency is equal or less than the figures it also outperforms the intel lineup in.

Best case it looks like about 9% better efficiency-wise. (also, blender is probably a bad benchmark to use as it performs better on gpu or hybrid; https://blog.render.st/blender-2-8-hybrid-cpugpu-rendering-speed-and-quality/ )



So did anyone test real world scenarios for power efficiency? one of the few was https://www.techpowerup.com/review/amd-ryzen-7-3700x/18.html

Here we can still see stress test slightly in ryzen3000's favor, but everything real-world intel is ahead (of course, full load over all cores can be a real world scenario but probably more rare on consumer parts).

To sum it up, ryzen 3000 might have a small efficiency advantage in full load all core workloads, and be slightly behind in everything else. Its very equal overall and seems to just have been hype/marketing.

Considering we are comparing 7nm to 14nm, a case can be made that matisse is either inherently extremely power hungry, or simply doesn't make use of the 7nm process power advantages.



Which brings us to temperatures. It's almost like the review kit asked reviewers to leave this out. Almost no none published temperature results.

For those that did, it looks quite bad, and might be the main reason the ryzen3000-series doesn't overclock almost at all.

Those tightly packed 7nm cores might just be really hard to cool, as power efficiency seems about equal.

These are the only two reviews i saw with temperatures, and the techspot one didn't include an intel reference.

https://www.techpowerup.com/review/amd-ryzen-9-3900x/19.html

https://www.techspot.com/review/1869-amd-ryzen-3900x-ryzen-3700x/



Personal thoughts

I am still strongly considering a 3900x as it is a unique 12-core product for a great price, and the chiplet/cluster issues might very well be optimized in the future.

If i decide against the 3900x i see no reason to buy the rest of the 3000 lineup, and will either go for a 9900k or wait for the next generations and re-evaluate as the 6700k is really still quite strong. The minor difference in price simply does not seem worth gambling on the chiplet/cluster design for.
 
Seems you're trying to convince yourself Intel is the way to go by finding the one review that says what you want it to say and labeling it as "real world" then pretending power consumption is your primary concern when looking at the highest end CPU's available on mainstream platforms. Which is doubly odd since it seems like overclocking is also a priority.

AMD will never make a processor that's good enough to beat priorities with shifting goalposts and cherry picked data points.

There's nothing wrong with buying Intel, so like the person above said, if you want Intel, buy Intel.
 
Personally, I'd wait. I mean the 6700k isn't really all that old (well, IMHO).

I'm not sure there's a "revolution" coming to "out do" what AMD has put up with Zen 2, but, you never know.

Let's just say, I'd be pretty happy to have a 6700K.

But... if you have a lot of money burning a hole in your pocket....I guess why not? You know? (I'm not going to tell my car enthusiast friend to not buy "whatever" if it's something he really wants)
 
Last edited:
Seems you're trying to convince yourself Intel is the way to go by finding the one review that says what you want it to say and labeling it as "real world" then pretending power consumption is your primary concern when looking at the highest end CPU's available on mainstream platforms. Which is doubly odd since it seems like overclocking is also a priority.

AMD will never make a processor that's good enough to beat priorities with shifting goalposts and cherry picked data points.

There's nothing wrong with buying Intel, so like the person above said, if you want Intel, buy Intel.

I think you need to re-read it. Power efficiency is of zero concern to me personally, just pointing out that people all over seem to repeat what seems to be untrue hype about efficiency (and that the opposite is possibly true). And I read and listened to and considered almost all major reviews, and pointed out that pretty much only the one i linked even bothered testing more normal use power efficiency, or temperatures - look around for yourself. I did not link any "cherry picked" reviews. In fact, the sweclockers power efficiency test is one of the best of all for ryzen 3000 with amd 3700x 9% or so ahead of 9900k in efficiency. (though blender is probably a bad test as described above, and also heavily favors ryzen3000's performance).

I am trying to convince myself to buy a 3900x :) I consider both the intel 9000-series and the 3900x great products, while I think that the rest of the 3000-lineup is a bad gamble for gaming on the chiplet/cluster design when you can simply have the same performance without potential issues from intel, and that the price difference in the lower/mid segment is minor. In the future, maybe all cpus will follow this design, but until then and until it has been properly optimized, it looks like a a dubious gamble.
 
I think you need to re-read it. Power efficiency is of zero concern to me personally, just pointing out that people all over seem to repeat what seems to be untrue hype about efficiency (and that the opposite is possibly true). And I read and listened to and considered almost all major reviews, and pointed out that pretty much only the one i linked even bothered testing more normal use power efficiency, or temperatures - look around for yourself. I did not link any "cherry picked" reviews. In fact, the sweclockers power efficiency test is one of the best of all for ryzen 3000 with amd 3700x 9% or so ahead of 9900k in efficiency. (though blender is probably a bad test as described above, and also heavily favors ryzen3000's performance).

I am trying to convince myself to buy a 3900x :) I consider both the intel 9000-series and the 3900x great products, while I think that the rest of the 3000-lineup is a bad gamble for gaming on the chiplet/cluster design when you can simply have the same performance without potential issues from intel, and that the price difference in the lower/mid segment is minor. In the future, maybe all cpus will follow this design, but until then and until it has been properly optimized, it looks like a a dubious gamble.

It's a very simple decision to make. You need to ask yourself this question. Is my system mainly for gaming? If your answer is yes, you should go with 9900K.
 
I think you need to re-read it. Power efficiency is of zero concern to me personally, just pointing out that people all over seem to repeat what seems to be untrue hype about efficiency (and that the opposite is possibly true).

The topic of efficiency is a complex one to discuss, and something that should be kept in line with specific use cases, to include software, hardware, environment, and performance need.

For systems that are largely closely matched, 'more efficient' or 'less efficient' can vary considerably between test subjects, as can whether the relative differences and totals are even impactful for a decision.
 
It's a very simple decision to make. You need to ask yourself this question. Is my system mainly for gaming? If your answer is yes, you should go with 9900K.

Ahh now this is not really 100% accurate. I game at 4k and I went with last gen Ryzen 2600X and this gen Ryzen 3600X because I could first off and I wanted to also. Well you can buy 2 or 3 Ryzen CPU's and have the same gaming experience as 9900K at 4K.
 
Sounds like you are looking for things to not like. As has been said its ok to still buy Intel if you have strong brand loyalty or if you even for some reason feel their product is still the best bang for your buck or worth a little bit of a premium whatever.

But let's get real on power... people are impressed because for instance a 2700x vs a 3700x you will be looking at at a 32% absolute power savings as per Anandtech (and others have found the same). As they say;
"We have to remember that we’re talking about overall absolute power, and not efficiency of the chip. When taking actual performance into account through the higher clock as well as Zen2’s increased performance per clock, the Performance/W figures for the new 3700X should be significantly higher than its predecessor."

With regard to the 3900x Anandtech said this;
"Following that, we see that this CPU’s per-core peak power consumption is quite notably higher than that of the 3700X, which is not a surprise given that the chip is clocked 200MHz higher at 4.6GHz versus “just” 4.4GHz. However even at this much higher clock, the 3900X’s power consumption remains notably lower than that of the 2700X."

https://images.anandtech.com/graphs/graph14605/111362.png

I mean I'm not sure what you are expecting but it looks to me like the 3700x uses about half as much power as a 9900k. The 3900x draws around 16% less then the 9900k sells for around the same price and has 4 more cores.

The i9-7920x does offer right around the same power use it seems as a 3900x which is impressive. Of course the 3900x demolishes it in a lot of tests... and the i9 equals it at best while never breeting 3900. Oh and 7920x is a $1200 CPU.
 
while I think that the rest of the 3000-lineup is a bad gamble for gaming on the chiplet/cluster design when you can simply have the same performance without potential issues from intel, and that the price difference in the lower/mid segment is minor. In the future, maybe all cpus will follow this design, but until then and until it has been properly optimized, it looks like a a dubious gamble.

Addressing this separately- I agree that this hasn't been thoroughly tested, and that it is a point of concern. The moment a latency-sensitive multithreaded application is pushed beyond the core complex boundary, and then the die complex boundary, dips in performance should be noticeable if not catastrophic. There is no 'absolute' solution to this problem as the latency involved exists by design. Generally, operating systems are being updated to be aware of core distribution and the latency involved, but applications must also become aware, and in neither case can the latency be mitigated to the point that it would be in a monolithic CPU with the same number of cores.
 
Ahh now this is not really 100% accurate. I game at 4k and I went with last gen Ryzen 2600X and this gen Ryzen 3600X because I could first off and I wanted to also. Well you can buy 2 or 3 Ryzen CPU's and have the same gaming experience as 9900K at 4K.

Not accurate if you're trying to push today's AAA games at the highest settings with today's fastest GPUs. Change any of those variables for the better, and you have a different picture.
 
I think you need to re-read it. Power efficiency is of zero concern to me personally, just pointing out that people all over seem to repeat what seems to be untrue hype about efficiency (and that the opposite is possibly true). And I read and listened to and considered almost all major reviews, and pointed out that pretty much only the one i linked even bothered testing more normal use power efficiency, or temperatures - look around for yourself. I did not link any "cherry picked" reviews. In fact, the sweclockers power efficiency test is one of the best of all for ryzen 3000 with amd 3700x 9% or so ahead of 9900k in efficiency. (though blender is probably a bad test as described above, and also heavily favors ryzen3000's performance).

I am trying to convince myself to buy a 3900x :) I consider both the intel 9000-series and the 3900x great products, while I think that the rest of the 3000-lineup is a bad gamble for gaming on the chiplet/cluster design when you can simply have the same performance without potential issues from intel, and that the price difference in the lower/mid segment is minor. In the future, maybe all cpus will follow this design, but until then and until it has been properly optimized, it looks like a a dubious gamble.

How is a chiplet design an issue again ? It's a proven tech that AMD has been using in the server space for awhile. (and so has Intel)
Chiplet design is the future. AMD simply beat Intel there at the consumer level. They beat Intel to an operating next gen fab, and design structure.

https://www.anandtech.com/show/14211/intels-interconnected-future-chipslets-emib-foveros

Nvidia also has chiplet based products already.
https://www.tomshardware.com/news/nvidia-msm-inference-chip,39780.html

We don't know specifics in regards to Intels XE GPUs or NVs next 7nm... but rumors are they will both be chiplet designs. I would also expect if zen2 continues to sell at the pace it has been... we will see intel move some of those server space design ideas into the mainstream a lot faster then anyone expects.
 
You have to qualify gaming. At 720p Intel 1080p it's really close 1440p and up don't matter.

Doesn't everyone game for serious at 720p ? ;)

Ya the higher you crank the resolution... the more stuff the game is doing the more intels small (and it is even if it looks big on a graph) advantage evaporates. Without a doubt no one boosting either system would see a difference in a blind A B test... cause the only differences anyone can show are at low resolutions with high end GPUs pushing 100s of frames per second. No one is seeing differences in frame rates when they are well over even a 144hz refresh cap.
 
Sounds like you are looking for things to not like. As has been said its ok to still buy Intel if you have strong brand loyalty or if you even for some reason feel their product is still the best bang for your buck or worth a little bit of a premium whatever.

But let's get real on power... people are impressed because for instance a 2700x vs a 3700x you will be looking at at a 32% absolute power savings as per Anandtech (and others have found the same). As they say;
"We have to remember that we’re talking about overall absolute power, and not efficiency of the chip. When taking actual performance into account through the higher clock as well as Zen2’s increased performance per clock, the Performance/W figures for the new 3700X should be significantly higher than its predecessor."

With regard to the 3900x Anandtech said this;
"Following that, we see that this CPU’s per-core peak power consumption is quite notably higher than that of the 3700X, which is not a surprise given that the chip is clocked 200MHz higher at 4.6GHz versus “just” 4.4GHz. However even at this much higher clock, the 3900X’s power consumption remains notably lower than that of the 2700X."

https://images.anandtech.com/graphs/graph14605/111362.png

I mean I'm not sure what you are expecting but it looks to me like the 3700x uses about half as much power as a 9900k. The 3900x draws around 16% less then the 9900k sells for around the same price and has 4 more cores.

The i9-7920x does offer right around the same power use it seems as a 3900x which is impressive. Of course the 3900x demolishes it in a lot of tests... and the i9 equals it at best while never breeting 3900. Oh and 7920x is a $1200 CPU.

That result from anandtech is interesting and not at all in line with the others. Anandtech is one I did not read very carefully as it was quickly pointed out that they used the wrong bios at first, but they have remedied that now.

However, the others measure power draw at wall, and anandtech measures from cpu package meaning the x570 motherboard is omitted.
If their results are correct, its probably as simple as ryzen 3000 uses less power, but x570 (that needs its own fan because of high power use), adds enough to reach intel levels - which is an interesting find.

Maybe ryzen3000 on an older motherboard would actually achieve the claimed efficiency increase.
 
That result from anandtech is interesting and not at all in line with the others. Anandtech is one I did not read very carefully as it was quickly pointed out that they used the wrong bios at first, but they have remedied that now.

However, the others measure power draw at wall, and anandtech measures from cpu package meaning the x570 motherboard is omitted.
If their results are correct, its probably as simple as ryzen 3000 uses less power, but x570 (that needs its own fan because of high power use), adds enough to reach intel levels - which is an interesting find.

Maybe ryzen3000 on an older motherboard would actually achieve the claimed efficiency increase.

It's not really that interesting. PCIx 4 doesn't power itself. I think it should be a pretty safe assumption that yes x570 uses more power then older PCI 3 based boards from either company.

I agree that it would be interesting to see X470 vs X570 3000 power comparisons. Still bottom line from everything I have read the 3000 ryens are much more power efficent vs the 2000s. They are also besting Intel a bit... and worse case equaling them. The problem is its very hard to really find fair comparisons imo. Intel right now is either selling chips with fewer cores and drastically smaller cache systems.... or people are comparing consumer level ryzen chips to server class Intel chips selling for 2x the price. (and AMD is equaling or winning... hench a lot of the hype). ;)
 
How is a chiplet design an issue again ? It's a proven tech that AMD has been using in the server space for awhile. (and so has Intel)
Chiplet design is the future. AMD simply beat Intel there at the consumer level. They beat Intel to an operating next gen fab, and design structure.

https://www.anandtech.com/show/14211/intels-interconnected-future-chipslets-emib-foveros

Nvidia also has chiplet based products already.
https://www.tomshardware.com/news/nvidia-msm-inference-chip,39780.html

We don't know specifics in regards to Intels XE GPUs or NVs next 7nm... but rumors are they will both be chiplet designs. I would also expect if zen2 continues to sell at the pace it has been... we will see intel move some of those server space design ideas into the mainstream a lot faster then anyone expects.

It adds latency between threads that need to be in sync, which applies to cpu for games, but not as often/never other products like servers, gpus etc. As long as only 4 physical cores+4smt are needed, it can be sorted. Today the windows scheduler doesn't seem to "know/care" about the chiplet design and randomly runs things across chiplets/clusters.
If you actually read what i wrote and click the links, the LTT video has an explanation/speculation at the point in the i linked the LTT video. They also made a later video expaning on the subject.
[edit: this sounds somewhat agressive but is unintended.]

The chiplet/cluster design is a bad thing for gaming consumers until it has been optimized for. we do not want it (yet). its is probably the future though, but for at least this generation, probably best to avoid. Not avoiding it is a gamble that windows and/or game devs will quickly optimize for it.
 
It adds latency between threads that need to be in sync, which applies to cpu for games, but not as often/never other products like servers, gpus etc. As long as only 4 physical cores+4smt are needed, it can be sorted. Today the windows scheduler doesn't seem to "know/care" about the chiplet design and randomly runs things across chiplets/clusters.
If you actually read what i wrote and click the links, the LTT video has an explanation/speculation at the point in the i linked the LTT video. They also made a later video expaning on the subject.
[edit: this sounds somewhat agressive but is unintended.]

The chiplet/cluster design is a bad thing for gaming consumers until it has been optimized for. we do not want it (yet). its is probably the future though, but for at least this generation, probably best to avoid. Not avoiding it is a gamble that windows and/or game devs will quickly optimize for it.

Theory crap. We don't want it ? WTF are you talking about. The only issue with clusters of cores is they can't share level 1 and level 2 cache... which anyone that can remember back to celeron proto days can attest effects performance. Having said that.... did you not notice the massive 64mb L3 Cache on the 3900x. Yes the solution to the "issue" you are describing is to increase memory cache usable by each core and cluster. I hate to break it to you but intel also uses core clustering... they are simply on the same die. Intel still has their own version of infinity fabric tying core clusters together on 4+ core chips.

Read the reviews. Looks at the results. AMDs design is NOT loosing anything. Their chiplet design is handedly besting their older 2000 series. If a chiplet design was an issue they wouldn't be easily beating up on their own previous generation running much the same arch on a traditional package.

Optimization isn't an issue because that is sort of the point of a chiplet. The chip has a controller on board that routes data where it needs to be. A chiplet isn't in anyway different from older non chiplet designs. The only difference is the controller is on its own bit of silicon... mainly to improve chip yields. (thus reducing cost and allowing the connecting of lots of inexpensive silicon to make a more capable package). Inside that intel chip things work exactly the same way... a 12 core intel chip has the same issues even if it is all on one bit of silicon. Those threads still need to use a L3 cache because they can't always share the L1 and L2 depending on the layout. Issues with multiple cores are not new to chiplets. The solutions are still the same.... AMD has cranked the simplest easy fixes for high core issues to 11. Namely speeding up infinity fabric... decoupling it from ram speeds... and massively increasing every level of chip cache to reduce the need for the controller to split threads.
 
How is a chiplet design an issue again ? It's a proven tech that AMD has been using in the server space for awhile. (and so has Intel)
Chiplet design is the future. AMD simply beat Intel there at the consumer level. They beat Intel to an operating next gen fab, and design structure.

https://www.anandtech.com/show/14211/intels-interconnected-future-chipslets-emib-foveros

Nvidia also has chiplet based products already.
https://www.tomshardware.com/news/nvidia-msm-inference-chip,39780.html

We don't know specifics in regards to Intels XE GPUs or NVs next 7nm... but rumors are they will both be chiplet designs. I would also expect if zen2 continues to sell at the pace it has been... we will see intel move some of those server space design ideas into the mainstream a lot faster then anyone expects.

The chiplet design is a cost compromise a monolithic die is indeed better for pure performance. AMD has done really well to advance the design with Zen 2 but it’s still a compromise. Being proven in the server space is one thing. Another thing that’s proven is that in latency sensitive workloads, there is a performance hit. It is, essentially, a bottleneck, but PCs have many bottlenecks and which one you happen to run into is going to depend on many factors including workloads. If a game scales past 6 cores or isn’t intelligent enough to keep the processing on one chiplet when it doesn’t need more than 6 cores you will likely come across this bottleneck.
 
Theory crap. We don't want it ? WTF are you talking about. The only issue with clusters of cores is they can't share level 1 and level 2 cache... which anyone that can remember back to celeron proto days can attest effects performance. Having said that.... did you not notice the massive 64mb L3 Cache on the 3900x. Yes the solution to the "issue" you are describing is to increase memory cache usable by each core and cluster. I hate to break it to you but intel also uses core clustering... they are simply on the same die. Intel still has their own version of infinity fabric tying core clusters together on 4+ core chips.

Read the reviews. Looks at the results. AMDs design is NOT loosing anything. Their chiplet design is handedly besting their older 2000 series. If a chiplet design was an issue they wouldn't be easily beating up on their own previous generation running much the same arch on a traditional package.

Optimization isn't an issue because that is sort of the point of a chiplet. The chip has a controller on board that routes data where it needs to be. A chiplet isn't in anyway different from older non chiplet designs. The only difference is the controller is on its own bit of silicon... mainly to improve chip yields. (thus reducing cost and allowing the connecting of lots of inexpensive silicon to make a more capable package). Inside that intel chip things work exactly the same way... a 12 core intel chip has the same issues even if it is all on one bit of silicon. Those threads still need to use a L3 cache because they can't always share the L1 and L2 depending on the layout. Issues with multiple cores are not new to chiplets. The solutions are still the same.... AMD has cranked the simplest easy fixes for high core issues to 11. Namely speeding up infinity fabric... decoupling it from ram speeds... and massively increasing every level of chip cache to reduce the need for the controller to split threads.

The increased L3 cache is a workaround for the increased latency, but it can not be solved as it is a physical design.
Again, click the LTT video which is at the correct point in time and listen for about 20 seconds. They had bad results and mitigated the issue by manually assigning cores. Users who do not, will lose sometimes lose performance until or rather IF windows scheduler and/or games in the future are optimized for this.

"The chip has a controller on board that routes data where it needs to be" - No. windows decides what cores runs what (and again, doesn't know/care about the ryzn3000 layout, at least yet, again, see the video - it's their info, not mine) , or an affinity mask at application startup/task manager overrides.
The chip does not analyze the application name, and tell windows scheduler "hey put it here please".
 
The chiplet/cluster design is a bad thing for gaming consumers until it has been optimized for. we do not want it (yet). its is probably the future though, but for at least this generation, probably best to avoid. Not avoiding it is a gamble that windows and/or game devs will quickly optimize for it.
Do you have anything to back this up? From what I've seen the new scheduler in the latest Win 10 version doesn't have much of an impact for Ryzen 3.
 
Do you have anything to back this up? From what I've seen the new scheduler in the latest Win 10 version doesn't have much of an impact for Ryzen 3.

Hmmmm, with a strange sense of deja vu, "click the LTT video above which is at the correct point in time" :) If they have patched the issue in the past days it is beyond my knowledge. Last windows patch was july 9?
 
The chiplet design is a cost compromise a monolithic die is indeed better for pure performance. AMD has done really well to advance the design with Zen 2 but it’s still a compromise. Being proven in the server space is one thing. Another thing that’s proven is that in latency sensitive workloads, there is a performance hit. It is, essentially, a bottleneck, but PCs have many bottlenecks and which one you happen to run into is going to depend on many factors including workloads. If a game scales past 6 cores or isn’t intelligent enough to keep the processing on one chiplet when it doesn’t need more than 6 cores you will likely come across this bottleneck.

It's not a new compromise is my point. Monolithic silicon CPUs have the same issues... Cores are clustered. That isn't new. The only difference is intel is using their off die interconnect instead of on die solutions. Intel or AMD older monolithic chips still have latency added when things are spread over multiple die complexes. However AMD has improved this behaviour in 3000 by drastically increasing the amount of shared cache. The entire point of a L3 cache is to allow multiple core complexes to share data. That isn't some new thing.

Again AMD is simply first to the party. Intel has started talking the last few weeks about their EMIB tech... as they are going to need their own infinity fabric like tech to enable their chiplet designs. Intel is talking about EMIB in relation to their current chiplet designs. The only difference between intel and amd right now... is amd has working tech to sell, and all intel has been doing the last few weeks is talk about tech they WILL be using. The tech is basically the same... although of course Intel is spinning theirs as superior. lol ;)

For games ya multiple cores are crap on all processors there is a reason why you get a FPS uplift when you disable hyperthreading.
 
  • Like
Reactions: blkt
like this
Chiplets get the components closer together but are you seriously trying to suggest that NUMA doesn't exist in the AMD sphere and call it theorycrafting?
 
The increased L3 cache is a workaround for the increased latency, but it can not be solved as it is a physical design.
Again, click the LTT video which is at the correct point in time and listen for about 20 seconds. They had bad results and mitigated the issue by manually assigning cores. Users who do not, will lose sometimes lose performance until or rather IF windows scheduler and/or games in the future are optimized for this.

"The chip has a controller on board that routes data where it needs to be" - No. windows decides what cores runs what (and again, doesn't know/care about the ryzn3000 layout, at least yet, again, see the video - it's their info, not mine) , or an affinity mask at application startup/task manager overrides.
The chip does not analyze the application name, and tell windows scheduler "hey put it here please".

My point is INTEL is not IMMUNE. The exact same issues effect many core intel chips. The fact that it is all on one piece of silicon is irrelevant. If core 0 wants to talk to core 7... it still will not have access to L1 or L2 cache space, and will have to store results of split math in L3. (that is the point of L3) If Core 0 and core 1 are doing work they can share L1 and L2.

Chiplet design hasn't changed core clustering and complexing. Those are old solutions that both intel and amd use. The conventions are different but the ideas are the same... and yes windows scheduler was never optimized well for any of the ryzens, although real world it makes very little difference. The real difference is seen in truely large work loads and on the thread ripper chips. All the hubbub about the windows scheduler a few months back was mostly smoke... yes some very big workloads on the larger thread rippers could see some real benchmark degradation but that was over hyped.
 
Chiplets get the components closer together but are you seriously trying to suggest that NUMA doesn't exist in the AMD sphere and call it theorycrafting?

No I'm saying show me one real world workload where you can say look... this effects AMD and AMD only.

The few games people point to that run better on reduced thread count... behave the same way with Intel hardware. Yes there are some games where if you disable SMT you will see a performance bump.... but the same theory applies to Intel. Turn off hyper threading and the same bump is seen. Issues with splitting work on high thread count chips... isn't new to chiplets. Its true of older monolithic designs as well.
 
Hmmmm, with a strange sense of deja vu, "click the LTT video above which is at the correct point in time" :) If they have patched the issue in the past days it is beyond my knowledge. Last windows patch was july 9?
Sorry, missed that. Haven't seen other reports so I'm not sure if this is repeatable. Were they testing with a proper BIOS?
 
My point is INTEL is not IMMUNE. The exact same issues effect many core intel chips. The fact that it is all on one piece of silicon is irrelevant. If core 0 wants to talk to core 7... it still will not have access to L1 or L2 cache space, and will have to store results of split math in L3. (that is the point of L3) If Core 0 and core 1 are doing work they can share L1 and L2.

This is true but with Ryzen if the other core is on another chiplet it has to traverse infinity fabric which is where the latency happens.
 
My point is INTEL is not IMMUNE. The exact same issues effect many core intel chips. The fact that it is all on one piece of silicon is irrelevant. If core 0 wants to talk to core 7... it still will not have access to L1 or L2 cache space, and will have to store results of split math in L3. (that is the point of L3) If Core 0 and core 1 are doing work they can share L1 and L2.

Chiplet design hasn't changed core clustering and complexing. Those are old solutions that both intel and amd use. The conventions are different but the ideas are the same... and yes windows scheduler was never optimized well for any of the ryzens, although real world it makes very little difference. The real difference is seen in truely large work loads and on the thread ripper chips. All the hubbub about the windows scheduler a few months back was mostly smoke... yes some very big workloads on the larger thread rippers could see some real benchmark degradation but that was over hyped.
With the risk of sounding like a broken record, listen to a few seconds of the LTT video on this. Its not "few months back" but now. Again, I have no deeper knowledge than this and truely no personal investment in neither amd nor intel. I have used products from both and both have reached amazing achivements in different points in time.

What i mean is simply that for intel, who obviously make electronics too and not magic, any such issues and well known and long optimized for. Ryzen with a miniscule (but hopefully growing) market share is not. As suggested by the video, and a followup video speculating in future windows optimizations.
 
Sorry, missed that. Haven't seen other reports so I'm not sure if this is repeatable. Were they testing with a proper BIOS?
To the best of my knowledge, there is no way the bios can affect this. The bios had some minor impact on clock frequencies etc.
 
So tired of hearing from Intel fanboys. Even if AMD had 25% more gaming performance than Intel's top cpu, these dumbasses would still find something to complain about.

I thought this was [H] a hardware enthusiast site that appreciated new PC technology. Way too many fanboys, (including Nvidia), here.
 
HT and SMT are not really related to chiplet design. HT and SMT are hardware features that allow you to use the same core (for lack of a better word) to process a second thread 'behind the scenes'. Unfortunately, HT and SMT can impact performance because HT and SMT are not good for doing the exact same task on the same core. Disabling HT and SMT prevents the OS scheduler from 'stealing' resources from the same core whilst divvying up the workload. It's not like you have part 1 of the core doing one thing and having to run out to the core on another chiplet to get part 2 to execute.

Schedulers, however, can have workloads spread out on other cores and some of those cores may or may not be in the same physical package. On my 12c/24t video rendering machine, I have six cores and twelve threads each in two separate packages. Obviously, this is a dual socket motherboard. Obviously, NUMA applies like a bitch when the scheduler assigns processes to the two separate packages. Turning off HT doesn't change NUMA.

AMD's chiplets are my computer in two chiplets instead of sockets. It's the same thing writ smaller. In Zen2, if I got my facts correct, allows you to have 8 cores in one chiplet. Anytime you can keep your workload on one chiplet (regardless of whether or not SMT is enabled) is going to be faster. If your workload is spread over two 'packages' it's going to impact speed. This is basic physics.
 
This is true but with Ryzen if the other core is on another chiplet it has to traverse infinity fabric which is where the latency happens.

Not really, Just because something is on the same bit of silicon that doesn't in fact make it always faster. Intel used to use a ring bus interconnect this works well on lower core chips... but the more cores you get the differing latency times you get per core. Which is why on older pre skylake intel chips cores where complexed. A complexed core with a ring bus is no better and perhaps worse on paper vs core connected via infinity fabric. (I think that argument goes either way depending on if your talking to AMD or Intel) Intel used a mesh system for knights landing very much like infinity fabric. Since skylake Intel uses their MoDe-X interconnect and I believe this is a 4x4 interconnect. Meaning 4 cores linked to the other 4 cores. Cores in one complex have the same issues with lower level cache sharing and added latency... as AMDs infinity fabric complex connection.

Anyway being in silicon really doesn't automatically make an interconnect better. Potentially but it really depends on the actual layout. At the end of the day its an issue that effects all current designs... and if you have me a choice of an intel chip with 2 x 4 core complexes all on the same bit of silicon with 1.35MB of L3 cache per core or AMDs 4 core complex design with 5.83MB of L3 per core. I think I know which I would prefer.

All theorycrafting about latancy difference between a silicon traced interconnect and a high speed off die interconnect aside. Having 4.3x the L3 Cache is going to mean using that interconnect A LOT less.

Actual benchmark results seem to bear out that the end performance is darn near identical... and if anything further software optimization favors AMD.
 
  • Like
Reactions: blkt
like this
So tired of hearing from Intel fanboys. Even if AMD had 25% more gaming performance than Intel's top cpu, these dumbasses would still find something to complain about.

We're discussing. Perhaps your bias causes you to perceive discussion as 'complaining', and to apply off-topic 'even if' judgements.

-my first CPU was from Cyrix, second was from AMD
 
Actual benchmark results seem to bear out that the end performance is darn near identical

I believe that this is an important bit of context to keep sight of while discussing inter-die latency with Zen 2. While the latency benchmarks show the expected increases in latency, we're not seeing significant performance drawbacks with current software, drivers, and operating systems, and:

and if anything further software optimization favors AMD.

This is absolutely true, again, for software, drivers, and operating systems going forward.
 
All theorycrafting about latancy difference between a silicon traced interconnect and a high speed off die interconnect aside. Having 4.3x the L3 Cache is going to mean using that interconnect A LOT less.

Which is not in dispute. What is disputable is your assertion that going off-die doesn't impact performance.
 
There are good choices (for performance) on both sides - We have a rare moment where its close performance with either team red or green (or blue) - Since gaming is a concern, find the game(s) you play most and matter most to you, pick whatever is the best performing for that game (again keeping money out of the discussion).

Oddly enough for me I have access to several of the latest systems (AMD and Intel), my fav. system to game with personally isn't even considered a top performer anymore (i9-7900X on asus apex board) - I enjoy reading all the reviews, however at a certain level is subjective to each's own use and if they care about temps and power usage. Example I like reading about temps and power draw, but would plug my PC into 220v if I needed to, lol... and always use some form of a higher performance cooler, so (when it comes to) temps are more of a gauge to me as to how hard it will be to hold the boost clocks or overclock a little.
 
Last edited:
HT and SMT are not really related to chiplet design. HT and SMT are hardware features that allow you to use the same core (for lack of a better word) to process a second thread 'behind the scenes'. Unfortunately, HT and SMT can impact performance because HT and SMT are not good for doing the exact same task on the same core. Disabling HT and SMT prevents the OS scheduler from 'stealing' resources from the same core whilst divvying up the workload. It's not like you have part 1 of the core doing one thing and having to run out to the core on another chiplet to get part 2 to execute.

Schedulers, however, can have workloads spread out on other cores and some of those cores may or may not be in the same physical package. On my 12c/24t video rendering machine, I have six cores and twelve threads each in two separate packages. Obviously, this is a dual socket motherboard. Obviously, NUMA applies like a bitch when the scheduler assigns processes to the two separate packages. Turning off HT doesn't change NUMA.

AMD's chiplets are my computer in two chiplets instead of sockets. It's the same thing writ smaller. In Zen2, if I got my facts correct, allows you to have 8 cores in one chiplet. Anytime you can keep your workload on one chiplet (regardless of whether or not SMT is enabled) is going to be faster. If your workload is spread over two 'packages' it's going to impact speed. This is basic physics.

This is all true... but SMT effects performance in a few ways. With games one other issue... is if the game is relying on the OS scheduler (which ideally it should be able to) SMT/HT will sometimes split work up into much smaller bits. This happens a lot for things like calculation of post effects on things like audio where the math can easily be threaded. What tends to happen is things get split accross complexes by this hyper (no pun intended) multi threading on specific "simple" workloads. Which games tend to have lots of. Chip complexes are a reality of every multi core design. In general both Intel and AMD complex cores into clusters of 4.

So ya your right SMT has other issues at times even if the math is all happening on the same single core. But it can more often cause issues with splitting work across complexes... as AMD and Intel depending on the generation of odd core links and ring type designs. Where the scheduler has no idea that cores 0 2 4 6 are linked instead of perhaps the expected 0 1 2 3. I think that was the main issue with early proto ryzen and windows scheduling specifically if I remember correctly.
 
Which is not in dispute. What is disputable is your assertion that going off-die doesn't impact performance.

Its negated almost completely by a superior cache design yes. Is going off die on paper worse... sure. It can be compensated for in a design that properly feeds those cores math by not forcing clearing of cache as often. Perhaps quadrupling the cache is a brute force fix ... but it's a fix.
 
Back
Top