AMD Ryzen 9 3950X Overclocked To 5 GHz Across All 16 Cores On LN2

Status
Not open for further replies.
IdiotInCharge said:
Not really.

And your claim of not just IPC parity, but of a significant IPC advantage for Ryzen 2 is quite laughable.



I am sure that you think your emotional quips are cute, but they're not. They're trolling.
Click to expand...

That is not my claim, it is what "so called" people reviewing it said, please don't infer their finding as something I claimed. Even if we comb through it and clock vs clock a 3ghz 9900k vs 3ghz 3600 we will probably come to the same deduction that Intel will still hold a slight advantage in single thread performance, granted that it was just over 2 years ago AMD was only competing with its old Phenom 2 parts, that is a long way. Since we are on "you claim" you were the one that made the asinine claim that AMD were delivering haswell performance but no proof of that on a clock vs clock basis has been shown.
 
Dan_D said:
Are you kidding me? Prior to the security issues, Intel still had a substantial lead on AMD in terms of IPC. That means clock for clock its faster. Period. It also clocks higher. The only reason why Ryzen holds its own in some applications is because it had more cores. When Ryzen launched, Intel was still selling us quad core CPU's. AMD often had twice the cores in a given price point and it was close enough in clocks and IPC that it would perform better in some tests. AMD has found success with Ryzen for two reasons. 1.) Price vs. performance. 2.) Due to offering more cores for the same or less money, AMD has found performance advantages in some cases. Intel isn't sacrificing efficiency to get faster. Its architectures have continued to improve on a clock for clock basis even if those improvements are small. Yet, clock speeds are going up. That's due to process maturity.

I've said it before, you have to look at the reasons why a given CPU wins a given benchmark to understand what's going on and to evaluate things in the correct context. AMD wins when and where more cores are a benefit. I do the benchmarks and work with the data all the time. That may certainly change with Zen 2. I don't have one on hand yet and haven't done the testing. Early indications are that AMD is going to close the gap significantly. Its also worth pointing out that AMD has increases Ryzen 3000's clocks a fair amount over the preceding generations of Ryzen CPU's. That's part of closing that gap with Intel.
Click to expand...


I have already addressed the first point if you run a clock vs clock say 3ghz test the 9900K will probably have the 9900K faster in single thread but I think compared to when the test was done by anandtech on release of the Ryzen 1800X back then the top CPU was the 7700K which scored 127 points while the 1800X scored 120 points which was on par with the 4770K single thread score with slightly lower clocks. Maybe on the new release the gap is closed a little more, will just have to wait and see. Of course most of AMD's gaming issues have not been clockspeed or single thread related but rather latency related and I imagine while it may still be an issue it will further be tweaked until it gets to optimal level to match the uarch performance.

the second point was listed above, the all core on the 2700 is around 3.8ghz sustained, if the new uarch can do 4.5ghz sustained that is going to result in massive performance gains given that Zen is already competitive with Intel.
 
IdiotInCharge said:
You didn't list any currency while posting on a US forum, so yes, you did say US dollars by omission. If you mean some other currency, of which there are hundreds in the world, then specify that. We're not going to guess, we're just going to laugh.
Click to expand...

laughing is already something we do enough of in this country to get by. It is a country that promised so much but pushes for socialism and other social justice politicking has set it back. In 2011 we traded around 6:1 with the USD, by 2013 and the first leaks of state capture and corruption that plummeted and in 2019 we are at 14:1 to the USD. Despite the problems the country has it remains a hotspot for foreigners, with american tourists being prominent.

I guess anyone can laugh at another persons country for whatever reason, but that doesn't always mean we live any worse than you do, I earn good money, I have beach side property, I have a retierment annuity and investments, I have a fantastic gun collection and despite living here I got a shot at professional baseball. I have thought about moving to the United States but immigration is a lot harder these days.
 
OrangeKhrush said:
That is not my claim, it is what "so called" people reviewing it said, please don't infer their finding as something I claimed. Even if we comb through it and clock vs clock a 3ghz 9900k vs 3ghz 3600 we will probably come to the same deduction that Intel will still hold a slight advantage in single thread performance, granted that it was just over 2 years ago AMD was only competing with its old Phenom 2 parts, that is a long way. Since we are on "you claim" you were the one that made the asinine claim that AMD were delivering haswell performance but no proof of that on a clock vs clock basis has been shown.
Click to expand...

Have you not had internet access for the past couple of years? I've done the board reviews during that time. I've looked at the data, graphed it out and I've had all these CPU's in hand save for the Ryzen 3000 series. When Ryzen launched originally, it delivered performance that roughly traded blows with Haswell. That's not a claim, that's fact. Go look at the reviews from that time frame. There were plenty of clock for clock comparisons done by many review outlets. Intel was still had generations of small improvements over Haswell. You had Haswell refresh, Broadwell, Skylake, and Kaby Lake, Even pulling out the stuff that wasn't an architectural improvement, you still had at least Broadwell and Skylake. Each brought an increase over Haswell. All the processor releases brought something to the table. Generally, clock speeds. Keep in mind that Intel still maintained a substantial clock speed advantage over AMD at this time and still does.

Even with the performance loss in some cases with all the security patches that hurt Intel's performance, it still has an IPC lead in some cases, if not most cases against Ryzen 1000 and 2000 series parts. I was running a Haswell-E based system using a 8c/16t Core i7 [email protected]. I switched to a 12c/24t Threadripper [email protected]. It was pretty much a lateral move outside of cases where an application can use more than 8c/16t. In those cases, Threadripper is obviously faster. I had the 5960X for years and I've been running the Threadripper since April. I can tell you it wasn't much of an upgrade.
 
Just read this thread today.... jesus... anyway

I wonder if we took the higher core count cpu, and manually disabled every other core, how much heat would get saved and as a result, how much better would the enabled cores overclock? Has anyone ever tried that? So the dual chiplet AMD, 16 core, disable 4 cores in each chiplet.

Be a fun experiment.

Compare the overclocking/PB/etc, as well as a few real world gaming tests, then re-anable the cores and do the same tests.

Regarding the undervolting, seems that I read here on [H] somewhere that when you undervolt a cpu (or might have been GPU), that the current drain and total power actually went up? Maybe something the silicon is doing internally, or maybe the VRM's? It's counter-intuitive but maybe that was a factor in this LN2 overclock... make a change that increases the total current and therefore power, but cool with LN2 to offset that. LN2 overclocking is something I have never done, but maybe one of those experts might be able to comment or confirm if lowering vcore can increase current, which seems backwards to me. Might not even need to be an LN2 enthusiast, but if someone had all the correct testing equipment, a few tests could probably be done to see if the current can go up with a lower vcore... Maybe I'm remembering what I read wrong too :)
 
  • Like
Reactions: blkt
like this
Dan_D said:
Have you not had internet access for the past couple of years? I've done the board reviews during that time. I've looked at the data, graphed it out and I've had all these CPU's in hand save for the Ryzen 3000 series. When Ryzen launched originally, it delivered performance that roughly traded blows with Haswell. That's not a claim, that's fact. Go look at the reviews from that time frame. There were plenty of clock for clock comparisons done by many review outlets. Intel was still had generations of small improvements over Haswell. You had Haswell refresh, Broadwell, Skylake, and Kaby Lake, Even pulling out the stuff that wasn't an architectural improvement, you still had at least Broadwell and Skylake. Each brought an increase over Haswell. All the processor releases brought something to the table. Generally, clock speeds. Keep in mind that Intel still maintained a substantial clock speed advantage over AMD at this time and still does.

Even with the performance loss in some cases with all the security patches that hurt Intel's performance, it still has an IPC lead in some cases, if not most cases against Ryzen 1000 and 2000 series parts. I was running a Haswell-E based system using a 8c/16t Core i7 [email protected]. I switched to a 12c/24t Threadripper [email protected]. It was pretty much a lateral move outside of cases where an application can use more than 8c/16t. In those cases, Threadripper is obviously faster. I had the 5960X for years and I've been running the Threadripper since April. I can tell you it wasn't much of an upgrade.
Click to expand...


Can't state the legitimacy of it but what we do know is that out the box the 3600 is nowhere near the clockspeed of the 8700K or 9900K

amd-ryzen-5-3600-x570-tests-2-jpg.jpg


amd-ryzen-5-3600-x570-tests-1-jpg.jpg


amd-ryzen-5-3600-x570-juegos-1-jpg.jpg


amd-ryzen-5-3600-x570-juegos-3-jpg.jpg


amd-ryzen-5-3600-x570-juegos-4-jpg.jpg


amd-ryzen-5-3600-x570-juegos-2-jpg.jpg


Again waiting for clarity, I don't really believe this sight but their original ryzen reviews were pretty solid. If the 3600 is running stock that is 4.2ghz max frequency, maybe it is sustained all core for that part as well but the results are impressive if that is the case.

The real only anomaly there is Far Cry 5 and maybe there is an explanation to it but everything else is very competitive.

If one is to assume that it is true for argument sake, then clearly clockspeed is not the most important factor.

I will like to wait on GamersNexus EPS12v readings for something like Cinebench where the 9900K chugs power on a crazy scale compared to the 2700, even thread ripper used half the power the 9980X was chugging which was around 500w. That is more interesting to me than clockspeed and if this is true it shows inefficiency even if the clockspeed is there to offset it.
 
I see LN2 overclocks as pretty much useless. This is not usable on a day to day basis. It's just for a brief test and showing off.

The question I have is how high you can get on all cores on a high end custom water loop.
 
Rockenrooster said:
I don't think anyone was expecting 5GHz on 16 cores...
I wonder if the 8 core can OC.
Also, don't forget the IPC difference. I'd rather have 15% more IPC than 15% more clock...
Click to expand...

I don't feel like this distinction matters.

I care about overall performance per core. If that is achieved by a higher clock at max overclock or by a much better IPC is totally irrelevant, IMHO.
 
Rockenrooster said:
I don't think anyone was expecting 5GHz on 16 cores...
I wonder if the 8 core can OC.
Also, don't forget the IPC difference. I'd rather have 15% more IPC than 15% more clock...
Click to expand...

I don't really care how the performance is achieved. I wouldn't care if the CPU was at its core a Pentium D @ 20GHz provided that it out performed everything else and we could actually keep it cool.

Zarathustra[H] said:
I see LN2 overclocks as pretty much useless. This is not usable on a day to day basis. It's just for a brief test and showing off.

The question I have is how high you can get on all cores on a high end custom water loop.
Click to expand...

Indeed. The only thing that the LN2 clocks tell us is that we should temper our expectations for its clocks on water.

Zarathustra[H] said:
I don't feel like this distinction matters.

I care about overall performance per core. If that is achieved by a higher clock at max overclock or by a much better IPC is totally irrelevant, IMHO.
Click to expand...

Agreed.
 
OrangeKhrush said:
laughing is already something we do enough of in this country to get by. It is a country that promised so much but pushes for socialism and other social justice politicking has set it back. In 2011 we traded around 6:1 with the USD, by 2013 and the first leaks of state capture and corruption that plummeted and in 2019 we are at 14:1 to the USD. Despite the problems the country has it remains a hotspot for foreigners, with american tourists being prominent.

I guess anyone can laugh at another persons country for whatever reason, but that doesn't always mean we live any worse than you do, I earn good money, I have beach side property, I have a retierment annuity and investments, I have a fantastic gun collection and despite living here I got a shot at professional baseball. I have thought about moving to the United States but immigration is a lot harder these days.
Click to expand...

Off topic, but, literally nothing about the current state of south afrika makes me want to laugh.
 
OrangeKhrush said:
laughing is already something we do enough of in this country to get by. It is a country that promised so much but pushes for socialism and other social justice politicking has set it back. In 2011 we traded around 6:1 with the USD, by 2013 and the first leaks of state capture and corruption that plummeted and in 2019 we are at 14:1 to the USD. Despite the problems the country has it remains a hotspot for foreigners, with american tourists being prominent.

I guess anyone can laugh at another persons country for whatever reason, but that doesn't always mean we live any worse than you do, I earn good money, I have beach side property, I have a retierment annuity and investments, I have a fantastic gun collection and despite living here I got a shot at professional baseball. I have thought about moving to the United States but immigration is a lot harder these days.
Click to expand...

I'm laughing at your expectation that the rest of the world would know where you are from, what currency you use, and what pricing looks like beyond the exchange rate. This is a fundamental inability to communicate.
 
BrotherMichigan said:
While it may be true that a die shrink doesn't necessarily reduce the voltage, it has certainly happened since 2006. As recently as 2011 for Intel, Vdd was 0.75V for their 22nm process (it was the same for 32nm), which dropped to 0.70V for 14nm and 10nm.
Click to expand...

With true Dennard scaling voltage would have reduced ~36%. Of course, foundry engineers can tweak nodes here and there to get some improvement as that 6% you quote. 12LP also bring us reduced voltages compared to 14LPP, but it wasn't due to shrink (geometries are identical), but due to optimizations in the node

7c02bb7b9df96f40c1ffae935087eead5ff59a804c563082d37c4f040bd6d9c4.png


7HPC will not bring lower voltages than 14LPP because it is a shrink, but because HPC is optimized for higher clocks than LPP.

We have to wait to Zen2 reviews, but we already have some datapoints for the 7HPC node

2.8 GHz <---> 0.775 V
3.0 GHz <---> 0.825 V
3.5 GHz <---> 0.950 V
4.0 GHz <---> 1.200 V
4.2 GHz <---> 1.375 V

4.2 GHz <---> 1.344 V
5.0 GHz <---> 1.608 V
5.2 GHz <---> 1.784 V
 
Dan_D said:
I don't really care how the performance is achieved. I wouldn't care if the CPU was at its core a Pentium D @ 20GHz provided that it out performed everything else and we could actually keep it cool.



Indeed. The only thing that the LN2 clocks tell us is that we should temper our expectations for its clocks on water.



Agreed.
Click to expand...


I haven't played with Ryzen at all yet. Does it typically allow you to set your own core count based overclocks?

You know, 5.2ghz for up to 4 cores loaded, 5.0ghz for up to 8 cores, 4.8 GHz for all core? Something like that?

I'd be happy with something like this, because generally when I need my best lightly threaded performance (gaming) my system isn't doing anything else.

I'd want to do it in the BIOS though, not in Ryzen Master or something like that. Most of the time I am not running Windows.
 
Last edited:
juanrga said:
With true Dennard scaling voltage would have reduced ~36%. Of course, foundry engineers can tweak nodes here and there to get some improvement as that 6% you quote. 12LP also bring us reduced voltages compared to 14LPP, but it wasn't due to shrink (geometries are identical), but due to optimizations in the node

7c02bb7b9df96f40c1ffae935087eead5ff59a804c563082d37c4f040bd6d9c4.png


7HPC will not bring lower voltages than 14LPP because it is a shrink, but because HPC is optimized for higher clocks than LPP.

We have to wait to Zen2 reviews, but we already have some datapoints for the 7HPC node

2.8 GHz <---> 0.775 V
3.0 GHz <---> 0.825 V
3.5 GHz <---> 0.950 V
4.0 GHz <---> 1.200 V
4.2 GHz <---> 1.375 V

4.2 GHz <---> 1.344 V
5.0 GHz <---> 1.608 V
5.2 GHz <---> 1.784 V
Click to expand...


Seems really disingenuous for GlobalFoundries to call those different processes different process sizes.

I wish they would return to defining nodes sizes by its original intent, the actual minimum feature size the process supports.

I'm surprised none of them have been sued over this yet.
 
OrangeKhrush said:
Can't state the legitimacy of it but what we do know is that out the box the 3600 is nowhere near the clockspeed of the 8700K or 9900K

View attachment 170301

View attachment 170302

View attachment 170303

View attachment 170304

View attachment 170305

View attachment 170306

Again waiting for clarity, I don't really believe this sight but their original ryzen reviews were pretty solid. If the 3600 is running stock that is 4.2ghz max frequency, maybe it is sustained all core for that part as well but the results are impressive if that is the case.

The real only anomaly there is Far Cry 5 and maybe there is an explanation to it but everything else is very competitive.

If one is to assume that it is true for argument sake, then clearly clockspeed is not the most important factor.

I will like to wait on GamersNexus EPS12v readings for something like Cinebench where the 9900K chugs power on a crazy scale compared to the 2700, even thread ripper used half the power the 9980X was chugging which was around 500w. That is more interesting to me than clockspeed and if this is true it shows inefficiency even if the clockspeed is there to offset it.
Click to expand...

You gotta remember that most of those can take advantage of multiple cores. i9 being on top since it is 8 cores vs the 3600's 6. But whats really missing is info regarding what clock speeds everything was at (I realize you mentioned it). Was anything overclocked?

The good news for AMD is that it appears they are closing the IPC gap. I suspect Intel will remain on top due to better overclocking/turbo-boost vs the AMD parts which seem already close to the bleeding edge.

Far Cry 5 only uses around 4 cores (from the quick little research I've done), so you would expect the performance to look similar to the single core graph... I know cinebench supposedly favors AMD, so perhaps the first few graphs make the performance appear closer (in per core performance) than it really is between the 2 architectures... I wouldn't use a single benchmark as a definitive indicator of performance.

More and better benchmarks will tell.

Someone needs to make an All In One LN2 self contained cooler, might be a market for such a thing :)
 
Last edited:
GoodBoy said:
I know cinebench supposedly favors AMD, so perhaps the first few graphs make the performance appear closer (in per core performance) than it really is between the 2 architectures...
Click to expand...

I'm not convinced there is any truth to this. Yes, it seems to scale better with more cores than many applications, and true, AMD currently has the advantage in this regard, but if you look at single threaded results, I think any claims of it being an AMD advantaged benchmark are untrue. I like Cinebench as a benchmark as it shows you both extremes of a CPU. Single Threaded results, and best case multithreaded results. It has its limits, but it is of value for that reason. I've been a proponent of using Cinebench for benchmarks for this reason for years. Back in the Bulldozer days when I proposed liking at Cinebench single threaded results, I was accused of peddling an INTEL biased benchmark :p
 
juanrga said:
The function relating voltage to frequency doesn't depend on the number of cores.
Click to expand...

This is only true if you ignore the thermal dynamics of electricity. Heat causes resistance, which means less cores equals less heat, which equals less resistance, which equals less voltage to achieve the same clock speeds. Hence, why you can't compare 6 core numbers with 16 core numbers, as well as to why they are hitting 6 Ghz on lower core count parts.


juanrga said:
4.2Ghz is far beyond the optimal range for 12LP,


Click to expand...

With this, and your spoiler tag, I am not sure what point you are trying to make here in retaliation to what I stated. Maybe you posted this to disprove your own math theory about R5 and R9 being comparable in voltage (1.344v vs 1.6v)

juanrga said:
Dennard scaling ceased to work around 2006. No longer a die shrink reduces the voltage.

I have given voltages for X470 boards, for X570 boards, and for non-AM4 boards.
Click to expand...

Again, this would only be true if you ignore the thermal dynamics of electricity. Die shrink equals less heat, which equals less resistance, which equals less voltage to achieve same clock speeds. This is one of the reasons that a die shrink gives us head room to clock higher.
 
Last edited:
NWRMidnight said:
... Die shrink equals less heat...
Click to expand...

Not exactly true. You are packing more transistors into a smaller space, the heat in that space intensifies. They've hit the power-wall, and the dennard scaling has ended.

Check out: https://cs.nyu.edu/courses/spring12/CSCI-GA.3033-012/lecture12.pdf Page 3, I love that a nuclear reactor is just up the graph from a pentium on the scale for power density... lol...

From a paper written in 2007:
Moore’s Law and Dennard Scaling
• Moore’s Law: transistor density doubles every N years (currently N ~ 2)
• Dennard Scaling:
• Shrink feature size by k (typ. 0.7), hold electric field constant
• Area scales by k 2 (1/2) , C, V, delay reduce by k
• P  CV 2f  P goes down by k 2

• We never saw this in practice:
• Generally kept area constant, used doubled transistor density to add more features, so C didn’t scale
• Aggressive pipelining, circuits, etc. to boost f beyond “natural” rate
• Leakage
• Power and power density went up, not down

• “Natural” frequency growth may slow even further due to increasing parasitics, parameter variations, etc.
Click to expand...

Natural Frequency growth has drastically slowed.. 2007 we had what, 2.0 to 3.0 Ghz cpu's, and we got near or at 4.0Ghz in 2010. Here it is 2019 and we are squeeing at the prospect of 5.0 Ghz.. and really, if we look at old overclocking records, they had hit 5 to 6 Ghz back then with LN2, now 10+ years later and it is still about the same...

Going to take a whole new material to replace silicon, or something like a processor that uses light instead of electricity, to get some kind of major boost to CPU performance...
 
Last edited:
GoodBoy said:
Not exactly true. You are packing more transistors into a smaller space, the heat in that space intensifies. They've hit the power-wall, and the dennard scaling has ended.

Check out: https://cs.nyu.edu/courses/spring12/CSCI-GA.3033-012/lecture12.pdf Page 3, I love that a nuclear reactor is just up the graph from a pentium on the scale for power density... lol...
Click to expand...

Are you sure about that? :

Die shrinks are beneficial to end-users as shrinking a die reduces the current used by each transistor switching on or off in semiconductor devices while maintaining the same clock frequency of a chip, making a product with less power consumption (and thus less heat production), increased clock rate headroom, and lower prices.
Click to expand...

https://en.wikipedia.org/wiki/Die_shrink
 
NWRMidnight said:
Are you sure about that? :



https://en.wikipedia.org/wiki/Die_shrink
Click to expand...

Overall power per transistor goes down, but we also increase transistor count, so overall power density per unit area of the core can go up, making it more difficult to cool.

In other words, these are not mutually exclusive.

Take two chips with the identical design and same clockspeed one in a smaller node and one on a larger node, the smaller node can get away with less voltage and thus less power.

The thing is, transistor count has drastically gone up as dies have shrunk, so we rarely see apples to apples comparisons between nodes like that.

Closest we got was probably the tock comparison to the tick on Intel's old tick-tock model, but even then the chips were not identical.
 
Last edited:
GoodBoy said:
Read the definition of dennard scaling (which was true for a long time) and how that scaling broke down in 2006: https://en.wikipedia.org/wiki/Dennard_scaling
Click to expand...

My understanding is Dennard scaling has broken down due to Moore's Law and the doubling of transistors. But the transistor count has not changed on Ryzen (even with the die shrink it is still only 4.8 billion transistors per 8 cores, which is identical to the first Ryzen series).
 
  • Like
Reactions: blkt
like this
NWRMidnight said:
Except the transistor count the Ryzen series hasn't changed from generation to generation: They are all 4.8 billion transistors per 8 cores:

https://en.wikipedia.org/wiki/Ryzen
Click to expand...

Fair, but they have upped the clocks on things and introduced a faster Infinity Fabric and PCIe bus, right?

My point was that if everything is apples to apples a smaller die can run on less voltage. I don't think we've ever seen an exact apples to apples product on different process nodes though.

Some have come very close, but usually the transition also involves some optimizations and other changes as well as increases in capability.

That, and as others have said already, with today's tiny process nodes, things don't scale as well as they used to as we are getting close to the limits of silicon.

They still scale somewhat, mind you, or there wouldn't be much reason to pursue a die shrink other than silicon wafer utilization savings, but nowhere near where it once was.
 
Zarathustra[H] said:
Fair, but they have upped the clocks on things and introduced a faster Infinity Fabric and PCIe bus, right?

My point was that if everything is apples to apples a smaller die can run on less voltage. I don't think we've ever seen an exact apples to apples product on different process nodes though.

Some have come very close, but usually the transition also involves some optimizations and other changes as well as increases in capability.

That, and as others have said already, with today's tiny process nodes, things don't scale as well as they used to as we are getting close to the limits of silicon.

They still scale somewhat, mind you, or there wouldn't be much reason to pursue a die shrink other than silicon wafer utilization savings, but nowhere near where it once was.
Click to expand...

Nothing is every apples to apples comparison anymore. :D
 
NWRMidnight said:
My understanding is Dennard scaling has broken down due to Moore's Law and the doubling of transistors. But the transistor count has not changed on Ryzen (even with the die shrink it is still only 4.8 billion transistors per 8 cores, which is identical to the first Ryzen series).
Click to expand...


It also says this:

The primary reason cited for the breakdown (of Dennard scaling since 2006-2007) is that at small sizes, current leakage poses greater challenges and also causes the chip to heat up, which creates a threat of thermal runaway and therefore further increases energy costs.[2
 
Zarathustra[H] said:
It also says this:

The primary reason cited for the breakdown (of Dennard scaling since 2006-2007) is that at small sizes, current leakage poses greater challenges and also causes the chip to heat up, which creates a threat of thermal runaway and therefore further increases energy costs.[2
Click to expand...
I think that is taking into account for Moore's Law of transistor count doubling every 2 years. They are not comparing equal transistor count, as the sentence before that states :

Since around 2005–2007 Dennard scaling appears to have broken down. As of 2016, transistor counts in integrated circuits are still growing, but the resulting improvements in performance are more gradual than the speed-ups resulting from significant frequency increases.
Click to expand...

Everything Dennard scaling slow down is referring to involves transistor count increases. Ryzen won't have any transistor count increases until Ryzen 3 using 7nm + node.

I could be wrong, but my take is a die shrink without a transistor count increase will result in a reduction in power requirements. Which hasn't been the case since 2006, as all die shrinks have included a transistor count increase till Ryzen. This could also be why we are not seeing the huge clock speed increases people where expecting.
 
Last edited:
sounds more like they're just trolling Intel which they deserve the right to do after all the abuse they received during bulldozer.
 
GoodBoy said:
The good news for AMD is that it appears they are closing the IPC gap.
Click to expand...

Compared to Skylake, not compared to Icelake.

Zarathustra[H] said:
I'm not convinced there is any truth to this. Yes, it seems to scale better with more cores than many applications, and true, AMD currently has the advantage in this regard, but if you look at single threaded results, I think any claims of it being an AMD advantaged benchmark are untrue. I like Cinebench as a benchmark as it shows you both extremes of a CPU. Single Threaded results, and best case multithreaded results.
Click to expand...

The reason why Cinenbech is an outlier doesn't have anything to do with scaling better with more cores. It has to do with abnormal low IPC gap and abnormal high SMT yield:

Cinebench R15 is some sort of a best case benchmark for AMD, that's why it's an outlier.
This. The IPC difference is abnormally low (5.6% vs. 14.4% average) and the SMT yield is abnormally high (41.6% vs. 28.7% average).
Click to expand...
Click to expand...

Cinebench is far from what rest of benchmarks show for Zen/Zen+.

NWRMidnight said:
This is only true if you ignore the thermal dynamics of electricity. Heat causes resistance, which means less cores equals less heat, which equals less resistance, which equals less voltage to achieve the same clock speeds.
Click to expand...

No, because heat depends on area, so the thermal density is the same.

NWRMidnight said:
This is one of the reasons that a die shrink gives us head room to clock higher.
Click to expand...

Already explained why this isn't true since 2006.
 
Last edited:
sirmonkey1985 said:
sounds more like they're just trolling Intel which they deserve the right to do after all the abuse they received during bulldozer.
Click to expand...

Well isn't the running gag now that all of the best Intel benchmarks can be discarded now due to their need for security fixes ;)
 
juanrga said:
Compared to Skylake, not compared to Icelake.



The reason why Cinenbech is an outlier doesn't have anything to do with scaling with more cores. It has to see with abnormal low IPC gap and abonrmal high SMT yield. Cinebench is far from what rest of benchmarks show for Zen/Zen+





No, because heat depends on area, so the thermal density is the same



Already explained why this isn't true since 2006.
Click to expand...

Uh, nope. You don't really have a clue about thermal dynamics do you?

As for the rest, may i suggest you read my other comments because your explanation is reliant on an increase in transistor count, which does not exist with zen, zen+, or zen 2.
 
Last edited:
juanrga said:
Compared to Skylake, not compared to Icelake.
Click to expand...

I guess it's a good thing for AMD that they won't be competing against Ice Lake in the desktop space then, eh? They'll have a relatively solid six months at least (including the holiday season) before any new product response from Intel, and we don't know yet if Comet Lake will present any real improvement over Coffee Lake other than two additional cores.
 
NWRMidnight said:
As for the rest, may i suggest you read my other comments because your explanation is reliant on an increase in transistor count, which does not exist with zen, zen+, or zen 2.
Click to expand...

No.

BrotherMichigan said:
I guess it's a good thing for AMD that they won't be competing against Ice Lake in the desktop space then, eh? They'll have a relatively solid six months at least (including the holiday season) before any new product response from Intel, and we don't know yet if Comet Lake will present any real improvement over Coffee Lake other than two additional cores.
Click to expand...

Desktop is a niche. The volume of sales is in mobile.
 
Status
Not open for further replies.
Back
Top