Zajimave. I u nas ve firme vidim velky narust VDI a dalsi sta az tisice se planuji. Sice ne vse bude potrebovat GPU akceleraci, ale i takove VDI se tu provozuji.
MSI MPG GUNGNIR 110R White | CPU AMD Ryzen 7 9700X Granite Ridge | DeepCool AK500 White | GPU Sapphire Pure RX 9070 XT 16GB plus UV | MB MSI MAG X670E GAMING PLUS WIFI | 32GB DDR5 A-DATA XPG LANCER RGB Dual KIT 7200 MHz | system HDD SSD M.2 Kingston FURY Renegade NVMe 1TB | Seagate Baracuda HDD 1TB SATA III | data HDD WD RED 1TB SATA III | Quad HD VA monitor 27" MSI Optix G27CQ4 Free Sync 165 Hz 10bit HDR | Soud Blaster Audigy Fx | PSU MSI MAG A850GL 850 W 80 PLUS Gold PCIe 5 II | Win 10-64 bit Pro
AMD tímto vydala Radeony PRO WX 4100, WX 5100 a WX 7100. Na tom by nebylo nic zvláštního, přislíbila to již v červnu. Jenže finální specifikace jsou výrazně lepší než původní červnové…
Wow, 99,6% CUDA kódu přeloženo automaticky, slušný
Using the HIPify Tools for automatic conversion, AMD was able to translate 99.6% of the code automatically. This is admittedly less than 100% – AMD has been upfront before that they may always be playing catch-up and won’t necessarily be able to replicate NVIDIA driver features 1-for-1 – but it’s significant progress. Ultimately AMD was able to finish the feature-complete port in less than 4 days for a 55K line project. And words can’t stress how important this is for AMD
AMD získalo kontrakty a dodává VDI FirePro S karty pro cloudová řešení společností Google (FirePro S9300 x2 - Fiji) a Alibaba (FirePro S7150 x2 - Tonga)
Příklad využítí "Boltzmann Initiative" v praxi - Naportování CAFFE Machine learning frameworku z CUDA na C++ pomocí HIP
99.6% success
Overall the CAFFE port was quite successful – 99.6% of the code was either automatically ported or was automatically converted. A single developer required less than a week to clean up the remaining 200 lines of code to produce a working version on ROCm. Previously we have tried to use OpenCL C-based API to port similar C++ applications. Porting C++ to C is quite a challenging problem, and the OpenCL CAFFE port in fact required tens of thousands of additional lines of code. Much of this code duplicates the CUDA kernels and runtime, and so significantly increases the maintenance cost. In comparison, the HIP port is a small delta from the original code, the conversion is largely automated, and as we will see can replace the CUDA port. The HIP port also includes all the features of the original application, including the multi-GPU, optimized peer-to-peer communication, and CUDNN support.
Tom's Hardware - OpenCL And CUDA Are Go: GeForce GTX Titan, Tested In Pro Apps
