In a YouTube teardown, the author takes apart the Huawei Atlas 300I Duo — a server AI accelerator Huawei built for China's domestic market. Two Ascend chips sit on a single slim board [00:09]. Together they carry 96 GB of memory, 48 GB per chip [00:17]. And it's not HBM — it's LPDDR4X, the same stuff you'd find in a laptop. That's why the card draws just 150 W [00:28] and fits in a single-slot chassis [00:35].
The author puts it head-to-head with the H100 PCIe: NVIDIA's card has a single Hopper GPU, 80–96 GB of HBM2e/HBM3, a TDP of around 350 W [00:42], and a two-slot footprint [00:49]. The H100 also costs roughly 10 times more [02:04]. According to the video, the Atlas 300I Duo goes for up to 300,000 rubles [00:00].
The media engine is a different story entirely. The card decodes up to 256 Full HD streams at 30 FPS simultaneously and chews through 4K JPEG at up to 1,024 frames per second [00:49, 01:21]. In China, it gets deployed for smart city and facial recognition systems, video moderation, recommendation engines, and clinics analyzing medical scans [01:05]. LLMs and image generation work too: the Ascend chips are supported by vLLM and ComfyUI [01:30].
The software, however, is closed. Instead of CUDA, you get CANN, Huawei's in-house equivalent. Windows isn't supported at all [01:58], and the plug-and-play fantasy dies right there [01:58]. After installing the card in an x16 slot, lspci | grep -i d100 should show two Huawei Device d100 devices (or d500, depending on the revision). Next, you create a system user HwHiAiUser with uid/gid 1000 and a home directory at /home/HwHiAiUser.
Then you download two.run files matched strictly to your architecture, x86_64 or aarch64: sudo ./Ascend-hdk-310p-npu-driver_*.run --full --install-for-all, followed by the firmware and a reboot. npu-smi info should show two chips with an OK status, temperature readings, and ~48 GB of memory each. On top of that, install the CANN Toolkit (e.g., 8.0.RC1) and add source /usr/local/Ascend/ascend-toolkit/set_env.sh. PyTorch only works paired with torch-npu: pip install torch==2.1.0 torch-npu==2.1.0.post3 — and the versions must match the CANN compatibility matrix.
Cooling is passive and server-grade. In a regular case, you'll need a blower fan pointed at the heatsink, or the chips will start throttling within minutes. The step-by-step guide with all the commands is in our Atlas 300I Duo setup guide.
The card isn't new — the news is the teardown itself. And here's why it took off, in our view: people don't need an "H100 replacement," they need cheap memory for inference. A model that won't fit into the 24 GB of a typical GPU gets 48 GB per chip here. At that price point, that's a rare option.
The "H100 killer at 1/10th the price" headline is clickbait, and here's our math. The 10x price gap applies to the board, not the cost per token: the H100 has a different memory class and mature software, while the Atlas has LPDDR4X and a CANN stack you still have to wrangle into working. Comparing TDP and price without showing tokens per second is comparing boxes, not computers.
The second wrinkle the video glosses over: the 96 GB isn't a single pool — it's two chips with 48 GB each. A model has to fit on one chip, or you'll have to slice it to fit CANN's parallelism scheme. This isn't "one big 96 GB GPU" — it's two accelerators on one board, and you need to plan your memory accordingly (our read on the card's architecture).
The media engine is the one area where the card doesn't just catch up — it pulls ahead: 256 simultaneous Full HD streams is a video-surveillance-and-moderation profile, not a chatbot one. That's exactly what it's bought for in China, and exactly where it pays for itself without a single LLM benchmark.
The winners are people comfortable with Linux and willing to fight with drivers. The losers are anyone hoping for a "buy and run" button: the docs and firmware are built for China [01:44], and hunting down the right versions is a quest of its own. What gives me pause here isn't the hardware — it's the support outside the domestic market.
The video contains not a single tokens-per-second benchmark — just specs and a TDP-versus-price comparison. "Up to 300,000 rubles" is a ceiling; where that figure came from and where the card is actually sold, the source doesn't say. One voice, one video, zero independent measurements. Two things are worth verifying yourself: real memory bandwidth in your actual workload, and whether drivers exist for your board revision.
Forecast: within the quarter, we'll see more Atlas 300I Duo breakdowns with real tokens-per-second numbers than exist today, but the card won't become a mainstream choice outside China — supply and documentation will see to that.
Disclaimer: single source — a YouTube video; official specs and pricing are unconfirmed.
Source: YouTube
Source: www.youtube.com · post in Telegram