AMD's MI455X delivers 432GB HBM4, enough for trillion‑parameter model weights in one server
In a July 24 X post, AMD disclosed a new MI455X GPU that packs 432GB of HBM4, enough to store FP8 weights of a trillion‑parameter model in a single server.
By Ryan Merket · Published
Why it matters
The MI455X’s 432 GB HBM4 memory could let a single server hold the FP8 weights of a trillion‑parameter model, potentially streamlining the deployment of ultra‑large AI systems and challenging Nvidia’s dominance in high‑memory GPUs.
Lysandre, an X user identified as Lysandre (@LysandreJik), announced on July 24 that AMD has released a new GPU, the MI455X, equipped with 432 GB of HBM4 memory. The post noted that four of these GPUs in a single chassis would deliver 1.7 TB of high‑bandwidth memory – a figure that exceeds the combined memory of five Nvidia H100 GPUs, each stocked with 80 GB of HBM2e.
"AMD just put more memory on one GPU than 5 H100s combined. The new MI455X has 432GB of HBM4. Put 4 in one server and you get 1.7TB," Lysandre wrote on X. The tweet also referenced early‑access testing and linked to a screenshot of the hardware spec sheet.
The claim marks a significant jump in GPU‑on‑board memory. HBM4, the latest generation of high‑bandwidth memory, promises higher capacity per stack and improved power efficiency over HBM2e. While Nvidia’s flagship H100 remains limited to 80 GB, AMD’s 432 GB per GPU represents a more than five‑fold increase, positioning the MI455X as a potential game‑changer for workloads that are memory‑bound.
The post further tied the memory capacity to the emerging class of trillion‑parameter AI models. FP8 quantization – a technique that compresses model weights to eight‑bit floating point – allows such massive models to fit in less memory than full‑precision formats. Lysandre calculated that a trillion‑parameter model such as the rumored Kimi K2 (or K3) could have its entire FP8 weight set stored on a single machine equipped with four MI455X GPUs, totaling 1.7 TB of HBM4.
If the calculation holds, the MI455X could simplify the deployment of extremely large foundation models. Today, training or inference of trillion‑parameter models typically requires multi‑node clusters linked by high‑speed interconnects, with complex model‑parallel pipelines to shuffle tensors across devices. By concentrating the entire weight tensor in a single server’s memory pool, developers could reduce latency, lower communication overhead, and possibly cut operational costs.
AMD has not issued an official press release at the time of writing, and the details come solely from Lysandre’s early‑access experience. The tweet includes a link to a PDF screenshot (https://t.co/pS5AmJON6W) that appears to be an internal spec sheet, but the contents have not been independently verified by AMD or third‑party benchmarks.
Industry observers note that the GPU memory race reflects a broader shift in AI hardware toward higher capacity, as model sizes continue to grow. Nvidia’s recent roadmap emphasizes faster interconnects (NVLink, Hopper) and software optimizations, but its memory ceiling remains at 80 GB per GPU for the H100 line. AMD’s move to HBM4 could force Nvidia to accelerate its own next‑gen memory roadmap or pursue alternative strategies such as memory‑offload pipelines.
Analysts also caution that raw memory capacity is only one axis of performance. Compute density, memory bandwidth, and software stack maturity each play a role in real‑world AI workloads. AMD’s MI series has historically lagged Nvidia in terms of CUDA ecosystem support, but the company has been expanding its ROCm stack and partnering with major cloud providers to improve adoption.
The early‑access claim suggests that AMD may be positioning the MI455X for enterprise AI customers who need to run exceptionally large inference workloads without the complexity of multi‑node setups. Cloud providers could also offer a “single‑node trillion‑parameter” tier that leverages the 1.7 TB memory pool, potentially differentiating their offerings from competitors that rely on GPU farms.
Until AMD releases an official specification sheet and performance benchmarks, the exact performance characteristics of the MI455X remain uncertain. Nonetheless, the announced 432 GB HBM4 capacity signals a clear strategic bet: AMD is aiming to capture a niche of memory‑intensive AI workloads that current GPU offerings struggle to accommodate.
For now, the industry will watch how AMD’s claim holds up under independent testing and whether the memory advantage translates into tangible cost and performance benefits for developers building next‑generation foundation models.