⚡ AOT startup: 1.43× faster with GraalVM Native Image, plus a Leyden AOT cache
🧩 Combined with native #GPU libraries from Java: #cuBLAS, #cuDNN and #MLX via the #TornadoVM API.
#Java #LLM #GPU👇
• 1.43× faster startup with @graalvm.org Native Image
• Up to 9.3× faster prefill
• Zero-copy weights on Apple silicon
Soon in @langchain4j.dev, @quarkus.io & @spring.io #SpringAI.
github.com/beehive-lab/jitllm
⚡ AOT startup: 1.43× faster with GraalVM Native Image, plus a Leyden AOT cache
🧩 Combined with native #GPU libraries from Java: #cuBLAS, #cuDNN and #MLX via the #TornadoVM API.
#Java #LLM #GPU👇
I've been developing a Python library inspired by FAISS, with its computational kernels implemented in Fortran. The CPU backend uses OpenMP, OpenBLAS, and thread autotuning. There's also a CUDA Fortran and cuBLAS backend for NVIDIA GPUs.
github.com/suwesh/fortr...
I've been developing a Python library inspired by FAISS, with its computational kernels implemented in Fortran. The CPU backend uses OpenMP, OpenBLAS, and thread autotuning. There's also a CUDA Fortran and cuBLAS backend for NVIDIA GPUs.
github.com/suwesh/fortr...
NVIDIA heeft in cuBLAS nieuwe algoritmen opgenomen voor matrixvermenigvuldigingen met dubbele precisie. De bibliotheek in CUDA Toolkit 13.4.2 gebruikt daarvoor integer Tensor Cores en behoudt vol…
Lees op techspeeltuin.nl
#technieuws #SoftwareEnOS
NVIDIA heeft in cuBLAS nieuwe algoritmen opgenomen voor matrixvermenigvuldigingen met dubbele precisie. De bibliotheek in CUDA Toolkit 13.4.2 gebruikt daarvoor integer Tensor Cores en behoudt vol…
Lees op techspeeltuin.nl
#technieuws #SoftwareEnOS
핵심 요약
• NVIDIA가 cuBLAS의 새 FP64 알고리즘 소개 글에서 CUDA Toolkit 13.4부터 들어간 Ozaki-II 방식과 2M 복소수 곱셈을 설명했습니다. FP64 행렬 곱(DGEMM, ZGEMM)을 정수 Tensor Core에서 계산하고 FP64와 같거나 더 나은 정확도를 목표로 합니다.
• NVIDIA 측정에서 DGEMM은 같은 GPU의 기본 FP64보…
핵심 요약
• NVIDIA가 cuBLAS의 새 FP64 알고리즘 소개 글에서 CUDA Toolkit 13.4부터 들어간 Ozaki-II 방식과 2M 복소수 곱셈을 설명했습니다. FP64 행렬 곱(DGEMM, ZGEMM)을 정수 Tensor Core에서 계산하고 FP64와 같거나 더 나은 정확도를 목표로 합니다.
• NVIDIA 측정에서 DGEMM은 같은 GPU의 기본 FP64보…
ai-news.ypetit.net/…
#AI #AIResearch #Biotech #NVIDIA
ai-news.ypetit.net/…
#AI #AIResearch #Biotech #NVIDIA
Join Dr. Tim Besard to learn how cuTile.jl 1.0 lets you write, verify and tune tensor-core kernels in Julia.
📅 Dec 9 | 12 PM EST
juliahub.com/events/cutil...
Join Dr. Tim Besard to learn how cuTile.jl 1.0 lets you write, verify and tune tensor-core kernels in Julia.
📅 Dec 9 | 12 PM EST
juliahub.com/events/cutil...
안녕하세요. Qwen3.8-Flash-Next를 로컬에서 실행하는 Strata를 포크해 Juud_engine을 개발하고 있어 공유합니다. 현재 공개한 소스 릴리스는 Strata v0.1.44.1 기반입니다.
추가한 기능
• Adaptive spin: CPU expert 배치가 끝난 뒤 워커가 불필요하게 대기 회전을 계속하는 시간을 줄입니다.
• 불필요한 CPU activation 양자화 생략: 단일 GPU의 혼합 다중 토큰 …
안녕하세요. Qwen3.8-Flash-Next를 로컬에서 실행하는 Strata를 포크해 Juud_engine을 개발하고 있어 공유합니다. 현재 공개한 소스 릴리스는 Strata v0.1.44.1 기반입니다.
추가한 기능
• Adaptive spin: CPU expert 배치가 끝난 뒤 워커가 불필요하게 대기 회전을 계속하는 시간을 줄입니다.
• 불필요한 CPU activation 양자화 생략: 단일 GPU의 혼합 다중 토큰 …
https://beefed.ai/en/choose-cublas-rocblas-vendor-blas
#CuBLAS #RocBLAS #VendorBlas #GpuBlasComparison #MultinodeGpu
https://beefed.ai/en/choose-cublas-rocblas-vendor-blas
#CuBLAS #RocBLAS #VendorBlas #GpuBlasComparison #MultinodeGpu
NVIDIA performance is a jar away from Java.
#Java #CUDA #TornadoVM
NVIDIA performance is a jar away from Java.
#Java #CUDA #TornadoVM
We also show demos with Apache Flink and @jitllm.io.
Watch our talk in @devoxx.com 👇
youtu.be/6SPnBmQrYEk?...
#opensource #AI #TornadoVM #Java #GPU
We also show demos with Apache Flink and @jitllm.io.
Watch our talk in @devoxx.com 👇
youtu.be/6SPnBmQrYEk?...
#opensource #AI #TornadoVM #Java #GPU
芝浦工大「Ozaki Scheme II」、NVIDIA CUDA 13.4に採用 低精度GPUで倍精度計算
novaist.jp/articles/nvi...
芝浦工大「Ozaki Scheme II」、NVIDIA CUDA 13.4に採用 低精度GPUで倍精度計算
novaist.jp/articles/nvi...
jitLLM uses NVIDIA CUDA features and libraries (cuBLAS, cuDNN) and reaches 0.90× to 1.12× of llama.cpp's throughput!
jitLLM is empowered by @tornadovm.org and will be soon in @langchain4j.dev and @quarkus.io.
github.com/beehive-lab/...
jitLLM uses NVIDIA CUDA features and libraries (cuBLAS, cuDNN) and reaches 0.90× to 1.12× of llama.cpp's throughput!
jitLLM is empowered by @tornadovm.org and will be soon in @langchain4j.dev and @quarkus.io.
github.com/beehive-lab/...
New TileContext API: Write #GPU kernels from #Java in tiles, not threads, via #NVIDIA #cuTile model.
Now, cuTile methods can be mixed with cuBLAS functions in one TaskGraph, optimizing data transfers.
👉 github.com/beehive-lab/TornadoVM/releases/tag/v7.0.0
#opensource #AI
New TileContext API: Write #GPU kernels from #Java in tiles, not threads, via #NVIDIA #cuTile model.
Now, cuTile methods can be mixed with cuBLAS functions in one TaskGraph, optimizing data transfers.
👉 github.com/beehive-lab/TornadoVM/releases/tag/v7.0.0
#opensource #AI
geni.us/NVIDIA-cuBLAS
geni.us/NVIDIA-cuBLAS
One GPU validated so far, early days. I keep watching people build real bridges out of the CUDA moat.
https://github.com/Speedstu/CUDA-for-AMD-Windows
#CUDA
One GPU validated so far, early days. I keep watching people build real bridges out of the CUDA moat.
https://github.com/Speedstu/CUDA-for-AMD-Windows
#CUDA