From: already5chosen@yahoo.com   
      
   On Wed, 26 Mar 2025 23:53:14 -0000 (UTC)   
   Lawrence D'Oliveiro  wrote:   
      
   > On Wed, 26 Mar 2025 23:45:54 +0200, Michael S wrote:   
   >    
   > > Top performing Arm CPUs are designed in USA.   
   > > 1. Cupertino, California (Apple)   
   > > 2. Raleigh, North Carolina (Qualcomm)   
   > > 3. Austin, Texas (Arm Inc)     
   >    
   > All outclassed by Fujitsu’s A64FX, as used in the Fugaku machine, and    
   > successors -- designed in Japan.   
      
   That's very OT in comp.os.vms, but I'd write it anyway. A64FX did not   
   outclass any of mentioned above.   
      
   A64FX is good chips for its intended applications, i.e. massively   
   parallel supercomputers. Its most important feature is Tofu comm links   
   that together with Tofu switches allow high-bandwidth communication   
   between tens of thousands of nodes with the latency that is not exactly   
   low, but can be called acceptable. It's second important feature is   
   built-in connection to HBM2 memory. Its third important feature is   
   rather good floating-point throughput when the application amenable to   
   vectorizing. Its fourth important feature is that said good throughput   
   is achieved without consuming too much power.    
   However the 4th point should be taken in time context. Consumed power   
   per FLOP was lower relatively to A64FX 2019 contemporaries. It is   
   significantly higher relatively to the next generation of GPGPUs. It is   
   true even if we only look at single and double precision calculations.   
   For lower precision calculations that are so popular nowadays and   
   becoming more popular with every passing minute, A64FX is hopelessly   
   behind state of the art.   
      
   Those were strong points of A64FX. Its weak points that make it a poor   
   choice for general purpose computing are:   
   (a) - integer performance. No benchmarks published, but description in   
   various papers could lead as to guess that in benchmarks like   
   SPECInt2017_speed A64FX core would be be at least 3 time slower than   
   top offerings of three design houses mentioned above.   
   (b) - memory capacity. Only 32 GB per 48-core node. For general-purpose   
   server application it's desirable to have at least 5 times more,   
   preferably 10 times more.   
      
   --- SoupGate-DOS v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)