Originally posted by edwaleni
View Post
Announcement
Collapse
No announcement yet.
Intel Launches Cooper Lake Xeons CPUs, New Optane Persistent Memory + SSDs
Collapse
X
-
Originally posted by smitty3268 View Post
It means poor performance per watt vs the competition, which is the key metric for server parts.
10 years ago, yes, one node difference implied considerable power reduction.
These days, static power (leakage due to a transistor being powered on) is much higher and increases considerably with lower feature sizes.
Comment
-
Originally posted by vladpetric View Post
Not really ...
10 years ago, yes, one node difference implied considerable power reduction.
These days, static power (leakage due to a transistor being powered on) is much higher and increases considerably with lower feature sizes.
While it is right that some time in a not too distant future it would become impossible to improve the performance per watt without switching to different semiconductor materials or to completely different electronic devices, we are not there yet.
AMD has 2 great advantages over Intel, one is the lower cost for more cores and more cache than Intel (due to the multiple chiplet design) and the much better performance per watt due to the TSMC 7-nm CMOS process.
If you would compare the specifications of any Intel and AMD current CPUs, you would see that at the same power consumption and at the same number of cores, the AMD CPUs always have a much higher base clock frequency.
The consequence of this fact is seen in all benchmarks. When an Intel CPU has most of the time no more than 1 or 2 active cores, so that the speed is limited by the maximum turbo core frequencies, then Intel may win the benchmark, but in all benchmarks where enough cores are active so that the power limits are reached, at the same number of cores and at the same power consumption the Intel CPUs are much slower than the AMD CPUs, because the average clock frequency during the benchmark is higher for the latter.
The static power was worst for the 90-nm Intel CPUs (Prescott/Nocona), where it could exceed a half of the total power consumption.
The next CMOS processes since 2006 had lower static power, due to innovations like using high-permittivity gate dielectrics, using FinFETs since Intel Haswell and designing the CPUs with a mixture of different kinds of transistors, some of which are high-speed transistors while others are low-leakage transistors.
While designing a CPU with a low static power has become quite a complex task in modern processes, it can be done, as it should be obvious because the current desktop computers have lower idle power values than ever and the current laptops have also better battery lifetimes than ever.
So the conclusion is that for now AMD has a considerable advantage over Intel at power efficiency. Just look that the best Intel laptop 8-core CPU has a pathetic 2.4 GHz base frequency, while its cheaper competitor from AMD has a 3.3 GHz base frequency and the frequency ratio is confirmed by all benchmarks (when they are done at the same power consumption; many published benchmarks are done with the Intel CPUs consuming up to a double power than AMD).
Cooper Lake is not competitive with Rome, because even with the double AVX-512 throughput it is cheaper to use twice as many AMD cores and also have lower power consumption and better connectivity.
Nevertheless, there are certain very specialized applications where Cooper Lake can be the best choice. While deep learning should have been better done on GPUs, professional NVIDIA GPUs are even more expensive for a given performance than Cooper Lake. AMD GPUs are much cheaper than both Cooper Lake and NVIDIA, but for most applications they require a much greater software development effort, due to lack of suitable libraries and tools, which might be prohibitive.
There are also other niche applications where Cooper Lake can be best, which take advantage of certain features of the Intel server CPUs that are not available yet for AMD CPUs, e.g. direct transfers between the cache memory and peripherals, e.g. networking cards, without passing through the main memory, or better performance counters for tuning or debugging certain programs.
- Likes 2
Comment
-
Originally posted by AdrianBc View Post
While it is right that some time in a not too distant future it would become impossible to improve the performance per watt without switching to different semiconductor materials or to completely different electronic devices, we are not there yet.
AMD has 2 great advantages over Intel, one is the lower cost for more cores and more cache than Intel (due to the multiple chiplet design) and the much better performance per watt due to the TSMC 7-nm CMOS process.
If you would compare the specifications of any Intel and AMD current CPUs, you would see that at the same power consumption and at the same number of cores, the AMD CPUs always have a much higher base clock frequency.
The consequence of this fact is seen in all benchmarks. When an Intel CPU has most of the time no more than 1 or 2 active cores, so that the speed is limited by the maximum turbo core frequencies, then Intel may win the benchmark, but in all benchmarks where enough cores are active so that the power limits are reached, at the same number of cores and at the same power consumption the Intel CPUs are much slower than the AMD CPUs, because the average clock frequency during the benchmark is higher for the latter.
The static power was worst for the 90-nm Intel CPUs (Prescott/Nocona), where it could exceed a half of the total power consumption.
The next CMOS processes since 2006 had lower static power, due to innovations like using high-permittivity gate dielectrics, using FinFETs since Intel Haswell and designing the CPUs with a mixture of different kinds of transistors, some of which are high-speed transistors while others are low-leakage transistors.
While designing a CPU with a low static power has become quite a complex task in modern processes, it can be done, as it should be obvious because the current desktop computers have lower idle power values than ever and the current laptops have also better battery lifetimes than ever.
So the conclusion is that for now AMD has a considerable advantage over Intel at power efficiency. Just look that the best Intel laptop 8-core CPU has a pathetic 2.4 GHz base frequency, while its cheaper competitor from AMD has a 3.3 GHz base frequency and the frequency ratio is confirmed by all benchmarks (when they are done at the same power consumption; many published benchmarks are done with the Intel CPUs consuming up to a double power than AMD).
Cooper Lake is not competitive with Rome, because even with the double AVX-512 throughput it is cheaper to use twice as many AMD cores and also have lower power consumption and better connectivity.
Nevertheless, there are certain very specialized applications where Cooper Lake can be the best choice. While deep learning should have been better done on GPUs, professional NVIDIA GPUs are even more expensive for a given performance than Cooper Lake. AMD GPUs are much cheaper than both Cooper Lake and NVIDIA, but for most applications they require a much greater software development effort, due to lack of suitable libraries and tools, which might be prohibitive.
There are also other niche applications where Cooper Lake can be best, which take advantage of certain features of the Intel server CPUs that are not available yet for AMD CPUs, e.g. direct transfers between the cache memory and peripherals, e.g. networking cards, without passing through the main memory, or better performance counters for tuning or debugging certain programs.- static power consumption being considerably better now, with FinFET technology. Ideally, if you had a ratio of static vs dynamic power for current technology, when the processor operates at full speed. Quantum tunneling leakage is still insane when the feature sizes are ~10 nm.
- performance per Watt - we're not talking about just frequency here, but frequency * IPC per Watt (yes, I think our beloved Michael Larabel has some on Phoronix, but they're not easy to find).
Comment
-
Originally posted by vladpetric View Post
I appreciate your comment (seriously), could you kindly provide references to:- static power consumption being considerably better now, with FinFET technology. Ideally, if you had a ratio of static vs dynamic power for current technology, when the processor operates at full speed. Quantum tunneling leakage is still insane when the feature sizes are ~10 nm.
- performance per Watt - we're not talking about just frequency here, but frequency * IPC per Watt (yes, I think our beloved Michael Larabel has some on Phoronix, but they're not easy to find).
In any case, do prove me wrong here (with some references).
Comment
-
Originally posted by vladpetric View PostWhy do you care that it's only 14nm? And do you even know what that means when it comes to semiconductor technology?
- Likes 3
Comment
Comment