GPU servers.
Without limits
for your ideas.

GLM-5.3

753 B — checkpoint parameters. 39 B — active per token.

Published checkpoint: FP8 / BF16 / F32. Weight payload from the published checkpoint or official deployment guide. This is stored payload, not measured GPU allocation; runtime conversion can increase it. Operating reserve is added separately.

Weight format: Start with the published checkpoint. Extra quantization may affect quality; converting quantized weights to BF16 does not restore lost precision.

Operating reserve: Memory beyond weights for KV cache, activations, runtime buffers and headroom across the selected configuration, not per GPU. Set it from context length and concurrent requests, then measure peak use. The presets are examples, not universal recommendations.

This is the limit declared in the model configuration, not a context verified on the selected hardware.

DSA uses kv_lora_rank=512 and full/shared indexer types. Cache sizing requires an architecture-specific calculation.

Context config checked: 06.10.2026.

Runtime validation required. Published deployment guide: HGX H200 · 8 GPU, Instinct MI355X · 8 GPU, Instinct MI325X · 8 GPU, Instinct MI300X · 8 GPU.

How to read this estimate

Weights = total parameters × bytes per parameter: BF16 2, FP8 1, INT4 0.5. Decimal GB. MoE models count every expert, not only active parameters. This is a lower bound: quantization scales, unquantized layers and metadata are not separately modelled; they must fit within the reserve.

A green result is a capacity estimate, not a tested deployment. GPU memory is distributed across devices. Tensor and expert parallelism, interconnects, model kernels and framework support must all be checked. Long context and parallel users can need much more reserve. Fine-tuning and training require a separate calculation.

KV cache depends on context length, concurrent requests, cache precision and the model’s attention architecture. Sliding-window, linear-attention and compressed-cache models need architecture-specific estimates; validate the peak in the selected runtime. This field changes the capacity estimate, not server settings.

• Capacity fits. The estimated memory budget fits. Validate software, sharding and performance.
• Limited headroom. Under 10% free at this reserve. Validate memory use before choosing.
• Exceeds capacity. This scenario exceeds total GPU memory. Change precision or platform.

BF16 / FP8 / INT4 are size scenarios. The format names do not assert native execution support or equal model quality.

NVL topology and sharding require separate validation; an 8-GPU PCIe server is not the same fabric as HGX/NVSwitch. Validate the exact checkpoint and kernels in ROCm before deployment.

Huawei Ascend uses CANN. Capacity results do not confirm that the selected checkpoint, operators or precision run on these stacks. A model port and runtime validation are required.

The calculation uses 270 GB per GPU for HGX B300 and 279 GB per GPU for GB300, following the NVIDIA datasheet. The 288 GB figure is not used. CPU RAM is not added to GPU memory. NVIDIA Blackwell Ultra (PDF) ↗