3 ms·

For performance / $ in HPC systems this will be a field day though.

by m_mueller 10y ago

erik 10y ago
Most HPC systems are focused on floating point though, and Zen's vector/AVX implementation doesn't match Intel's.
- m_mueller 10y ago
  That only matters for applications that are fully vectorized however. Most real world HPC applications depend much more on memory bandwidth and/or multicore performance and how well the compilers can optimize for the vector units (typically not very well). If I can choose two cores with 128 bit vectors for the same money as one core with 256bit, I'll choose the two core system any day of the week.
  - erik 10y ago
    I certainly agree that vector performance is overrated. But it seems to dominate HPC spending. At least on the top500. But maybe AMD can pick up some HPC spending for smaller clusters designed to run less specialized code? I don't know how big that market is.
    - m_mueller 10y ago
      I do think it's big. What you see on top500 is the high end of the market, the big bulk is in small to medium sized clusters or in cloud infrastructure where people rather care for benchmarks like requests per second of some specific servers. There again memory throughput, non-vectorized performance and cores per dollar are probably the most relevant characteristics of a processor. And those who chase to have the highest flop numbers have switched to accelerators and custom chips anyways.
geezerjay 10y ago
For performance/$ in HPC systems, nowadays anyone can buy used Opteron 61XX processors for less than 20$/each, and even Opteron 62XX for around 50$.
- m_mueller 10y ago
  I think this is a bit too simplistic for HPC. Yes, you can go really low cost if all you care is theoretical performance / $. In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year). Even 20% higher memory bandwidth or computational performance per socket matters for the time to solution, since a higher number of nodes is uncertain to replace that due to the communication overhead. As does the performance per watt (costs of cooling). Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer. HBM means an order of magnitude higher memory bandwidth than what we're used to from CPUs, matching what the high end Nvidia Pascal cards are currently offering. By coupling this to a capable multicore X86 CPU with fat cores (as opposed to Knights landing with a higher number of slow cores), could mean tremendous speedups without any programming work for bandwidth bound applications - which includes pretty much any stencil application, e.g. atmospheric models, ocean models, earthquake prediction etc. By 'tremendous' I mean 5-10x per Socket, which is a big deal in these fields.
  - geezerjay 10y ago
    > Yes, you can go really low cost if all you care is theoretical performance / $. Yes that's a fact, and if you're trying to evaluate a solution based on performance/$ then the comparison is made with regards to this factor, and not others. > In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year). If you believe that criteria is relevant then you should know that Ryzen's advertised TDP is around 95W while some Opteron 61XX processors have a TDP of 85W. > Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer. For 40$ you can buy 4 opterons with 12 cores each. This means that for less than 1000$ you can put together a 48-core system that supports up to 512 GB of RAM. For around 500$, anyone can put together a 24-core system that also supports a couple of GPUs. Anyone that cares about HPC on a budget knows quite well that used Opterons is where the optimal price/performance ratio can be found, particularly as Opterons already out-perform Xeons in BLAS-based work.

For performance / $ in HPC systems this will be a field day though.

erik 10y ago

Most HPC systems are focused on floating point though, and Zen's vector/AVX implementation doesn't match Intel's.

m_mueller 10y ago
That only matters for applications that are fully vectorized however. Most real world HPC applications depend much more on memory bandwidth and/or multicore performance and how well the compilers can optimize for the vector units (typically not very well). If I can choose two cores with 128 bit vectors for the same money as one core with 256bit, I'll choose the two core system any day of the week.
- erik 10y ago
  I certainly agree that vector performance is overrated. But it seems to dominate HPC spending. At least on the top500. But maybe AMD can pick up some HPC spending for smaller clusters designed to run less specialized code? I don't know how big that market is.
  - m_mueller 10y ago
    I do think it's big. What you see on top500 is the high end of the market, the big bulk is in small to medium sized clusters or in cloud infrastructure where people rather care for benchmarks like requests per second of some specific servers. There again memory throughput, non-vectorized performance and cores per dollar are probably the most relevant characteristics of a processor. And those who chase to have the highest flop numbers have switched to accelerators and custom chips anyways.

geezerjay 10y ago

For performance/$ in HPC systems, nowadays anyone can buy used Opteron 61XX processors for less than 20$/each, and even Opteron 62XX for around 50$.

m_mueller 10y ago
I think this is a bit too simplistic for HPC. Yes, you can go really low cost if all you care is theoretical performance / $. In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year). Even 20% higher memory bandwidth or computational performance per socket matters for the time to solution, since a higher number of nodes is uncertain to replace that due to the communication overhead. As does the performance per watt (costs of cooling). Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer. HBM means an order of magnitude higher memory bandwidth than what we're used to from CPUs, matching what the high end Nvidia Pascal cards are currently offering. By coupling this to a capable multicore X86 CPU with fat cores (as opposed to Knights landing with a higher number of slow cores), could mean tremendous speedups without any programming work for bandwidth bound applications - which includes pretty much any stencil application, e.g. atmospheric models, ocean models, earthquake prediction etc. By 'tremendous' I mean 5-10x per Socket, which is a big deal in these fields.
- geezerjay 10y ago
  > Yes, you can go really low cost if all you care is theoretical performance / $. Yes that's a fact, and if you're trying to evaluate a solution based on performance/$ then the comparison is made with regards to this factor, and not others. > In practice, the cost efficiency of the system should be seen as 1 / (time to solution for typical applications) / (cost of ownership per year). If you believe that criteria is relevant then you should know that Ryzen's advertised TDP is around 95W while some Opteron 61XX processors have a TDP of 85W. > Furthermore, the biggest news for me regarding Ryzen is what they're doing with APUs. Affordable APUs with unified memory supporting HBM (stacked memory) could be a game changer. For 40$ you can buy 4 opterons with 12 cores each. This means that for less than 1000$ you can put together a 48-core system that supports up to 512 GB of RAM. For around 500$, anyone can put together a 24-core system that also supports a couple of GPUs. Anyone that cares about HPC on a budget knows quite well that used Opterons is where the optimal price/performance ratio can be found, particularly as Opterons already out-perform Xeons in BLAS-based work.