That works for marketing only because there isn't a mainstream 8C/8T chip to compare it to (yet).I don't buy much of that. 6C/12T is a multiple of three and customers have no problem with it. The 7700K will continue to fill the 4C/8T position. IIRC, HT is only ineffective in a couple of isolated instances, and even then the performance hit is negligible.
Because of the highly synchronized nature of computationally intensive HPC parallel codes, the usefulness of hyper-threading can be very limited. It should also be noted that with commercial CFD applications, a user should weigh the technical and performance benefits of hyper-threading against the possible additional licensing costs that might be incurred. ANSYS’ suggestion is to disable HyperThreads in the system BIOS on each node. This is discussed in ANSYS’ Installation and Licensing Documentation. In our testing, it was found to be adequate to request PBS resources in such a way that the HyperThreads were not included in the cpuset.
Nah. Coffee Lake went to 6 core because Cannon Lake was going to be 8 cores, but Intel needed a backup plan in case 10nm was as prickly as 14nm was. These CPUs were in development long before a competitor came up with a octocore CPU. The CFL hexacore was clearly moved forward - that's all Intel could do in response to Summit Ridge.Seriously, the only thing 6/12 is good for is so we have something cheaper than 8/16 to buy. That's it. These days, 8/16 is where the upgrade should be focused coming from the ancient 4/8. Coffee Lake is a stopgap chip and yet again represents Intel's laziness and lack of innovation since all they did in response to the competition was rush forward some stuff they already had cooking for much later. Intel would have waited another 10 years before delivering an 8/16 chip, but now I suspect we'll get it in a year or two.
But the 7700K is not socket-compatible with 8th-gen Core motherboards. ?I don't buy much of that. 6C/12T is a multiple of three and customers have no problem with it. The 7700K will continue to fill the 4C/8T position. IIRC, HT is only ineffective in a couple of isolated instances, and even then the performance hit is negligible.
^ This. HT is disabled on the servers I run all my physical design jobs on (placement, routing, STA, CTS, extraction, LEC, PDV-DRC, PDV-LVS, etc.)That works for marketing only because there isn't a mainstream 8C/8T chip to compare it to (yet).
As for HT being ineffective, I'm biased since my job used to depend on intensive multi-processor, many core simulations. There HT was pretty disastrous in my uses. Here is HP's take on it (see page 20):
https://www.hpe.com/h20195/V2/getpdf.aspx/a00007032enw.pdf?ver=1.0
By the same argument, what can you do with 8 cores that you can't do with six faster ones? It obviously depends on the application. Eight cores is not inherently better than six, especially when the six cores have a faster clockspeed and an ipc advantage in most applications.Concerning 4/8 thread vs 6/6, I almost care enough to do a simulated test, which would be easy to do. I'm curious as to how they perform when compared, but honestly I'm hoping someone else cares enough to do the test. I care about 80% and someone needs to care 100% in order to trigger the necessary motivation to actually test it.
Also, I feel 6/12 just isn't enough coming from 4/8. These days, cores are being thrown at us by the bucket full and I don't see what you can do with 6 that you can't sort of already do with 4. 8/16 makes much more sense because there you have substantial differences in streaming, productivity etc, and in a meaningful way. But 6/12? That's just a weak half measure and won't help with much really.
The only reason I have gone with 6/12 in the past is because I want those extra cores/threads for the games that use them, even if they only use them a little, and I wasn't going to spend $1000 on an 8 core just to game.
Seriously, the only thing 6/12 is good for is so we have something cheaper than 8/16 to buy. That's it. These days, 8/16 is where the upgrade should be focused coming from the ancient 4/8. Coffee Lake is a stopgap chip and yet again represents Intel's laziness and lack of innovation since all they did in response to the competition was rush forward some stuff they already had cooking for much later. Intel would have waited another 10 years before delivering an 8/16 chip, but now I suspect we'll get it in a year or two.
By the same argument, what can you do with 8 cores that you can't do with six faster ones? It obviously depends on the application. Eight cores is not inherently better than six, especially when the six cores have a faster clockspeed and an ipc advantage in most applications.
I think streaming will be better on a slower 8/16 chip in terms of smoothness. There are some things that just need the cores and threads and having less faster cores won't make up for it. I .
No, not really. Faster cores can make up the difference.
It simply depends on how much faster the the lower core count machines are.
Example at identical IPC: If you have a 6/12 machine running 5 GHz, it will ALWAYS be faster than and 8/16 machine running 3.5 GHz, even in the worse case of near 100% parallel loads, and the further from parallel the load gets the greater the advantage will be.
For the Intel CL 6/12 vs Ryzen 8/16 situation it will be messier. Intel has better IPC/clockspeed, AMD sometimes gets a bigger boost from SMT than Intel gets from HT. If Intel clocks close to 5GHz, it should be a dogfight on the near 100% parallel loads, with Intel winning just about everything else.
It should be a very interesting review.
I think streaming will be better on a slower 8/16 chip in terms of smoothness. There are some things that just need the cores and threads and having less faster cores won't make up for it. I know the 8700K will be great. I just think for Intel's big move to finally get away from quad cores they should have gone for an even doubling of cores and not some half measure 6 core move. 8/16 would have been much, much better and brought them inline with what's expected these days from a mainstream high-end chip. Like I said, I voted no.
No, not really. Faster cores can make up the difference.
It simply depends on how much faster the the lower core count machines are.
Example at identical IPC: If you have a 6/12 machine running 5 GHz, it will ALWAYS be faster than and 8/16 machine running 3.5 GHz, even in the worse case of near 100% parallel loads, and the further from parallel the load gets the greater the advantage will be.
For the Intel CL 6/12 vs Ryzen 8/16 situation it will be messier. Intel has better IPC/clockspeed, AMD sometimes gets a bigger boost from SMT than Intel gets from HT. If Intel clocks close to 5GHz, it should be a dogfight on the near 100% parallel loads, with Intel winning just about everything else.
It should be a very interesting review.
I don't know about that. Many game engines these days are built around, or trending towards the concept of task based parallelism in which a single or multiple tasks are broken up and run on a single thread or across multiple threads. In these kinds of games built on these engines, thread counts matter a lot. Give you an example. Ubisoft engines typically can scale to very high thread counts, and Watch Dogs 2 is a great example. Here you see the deca-core 6950x which has a base clock of 3ghz, thoroughly trounce the much higher clocked quad core 7700K which has a base clock of 4.2ghz and better IPC. Even the 1800x manages to beat the 7700K in this CPU limited scenario.
Now I'm not saying every game is going to be like Watch Dogs 2 of course. My point is that the industry is trending towards greater levels of parallelism, in which core or thread counts will be perhaps the most important factor for performance.
Also, I still think your chip is better than all this other crap out there, even the 8700K. Too bad that chip still cost $1,000. I'd be tempted to just grab one of those. You have the last no compromise 8/16 chip until they hit Intel's mainstream with hopefully a ring bus.
1,280x720 i fail to see the relevance of such a test for your argument. Most of us here, never use that resolution. Do you have a graph at, at least 1080p or higher?
These comparisons make sense only in Cinebench-type workloads, when running in a benchmarking environment. Fewer fast cores != More, less fast cores in many situations that are reflective of real-world usage.I think you missed the point.
I was addressing a claim that clock speed could NOT make up for cores.
This is a False point. Obviously if the clock speed is not fast enough it can't overcome core count, but there is always a point were clock speed can over come cores if you can run that fast.
To get a starting estimate on how they will perform on parallel loads, just multiply core count clock speed to see which number is higher.
Your example
6950K 10 cores * 3 GHz = 30
7700K 4 cores * 4.2 GHz = 17
So obviously the 6950 has a massive performance advantage on parallel loads.
OTOH if you could run the 7700K at 8 GHz (given enough memory bandwidth) it would be faster than the 6950K (4 cores X 8 GHz = 32 ).
Enough clock speed can overcome core counts.
For the 7700K vs 8700K vs Ryzen assuming 5GHZ Intel and 4GHz AMD:
7700K 4 cores * 5 GHz = 20
8700K 6 cores * 5 GHz = 30
Ryzen 8 cores * 4 GHz = 32
Note how close the 8700K comes to the Ryzen in this case, and that doesn't include Intels IPC advantage.
If 8700K overclocks to 5GHz it will be very close to Ryzen on near parallel loads, and destroy it on less parallel loads. This is why it should make for a very interesting set of reviews and a very interesting processor.
I think you missed the point.
I was addressing a claim that clock speed could NOT make up for cores.
This is a False point. Obviously if the clock speed is not fast enough it can't overcome core count, but there is always a point were clock speed can over come cores if you can run that fast.
To get a starting estimate on how they will perform on parallel loads, just multiply core count clock speed to see which number is higher.
Your example
6950K 10 cores * 3 GHz = 30
7700K 4 cores * 4.2 GHz = 17
So obviously the 6950 has a massive performance advantage on parallel loads.
OTOH if you could run the 7700K at 8 GHz (given enough memory bandwidth) it would be faster than the 6950K (4 cores X 8 GHz = 32 ).
Enough clock speed can overcome core counts.
For the 7700K vs 8700K vs Ryzen assuming 5GHZ Intel and 4GHz AMD:
7700K 4 cores * 5 GHz = 20
8700K 6 cores * 5 GHz = 30
Ryzen 8 cores * 4 GHz = 32
Note how close the 8700K comes to the Ryzen in this case, and that doesn't include Intels IPC advantage.
If 8700K overclocks to 5GHz it will be very close to Ryzen on near parallel loads, and destroy it on less parallel loads. This is why it should make for a very interesting set of reviews and a very interesting processor.
I thought big little & the latest A11 (2+4) would show some the merits of more cores, but apparently not.These comparisons make sense only in Cinebench-type workloads, when running in a benchmarking environment. Fewer fast cores != More, less fast cores in many situations that are reflective of real-world usage.
I'd take a Xeon 2650 v2 over an i7 3770 any day, because unlike the i7 the Xeon doesn't throw up when a colleague runs a FORTRAN code using SSH, which makes the 40-tab Firefox session of the previous night crash as soon as you log in on the 3770. The Xeon on the other hand is happy to oblige.
An interesting explanation, but it's overly simplistic in several areas, the most obvious being that it assumes that clockspeed always scales linearly, when we know for a fact that it doesn't. The P4 experiment taught us that much at least, because very high clock speeds require sacrifices in other areas, namely IPC, as the instruction pipeline has to be very deep to achieve those speeds. Every CPU and GPU architecture has a sweet spot when it comes to frequency, and going above that results in diminishing returns, or zero performance improvement. Even if you could theoretically push a 7700K to 8ghz, there is no guarantee that it would outperform the 6950x in highly parallel workloads.
The lack of linear scaling from clock speed along with the difficulty of increasing clock speeds due to thermodynamics is the main reason why CPU manufacturers have had to resort to increasing core count and instruction level parallelism to increase performance. It's easier to design an octa core CPU that runs at 3ghz, rather than a quad core CPU that runs at 8ghz. And with the programming models becoming more and more parallel, it doesn't even make sense to focus on clock speeds anymore.
It will always make sense to focus on clock speeds for a general purpose CPU.
There is no great Parallel Computing Nirvana just waiting for us, if enough people buy multi-core CPU's.
This is just wishful thinking.
An interesting explanation, but it's overly simplistic in several areas, the most obvious being that it assumes that clockspeed always scales linearly, when we know for a fact that it doesn't. The P4 experiment taught us that much at least, because very high clock speeds require sacrifices in other areas, namely IPC, as the instruction pipeline has to be very deep to achieve those speeds. Every CPU and GPU architecture has a sweet spot when it comes to frequency, and going above that results in diminishing returns, or zero performance improvement. Even if you could theoretically push a 7700K to 8ghz, there is no guarantee that it would outperform the 6950x in highly parallel workloads.
The lack of linear scaling from clock speed along with the difficulty of increasing clock speeds due to thermodynamics is the main reason why CPU manufacturers have had to resort to increasing core count and instruction level parallelism to increase performance. It's easier to design an octa core CPU that runs at 3ghz, rather than a quad core CPU that runs at 8ghz. And with the programming models becoming more and more parallel, it doesn't even make sense to focus on clock speeds anymore.
Performance of a specific core does scale linearly with Clock speed, as long as memory bandwidth doesn't bottleneck it, which was the stipulation in my post.
You are talking about something else entirely on the P4, which was changing the core to try and make it run faster, which is NOT what I am talking about at all. I am simply talking about cranking up the speed on the same core.
Or on Coffee Lake vs Ryzen which is the case most people are interested in. Coffe Lake has both higher clock speed and higher IPC than Ryzen. Which is why it may challenge Ryzen while having less cores.
In highly parallel workloads, core count is always going to trounce clock speed, much like core speed always trounces core count in serial workloads.