Confirmed - i9 9900k will have soldered IHS, no more toothpaste TIM

[Hitman928]

It is not superior, To say it is superior is to imply Intel's implementation is not efficient; which is to say, don't talk about SMT in a vacuum.

It's hard to compare the two without a lot of details which probably aren't provided anywhere and even if they were, would take a lot of time and technical knowledge to break down. With that said, it's not true with modern SMT designs that greater SMT gains means more waste in the pipeline, that hasn't been a hard truth in quite a while. SMT is part of the design now from the onset. What we can say is true is that AMD's SMT implementation on average provides more of an uplift in well threaded applications than intel's does. This is easy to show and has been shown. You still have to take the overall performance into account, but that doesn't prevent the prior sentence from being any less true.

 

[TheELF]

Well, it is superior, and has been proven many times. You are trying to say its not ?

Edit, its not 50% boost, maybe 40%, I don't remember the finding

I said 33% because that's the difference in real cores between the i7-8700 and the 2700.
For the rest of the CPUs the difference is in SMT only which would also only be ~30-35% for HT,I put 50% in parenthesis because I have seen this figure being thrown around a lot of times,how does this make me sound desperate,as if 30% (or 40% as you state) isn't enough of a difference...

Also HTT/SMT only reflects how many instructions are being left unutilized from a core from a single thread,zen has more instructions staying idle,witnessed by the lower single thread speed.

 

[Markfw]

I said 33% because that's the difference in real cores between the i7-8700 and the 2700.
For the rest of the CPUs the difference is in SMT only which would also only be ~30-35% for HT,I put 50% in parenthesis because I have seen this figure being thrown around a lot of times,how does this make me sound desperate,as if 30% (or 40% as you state) isn't enough of a difference...

Also HTT/SMT only reflects how many instructions are being left unutilized from a core from a single thread,zen has more instructions staying idle,witnessed by the lower single thread speed.

Just to make myself clear. HT provides about a 20% boost in performance on vs off. AMD's SMP provides about 30-35% boost, that was my only point.
As far as cores, the performance depends on if its gaming, and the 8700k wins, or a multi-threaded app, in which case the 2700x usually wins.

 

[frozentundra123456]

Just to make myself clear. HT provides about a 20% boost in performance on vs off. AMD's SMP provides about 30-35% boost, that was my only point.
As far as cores, the performance depends on if its gaming, and the 8700k wins, or a multi-threaded app, in which case the 2700x usually wins.

No, the *per core* performance even in multithreaded apps is still in favor of the 8700k. That is because you are comparing 8 cores vs six. The 8700k also has more overclocking headroom than the 2700x.

 

[Markfw]

No, the *per core* performance even in multithreaded apps is still in favor of the 8700k. That is because you are comparing 8 cores vs six. The 8700k also has more overclocking headroom than the 2700x.

Yes, I know that. But I am comparing these two because they are very close in price, and the overall MT is usually won by the 2700x

 

[ehume]

This seems to be the first Big News in tech since Ryzen was released.

 

[The Stilt]

Just to make myself clear. HT provides about a 20% boost in performance on vs off. AMD's SMP provides about 30-35% boost, that was my only point.
As far as cores, the performance depends on if its gaming, and the 8700k wins, or a multi-threaded app, in which case the 2700x usually wins.

If you exclude the scenarios where SMT hurts the performance of either of the CPUs, AMD's and Intel's implementations are not that much apart (~33% vs. ~29% average yield).

 

[epsilon84]

Just to make myself clear. HT provides about a 20% boost in performance on vs off. AMD's SMP provides about 30-35% boost, that was my only point.
As far as cores, the performance depends on if its gaming, and the 8700k wins, or a multi-threaded app, in which case the 2700x usually wins.

As Stilt has pointed out above, your HT/SMT numbers are well off, 4% is hardly 10-15%. You've basically turned a negligible difference between the two into a significant one.

2700X wins most MT tests due to having 8 cores vs 6, not because of a vastly superior SMT implementation from AMD.

 

[Markfw]

As Stilt has pointed out above, your HT/SMT numbers are well off, 4% is hardly 10-15%. You've basically turned a negligible difference between the two into a significant one.

2700X wins most MT tests due to having 8 cores vs 6, not because of a vastly superior SMT implementation from AMD.

First I said I was unsure of the exact percentages, I thought it was was more than 4%, and I still think thats up for debate.. Second, I never said ANYTHING about why the 2700x won most MT benchmarks other than it does (most likely due to more cores)

 

[epsilon84]

First I said I was unsure of the exact percentages, I thought it was was more than 4%, and I still think thats up for debate.. Second, I never said ANYTHING about why the 2700x won most MT benchmarks other than it does (most likely due to more cores)

I never noticed where SMT had significantly better scaling than HT with the one exception of Cinebench, where the SMT yield is exeptionally high for Zen. For most other benchmarks/apps the difference really isn't much, Stilts done in depth testing on relative SMT yield so I'm not sure what is 'up for debate', it's not like he just makes up his numbers.

I know you didn't say anything about how the 2700X generally wins MT benches but it was the way you framed your post - mentioning a supposed 10 - 15% advantage from AMD's SMT vs Intel HT and then saying a 2700X wins MT benches in the following paragraph. Sometimes its what you *don't* say... I'm not having a go at you, just trying to be factually accurate here. AMD's slightly more efficient SMT implementation certainly doesn't hurt in MT apps, but at the end of the day 33% more cores has a far bigger difference to MT performance than ~4% better SMT scaling.

 

[Markfw]

I never noticed where SMT had significantly better scaling than HT with the one exception of Cinebench, where the SMT yield is exeptionally high for Zen. For most other benchmarks/apps the difference really isn't much, Stilts done in depth testing on relative SMT yield so I'm not sure what is 'up for debate', it's not like he just makes up his numbers.

I know you didn't say anything about how the 2700X generally wins MT benches but it was the way you framed your post - mentioning a supposed 10 - 15% advantage from AMD's SMT vs Intel HT and then saying a 2700X wins MT benches in the following paragraph. Sometimes its what you *don't* say... I'm not having a go at you, just trying to be factually accurate here. AMD's slightly more efficient SMT implementation certainly doesn't hurt in MT apps, but at the end of the day 33% more cores has a far bigger difference to MT performance than ~4% better SMT scaling.

Fair enough. I read other places (that I can't find by googling now) that said it was more like 10-15%, but for the moment, lets just leave it at that.

My main reason for this whole line was Zucker2's reply, that I thought was out of line.

 

[Hitman928]

If you exclude the scenarios where SMT hurts the performance of either of the CPUs, AMD's and Intel's implementations are not that much apart (~33% vs. ~29% average yield).

As a comparison, hardwarecanucks tested SMT on and off for a Ryzen 1700x and a 6900k. If you average the results for tests which are decently multithreaded beyond 8 cores, I get an average SMT uplift of:

AMD 28.9% vs intel 21.7%

This average does include 1 test where intel had negative scaling but AMD had a decent uplift. If you take that out, you get:

AMD 30.9% vs intel 26.3%

There also exists 1 outlier where intel seemed to scale unusually high. If you take that out as well, you get:

AMD 30.2% vs intel 24.0%

https://www.hardwarecanucks.com/for...0-amd-ryzen-7-1700x-review-testing-smt-3.html

I should also mention that there were multiple examples (mostly in games) where intel's HT resulted in negative scaling whereas AMD's SMT had no effect or a positive effect. I didn't include those in the averages but it should probably be considered in the discussion overall.

 

[epsilon84]

As a comparison, hardwarecanucks tested SMT on and off for a Ryzen 1700x and a 6900k. If you average the results for tests which are decently multithreaded beyond 8 cores, I get an average SMT uplift of:

AMD 28.9% vs intel 21.7%

This does include 1 test where intel had negative scaling but AMD had a decent uplift. If you take that out, you get:

AMD 30.9% vs intel 26.3%

There also exists 1 outlier where intel seemed to scale unusually high. If you take that out as well, you get:

AMD 30.2% vs intel 24.0%

https://www.hardwarecanucks.com/for...0-amd-ryzen-7-1700x-review-testing-smt-3.html

I should also mention that there were multiple examples (mostly in games) where intel's HT resulted in negative scaling whereas AMD's SMT had no effect or a positive effect. I didn't include those in the averages but it should probably be considered in the discussion overall.

6900K is an older gen architecture though, would have been more noteworthy to compare it to any ~Lake series architecture IMO. I'm sure once the 9900K hits and we have direct comparisons to a 2700X everything will become clear.

I'm also fairly sure the Skylake based HT enabled CPUs like the 8700K show very little negative scaling from HT in gaming, in any case including gaming numbers into an 'SMT yield' analysis is somewhat ingenious because Intel CPUs are actually strongest in gaming to begin with and is the one area where AMDs core/thread advantage generally doesn't translate to better performance.

It's like saying Vega 64 gains a bit more from DX12 than a 1080 Ti does, but then not mentioning the 1080 Ti is actually 20% faster still.

 

[Hitman928]

6900K is an older gen architecture though, would have been more noteworthy to compare it to any ~Lake series architecture IMO. I'm sure once the 9900K hits and we have direct comparisons to a 2700X everything will become clear.

This is true, but it's the best, most recent data I could find. If you have a more data, I'd be happy to look through it as well. It also correlates with Stilt's data as well. When you take out the outliers, he showed 3% higher relative improvement for AMD and hardwarecanucks showed 5%. Not exactly the same, but pretty close and it will obviously change depending on the test suite used.

I'm also fairly sure the Skylake based HT enabled CPUs like the 8700K show very little negative scaling from HT in gaming

Probably true, but again, I have no data on that.

in any case including gaming numbers into an 'SMT yield' analysis is somewhat ingenious because Intel CPUs are actually strongest in gaming to begin with and is the one area where AMDs core/thread advantage generally doesn't translate to better performance.

I don't think anyone included gaming numbers when presenting SMT yield, or at least, I didn't. If I had data from 4 core / 8 thread processors I might as that might actually yield interesting data. As it is, there's maybe one or two games out there that will scale beyond 8 cores, so using games when comparing SMT on / off with a 1700x and 6900k would be pointless, outside of the observation which I mentioned.

It's like saying Vega 64 gains a bit more from DX12 than a 1080 Ti does, but then not mentioning the 1080 Ti is actually 20% faster still.

No it's not, because you don't care about absolute performance when discussing SMT uplift, you're just calculating SMT uplift, that's all. Obviously you don't make any kind of performance or buying conclusion based on that one metric, it's just an interesting point of comparison between two architectures.

 

[frozentundra123456]

I never noticed where SMT had significantly better scaling than HT with the one exception of Cinebench, where the SMT yield is exeptionally high for Zen. For most other benchmarks/apps the difference really isn't much, Stilts done in depth testing on relative SMT yield so I'm not sure what is 'up for debate', it's not like he just makes up his numbers.

I know you didn't say anything about how the 2700X generally wins MT benches but it was the way you framed your post - mentioning a supposed 10 - 15% advantage from AMD's SMT vs Intel HT and then saying a 2700X wins MT benches in the following paragraph. Sometimes its what you *don't* say... I'm not having a go at you, just trying to be factually accurate here. AMD's slightly more efficient SMT implementation certainly doesn't hurt in MT apps, but at the end of the day 33% more cores has a far bigger difference to MT performance than ~4% better SMT scaling.

Exactly. The discussion at the time was per core performance, not total MT throughput. Actually the comment from Mark reinforces Elf's comment that an intel core is about 33% more powerful than Zen, since performance is close with a slight edge to Zen, but 33% more cores.

 

[Shmee]

I may end up getting an i9 9900k for gaming, depending on prices and OC performance/reviews.

 

[french toast]

I may end up getting an i9 9900k for gaming, depending on prices and OC performance/reviews.

Looks to be one of the best intel CPU buys for years, good longevity and a nice balance of ST performance and MT for productivity and 'future proofing'.

Everything depends on price though, realistically it should be $350... But in current climate Intel can get away with putting it into a higher price category, due to it's single thread performance alone.
Next year we wont get this nonsense as ST and gaming performance will be approx equal...

If can pick up a i9 9900k for $400, well that is a bargain imo.

 

[TheELF]

I should also mention that there were multiple examples (mostly in games) where intel's HT resulted in negative scaling whereas AMD's SMT had no effect or a positive effect. I didn't include those in the averages but it should probably be considered in the discussion overall.

As you already noticed yourself later on, SMT/HT tests on games should be done on less then 4 cores/threads.
HTT can give you up to 100% gains under the right circumstances,because that's what it is,it uses idle resources and games (as well as all of your common daily progs) run much less instructions than any DC/scientific app does.

 

[TheELF]

Just to make myself clear. HT provides about a 20% boost in performance on vs off. AMD's SMP provides about 30-35% boost, that was my only point.
As far as cores, the performance depends on if its gaming, and the 8700k wins, or a multi-threaded app, in which case the 2700x usually wins.

The bench I posted includes heavy multithreaded as well as less heavy and lightly threaded progs,it doesn't concentrate on one or the other to hammer a point through,it shows you an average for someone that will run all sorts of things on hishersits system.

 

[The Stilt]

As a comparison, hardwarecanucks tested SMT on and off for a Ryzen 1700x and a 6900k. If you average the results for tests which are decently multithreaded beyond 8 cores, I get an average SMT uplift of:

AMD 28.9% vs intel 21.7%

This average does include 1 test where intel had negative scaling but AMD had a decent uplift. If you take that out, you get:

AMD 30.9% vs intel 26.3%

There also exists 1 outlier where intel seemed to scale unusually high. If you take that out as well, you get:

AMD 30.2% vs intel 24.0%

https://www.hardwarecanucks.com/for...0-amd-ryzen-7-1700x-review-testing-smt-3.html

I should also mention that there were multiple examples (mostly in games) where intel's HT resulted in negative scaling whereas AMD's SMT had no effect or a positive effect. I didn't include those in the averages but it should probably be considered in the discussion overall.

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/#post-38770120

https://forums.anandtech.com/threads/ryzen-strictly-technical.2500572/#post-38770122

With the workloads showing negative scaling on either of the CPUs included, the average for AMD is 28.72% and 25.37% for Intel.
With the workloads showing negative scaling excluded, the average for AMD is 33.60% and 29.08% for Intel.

This is Summit Ridge and Kaby Lake.

In my test suite AMD has two workloads which show negative scaling, Intel has one.

 

[Zucker2k]

Fair enough. I read other places (that I can't find by googling now) that said it was more like 10-15%, but for the moment, lets just leave it at that.

My main reason for this whole line was Zucker2's reply, that I thought was out of line.

I don't know if I said anything out of line. The premise is simple; the stronger core (ipc, per thread) stands to gain less from HT/SMT because it can scale higher and have less resources (relative to overall core performance) left to tackle a second thread. For Intel, this is typically a ratio of 3:1 or 75/25. For AMD, it's about 67/33. So, to say AMD's SMT gains is "superior" is simplistic, and frankly not true. A lot of points have been made already (above) so I won't go into details, but at least you have to realize the blunder of picking two random numbers from two different architectures and drawing direct conclusions from the data. That's a real minefield if there ever was one.

 

[ehume]

Here's a Question: what code name will the i9 9900K come out using? My i7 8700k is a 14nm++ Coffee Lake. More of the same for the i9 9900k?

 

[TheELF]

In my test suite AMD has two workloads which show negative scaling, Intel has one.

If you got the same version of himeno as I found,it does a little check before it runs so if you run two iterations on the same core it decides to only run half the times on each...this completely defeats the purpose of HTT,it should run the same amount of instructions everything each time.

 

[Brotate]

Noob question but does this mean that you can't use a watercooler with these, that you wouldn't want to, or both?

 

[Markfw]

Noob question but does this mean that you can't use a watercooler with these, that you wouldn't want to, or both?

This has no effect with whether you can or not. All it means is that its MOST LIKELY not going to need as much cooling. but since it has 2 more cores, I would guess that the cooling required would be close to what the 8700k requires.