Skip to content
RobotWorld
← Back to Videos

DeepGEMM-Ascend Brings DeepSeek Kernels to Huawei NPUs

Loading video
Loading video

DeepSeek ports its DeepGEMM kernel library to Huawei Ascend 950 NPUs, with BF16, FP8, FP4 GEMM, MQA and MegaMoE reaching up to 99.8% dense-GEMM hardware utilization.

DeepGEMMDeepGEMM-AscendDeepSeekHuawei AscendNPULLM InferenceGEMMMoEOpen Source
Category: llm
Author: @EngMoElgaraihy
Date: 2026-10-05T00:00:00
Duration: 15.0s
Source: https://x.com/EngMoElgaraihy/status/2106934930106417544
Reference: https://github.com/deepseek-ai/DeepGEMM-Ascend

Related Videos

"Welcome to the era of Superintelligence": Fable 5.5 apparently goes live — devs report auto-routing on claude.ai, and a Superman continuous-animation demo in multiple art styles
0:40

"Welcome to the era of Superintelligence": Fable 5.5 apparently goes live — devs report auto-routing on claude.ai, and a Superman continuous-animation demo in multiple art styles

@chetaslua · 220 views · 2026-10-01
LLMAnthropicClaude
AutoTrust open-sources JEV-27B-VL: the first open-weight near-SOTA multimodal decision model — it watches Mario and decides whether to jump
0:25

AutoTrust open-sources JEV-27B-VL: the first open-weight near-SOTA multimodal decision model — it watches Mario and decides whether to jump

@AutoTrustAI · 353 views · 2026-10-01
LLMJEV-27B-VLAutoTrust AI
UniMate Open-Sources Text-to-Animation for Any Rigged Skeleton
0:31

UniMate Open-Sources Text-to-Animation for Any Rigged Skeleton

@Stefan_3D_AI · 346 views · 2026-09-29
Text-to-Motion3D AnimationOpen Source
OmniChar: One Character Across Every Model, Carried in a Portable .char File
0:23

OmniChar: One Character Across Every Model, Carried in a Portable .char File

@aiaicreate · 672 views · 2026-09-28
OmniCharCharacter ConsistencyComfyUI
ROBOTIS Mission Canvas: Composing Reusable Robot Skills into Repeatable Mobile Missions
1:49

ROBOTIS Mission Canvas: Composing Reusable Robot Skills into Repeatable Mobile Missions

@passionvirus · 475 views · 2026-09-27
ROBOTISMission CanvasAction Canvas
Open-RAIL: China Mobile's Real-Time Execution Layer Between VLA Models and Humanoids
1:49

Open-RAIL: China Mobile's Real-Time Execution Layer Between VLA Models and Humanoids

@CyberRobooo · 658 views · 2026-09-23
VLAHumanoidReal-time control

As an Amazon Associate, we earn from qualifying purchases.