AMD unveils new AI chip to compete with NVIDIA Blackwell
What's the story
AMD has unveiled its latest artificial intelligence (AI) chip, the Instinct MI325X.
The new offering will take on NVIDIA's data center graphics processors (GPUs) head-on.
The Instinct MI325X will go into production before 2024 ends, and is expected to be used by Dell, HP, Lenovo, etc. starting in Q1 2025.
With this move, AMD could shake NVIDIA's pricing strategy as the latter has been enjoying some 75% gross margins due to high demand for its GPUs over the last year.
Comparison
AMD's Instinct MI325X vs NVIDIA's H200
AMD claims its 256GB Instinct MI325X GPU outperforms NVIDIA's 141GB H200 processor in AI inference workloads.
Powered by the CDNA 3 architecture, the MI325X offers 256GB of HBM3E with 6.0TB/s bandwidth, providing 1.8x more capacity and 1.3x more bandwidth than the H200.
It also delivers 1.3x higher peak FP16 and FP8 compute performance.
These advantages translate into up to 1.3x better inference performance on Mistral 7B, 1.2x on Llama 3.1 70B, and 1.4x on Mixtral 8x7B models.
Market potential
Instinct MI325X targets $500 billion market
The launch of Instinct MI325X is AMD's move to take on NVIDIA's reign in the data center GPU market.
The company hopes to either dethrone or capture a large chunk of this market, which is expected to be worth $500 billion by 2028.
"AI demand has actually continued to take off and actually exceed expectations," AMD CEO Lisa Su said at the product announcement event.
AMD's stock has only gained 20% so far in 2024 against NVIDIA's stellar 175% jump.
Product roadmap
AMD accelerates product schedule to compete with NVIDIA
With the MI325X, AMD is accelerating its product release cycle to introduce new chips every year.
The move is intended to boost competition with NVIDIA and take advantage of the AI chip boom. The newly launched AI chip replaces the MI300X, which started shipping late last year.
Besides the MI350 due in 2025, the upcoming releases include MI400 in 2026.
Software improvement
Enhanced software to attract AI developers
In response to NVIDIA's proprietary programming language, CUDA, AMD has been improving its competing software, ROCm.
The move is designed to lure AI developers into moving more of their AI models to AMD's chips, aka accelerators.
The company markets its AI accelerators as more competitive for use cases where AI models are generating content or making predictions instead of processing huge amounts of data.
Performance edge
AMD's AI accelerators outperform some NVIDIA chips
AMD's AI accelerators have proven to be more efficient in some applications.
For example, they can serve Meta's Llama AI model faster than some NVIDIA chips, thanks to their advanced memory.
"What you see is that MI325 platform delivers up to 40% more inference performance than the H200 on Llama 3.1," said Su, referring to Meta's large language AI model.
Sales growth
Data center sales more than doubled in the past year
AMD's data center sales during the June quarter more than doubled in the past year to $2.8 billion, with AI chips accounting for only about $1 billion, the company said in July.
AMD takes about 34% of total dollars spent on data center CPUs, the company said.
That's still less than Intel, which remains the boss of the market with its Xeon line of chips.