Quite so! This could be quite useful to help navigate the environmental regulation rodeo as explained by BASF when presenting their new 3 PetaFlop/s Quriosity — “the world’s largest supercomputer used in industrial chemical research”. Those reconfigurable FP64 NextSilicon-branded Maverick-2 chips could get them to assess the “Potential impact of crop protection products on groundwater quality” in even less than a few hours (instead of years), or through more complex environment models: https://www.basf.com/global/en/who-we-are/innovation/how-we-innovate/our-RnD/Digitalization_in_R-D/supercomputer

I reckon it’d be quite a sight to see the rancheros at Corteva (Dow/Elanco-Dupont), Bayer-Monsanto, and Syngenta, down the Arbuckle and saddle-up to join this here computational steer wrestling contest, for everyone’s improved environmental protection, with great crop production!

Also, thanks for linking to “A recent study from Oak Ridge National Laboratory” where Table 3 shows that a 10x speedup with MxP only occurs (so far at least) for dense matrix situations, while the more relevant sparse matrix computations may see just 1.5x (eg. CFDand Climate/Weather in their Table 2, and I expect contaminant transport and environmental flows, unfortunately).

Maverick-2’s mill cores that automatically identify computational hotspots and optimize the corresponding compute graphs (reducing data movement overhead among others) sounds like quite the ticket imho. I like the 4x better perf-per-watt vs GPUs and hope it gets realized in sparse computations (eg. both direct and iterative solvers). In particular, I expect the distributed HBM3E approach, with dynamic reconfiguration of compute, to help out with the HPCG and Graph500 memory access challenges (here in HPC-relevant FP64, rather than the more common FP32 of dataflow devices).

Looking forward to seeing some conference presentations or papers related to Sandia’s NNSA Vanguard Penguin Tundra testing of this quite promising ASIC ( https://www.sandia.gov/research/news/sandia-partners-with-nextsilicon-and-penguin-solutions-to-deliver-first-of-its-kind-runtime-reconfigurable-accelerator-technology/ )!

A cloud service should be created that allows potential customers to easily upload source code and evaluate the performance of Maverick-2. The runtime limit could be 60 seconds. The uploaded source code should be prevented from accessing the internet or the local file system. The uploaded source code should be able to print a maximum of 5000 characters. Potential customers would submit a job to a queue and see the results on the website when the job completes. No registration or email address should be required.

The NextSilicon website should have detailed documentation describing the microarchitecture of Maverick-2, especially the memory subsystem. The latency, bandwidth and size of each level in the memory hierarchy should be listed. Techniques for optimizing the performance of Maverick-2 should be described. When Maverick-2 becomes available for sale, there should be links on the NextSilicon website to webpages where the product can be bought.

NextSilicon should work with HPC software providers, like Q-Chem, to get their applications running on Maverick-2. The NextSilicon website should show the performance of these real-world HPC applications on Maverick-2 and alternative platforms. Maverick-2 should provide at least a 3x better price/performance ratio on real-world HPC applications than the best available NVIDIA GPUs and x86 CPUs.

Where are the benchmarks, feeds and speeds?

For example it would be interesting to see to what extent, if any, SPECfp benefits from Maverick-2 acceleration.