#cdna4
Zyphra has developed ZYFA, an innovative FP8 attention kernel for AMD's MI355X GPU, achieving 1.76x greater throughput over AITER kernels. By directly computing probabilities from scores and optimizing execution, ZYFA improves attention mechanisms in large models. https://arxiv.org/abs/2611.07496
ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4
ArXiv link for ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4
arxiv.org
November 7, 2026 at 11:30 AM
Zyphra has unveiled ZYFA, an FP8 attention kernel for AMD’s MI355X, achieving 1.16–1.76 times the throughput of previous kernels. By bypassing the costly softmax operation, ZYFA sets a new efficiency benchmark in long-context AI models. https://arxiv.org/abs/2611.07496
ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4
ArXiv link for ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4
arxiv.org
November 7, 2026 at 12:40 AM
Quentin Anthony, Ganesh Bikshandi, Praneeth Medepalli, Stephen Ebert: ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4 https://arxiv.org/abs/2611.07496 https://arxiv.org/pdf/2611.07496 https://arxiv.org/html/2611.07496
November 6, 2026 at 6:40 AM
rocm-gfx950-bin
ROCm Core SDK - CDNA4
aur.archlinux.org
October 4, 2026 at 8:53 PM
揭秘AMD全新MI350加速器:访谈首席架构师解析CDNA4架构创新

https://qian.cx/posts/0B9E4A50-E9E2-4872-9A1E-BE6A36F482C6
September 14, 2025 at 5:27 AM
Punjabi Folk Instrumental Music
YouTube video by SK World Punjabi
youtu.be
June 30, 2026 at 1:17 PM
Punjabi Folk Instrumental Music
YouTube video by SK World Punjabi
youtu.be
May 31, 2026 at 3:21 AM
youtu.be/cdnA4-CfJm4
Please like share and subscribe to my YouTube channel
Punjabi Folk Instrumental Music
YouTube video by SK World Punjabi
youtu.be
March 26, 2026 at 7:37 AM
AMD MI350P発表 既存ラックへCDNA4 AIを持ち込む

#upppp
https://upppp.jp/technology/20260511-145133/
upppp.jp
May 11, 2026 at 5:51 AM
이 와중 눈물의 똥꼬쇼를 펼치고 있지만 아직 좀 시원찮은 AMD.

거듭 말하는데, 눈물의 똥꼬쇼를 펼치고 있고 실제로 성과가 나오긴 나오고 있음. 한참 부족해서 그렇지.

전에도 말했지만 지금 대세는 4빗인데 AMD는 CDNA4(RDNA 아님)만 4빗을 지원함. RDNA5에선 지원한다는데 그건 27년에나 나옴.

그니까 엔비디아는 충분히 쉴 수 있는 것임. 죽어라 따라오려고 하는 경쟁상대는 올 해 할 수 있는 게 없지, 애플은 아무리 팔아도 엔비디아 따라잡기엔 범위가 다르지…

어떤 의미론 지금의 메모리 대란이 엔비디아에겐 호기임
- 2026년은 30년 만에 처음으로 새로운 엔비디아 게이밍 GPU가 출시되지 않는 해

- 엔비디아가 웨이퍼와 메모리 공급을 AI 데이터센터 칩으로 전환하면서 오랜 연간 출시 전통이 깨졌음

This is the first year in three decades without a new Nvidia gaming GPU

www.xda-developers.com/this-is-the-...
This is the first year in three decades without a new Nvidia gaming GPU
Is this the year we'll look back on as the point when everything changed?
www.xda-developers.com
February 26, 2026 at 9:25 AM
이는 ai에서 fp4가 현재 가장 고성능에 고효율 연산이기 때문입니다. 작고 빠르고 강력하며 정밀도도 높은, “최고 가성비” 라는거죠. 이보다 비트를 높이면 결과물은 좋아도 느려지고, 비트를 낮추면 빠르지만 결과물이 구려집니다.

이미 엔비디아는 nvfp4라는, fp4보다 더 강력한 독자적 규격을 사용합니다만, 이게 지원되는 것은 5000번대부터입니다. amd의 경우 민간용은 없고. cdna4가 fp4를 지원합니다. rdna는 5부터 지원 예정이고요.

즉 차세대 엑박/플스와 다음세대 라뎅부터 지원한단거죠.
February 13, 2026 at 4:48 PM
AMD MI350: Новое поколение ускорителей с архитектурой CDNA4 и инновационными технологиями

https://kripta.biz/posts/87598379-D89C-457C-ACA2-C3590C31F3EA
September 14, 2025 at 5:28 AM
Matrix Core Programming on AMD CDNA Architecture | Discussion
Matrix Core Programming on AMD CDNA™3 and CDNA™4 architecture
This blog post explains how to use Matrix Cores on CDNA3 and CDNA4 architecture, with a focus on low-precision data types such as FP16, FP8, and FP4
rocm.blogs.amd.com
November 30, 2025 at 5:00 AM
Occupancy Math on the AMD MI355X: A From-First-Principles Guide
Discussion | hackernews | Author: skidrow
Occupancy Math on the AMD MI355X: A From-First-Principles Guide
A from-first-principles guide to wavefront occupancy on AMD's MI355X (CDNA4): the hardware resource budget, the four limiters that cap it, worked MXFP8 GEMM examples, and why peak throughput often lives at low occupancy.
indianspeedster.github.io
June 21, 2026 at 5:02 PM
Peak Scale & New Architecture (CDNA4)
New architecture improvements, more modern process node etc., meaning better efficiency (though at very high power) and more “future proof” in features. Also, AMD claims big gains in throughput vs previous generation for both training and inference.
September 11, 2025 at 1:50 PM
=>
AMD "CDNA5" Instruction Set Architecture: Reference Guide, Jul 27, 2026 (832 pp) www.amd.com/content/dam/...

x.com/ogawa_tter/s...
CDNA3 Arch, Apr 2024
CDNA3 ISA,, Aug 5, 2025
CDNA4 Arch, Jun 2025
CDNA4 ISA, Aug 5, 2025 www.amd.com/content/dam/...

CDNA 5 Arch (24 pp) bsky.app/profile/ogaw...
August 1, 2026 at 2:55 PM
=>
Matrix Core Programming on AMD CDNA3 & CDNA4 Architecture, Sep 30, 2025, ROCm rocm.blogs.amd.com/software-too...

CDNA3
Arch, Apr 9, 2024 www.amd.com/content/dam/...
ISA, Aug 5, 2025 www.amd.com/content/dam/...
CDNA4
Arch, Jun 23 www.amd.com/content/dam/...
ISA, Aug 5 www.amd.com/content/dam/...
October 14, 2025 at 8:20 AM
AMD's new MI355X CDNA4 GPU is punching above its weight in MLPerf v6.0, delivering training times that rival the big players. Curious how it stacks up? Dive into the benchmark details. #AMDMI355X #CDNA4 #MLPerf6.0

🔗 aidailypost.com/news/amds-mi...
June 16, 2026 at 3:53 PM