#tflops — Public Fediverse posts

Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s

https://www.cocoawithlove.com/blog/matrix-multiplications-swift.html

#HackerNews #Swift #LLM #matrix #multiplication #Gflops #Tflops #machinelearning

New benchmark shows that larger CUDA tiles can cut Flash Attention throughput by 18‑43 % across sequence lengths. The study dives into kernel design, TFLOPS loss, and what it means for transformer model efficiency on NVIDIA GPUs. Open‑source researchers can use these insights to tune their kernels and reclaim performance. #FlashAttention #CUDATiles #GPUPerformance #TFLOPS

🔗 https://aidailypost.com/news/large-cuda-tiles-reduce-flash-attention-tflops-by-1843-across

New benchmark shows that larger CUDA tiles can cut Flash Attention throughput by 18‑43 % across sequence lengths. The study dives into kernel design, TFLOPS loss, and what it means for transformer model efficiency on NVIDIA GPUs. Open‑source researchers can use these insights to tune their kernels and reclaim performance. #FlashAttention #CUDATiles #GPUPerformance #TFLOPS

🔗 https://aidailypost.com/news/large-cuda-tiles-reduce-flash-attention-tflops-by-1843-across

New benchmark shows that larger CUDA tiles can cut Flash Attention throughput by 18‑43 % across sequence lengths. The study dives into kernel design, TFLOPS loss, and what it means for transformer model efficiency on NVIDIA GPUs. Open‑source researchers can use these insights to tune their kernels and reclaim performance. #FlashAttention #CUDATiles #GPUPerformance #TFLOPS

🔗 https://aidailypost.com/news/large-cuda-tiles-reduce-flash-attention-tflops-by-1843-across

Fp8 runs ~100 tflops faster when the kernel name has "cutlass" in it

https://github.com/triton-lang/triton/pull/7298

#HackerNews #Fp8 #cutlass #tflops #performance #optimization #HackerNews #triton

😂 Ah, the classic tale of tech sorcery where simply naming your kernel "cutlass" magically unlocks 100 #tflops of speed! Meanwhile, x.com is still busy booting you off your browser faster than you can say "incompatibility." 🏴‍☠️🔗📉
https://twitter.com/cis_female/status/1943069934332055912 #techhumor #cutlass #xcom #incompatibility #HackerNews #ngated

FP8 is ~100 tflops faster when the kernel name has "cutlass" in it

https://twitter.com/cis_female/status/1943069934332055912

#HackerNews #FP8 #tflops #cutlass #performance #optimization #AI

#China's secretive #Tianh 3 #supercomputer uses homegrown hybrid #CPU — rivals US systems with 1.57 #Exaflops of performance. #NUDT #MT3000 features a unique heterogeneous architecture that includes general-purpose CPU cores with 96 control cores and 1,536 accelerator cores. Netting the MT-3000 processor reportedly achieves 11.6 FP64 #TFLOPS of peak performance and demonstrates a power efficiency of 45.4 #GigaFLOPS/Watt at operational frequency of 1.20 GHz https://www.tomshardware.com/tech-industry/supercomputers/chinas-secretive-tianhe-3-supercomputer-uses-homegrown-hybrid-cpu-rivals-us-systems-with-157-exaflops-of-performance-report #hpc #sanctions