Chuyển đến nội dung
MZGrid

Trang này chỉ có bằng tiếng Anh. Phiên bản tiếng Anh là bản có giá trị chính thức.

GPU catalogue

The supported GPU models, their reference prices in MZG per GPU-hour, how live prices relate to them, and how to choose a GPU for your workload.

MZGrid supports thirteen NVIDIA GPU models. Each is identified by a stable sku used in the API, the SDK and provider pricing.

SKUGPUVRAMKiến trúcFP16 TFLOPSGiá tham chiếu MZG/h
b200NVIDIA B200192 GBBlackwell2.2501.500
h200NVIDIA H200 SXM141 GBHopper9891.000
h100-sxmNVIDIA H100 SXM80 GBHopper989750
h100-pcieNVIDIA H100 PCIe80 GBHopper756640
a100-80NVIDIA A100 SXM 80GB80 GBAmpere312460
a100-40NVIDIA A100 PCIe 40GB40 GBAmpere312320
l40sNVIDIA L40S48 GBAda362290
rtx6000-adaNVIDIA RTX 6000 Ada48 GBAda364240
rtx5090GeForce RTX 509032 GBBlackwell210220
a6000NVIDIA RTX A600048 GBAmpere155160
rtx4090GeForce RTX 409024 GBAda165150
rtx4080sGeForce RTX 4080 SUPER16 GBAda10490
rtx3090GeForce RTX 309024 GBAmpere7175

Reference prices vs. live prices

The reference price is a catalogue anchor, not what you pay. Every provider sets its own price per SKU, and the marketplace shows live offers, which typically sit between 0.85× and 1.20× the reference:

SKUReference MZG/hTypical live range MZG/h
h100-sxm750637.50 – 900
a100-80460391 – 552
l40s290246.50 – 348
rtx4090150127.50 – 180
rtx30907563.75 – 90

Offers at the lower end usually come from newer providers with less reputation or from community-tier machines; datacenter-tier providers tend to price toward the upper end. Current lowest and median prices per SKU are returned by GET /v1/catalog/gpus.

The FP16 reference column

The FP16 dense TFLOPS figure is the benchmark reference used for verification. A provider's measured throughput must fall within 70–115% of this value for the machine to be listed as verified. It is also a useful rough guide to relative compute speed between SKUs.

Choosing a GPU

VRAM first

VRAM decides what fits. As a rule of thumb, model weights need about 2 GB per billion parameters in BF16/FP16, 1 GB in FP8/INT8 and 0.5–0.6 GB with 4-bit quantisation (AWQ/GPTQ). Add headroom for the KV cache, which grows with context length and the number of concurrent requests — plan for at least 20–30% on top of the weights for serving.

8B models

An 8B model in BF16 needs about 16 GB for weights. A 24 GB card (rtx4090, rtx3090) runs it comfortably with a moderate context; rtx5090 (32 GB) or l40s (48 GB) give room for long contexts and many concurrent requests. The RTX 4090 usually offers the best throughput per MZG; the RTX 3090 is the cheapest option.

32B models

A 32B model in 4-bit AWQ needs roughly 18–20 GB, so it fits on 24 GB cards with a short context (qwen2.5-32b template). For production traffic or longer contexts, a 48 GB card (l40s, rtx6000-ada, a6000) is the sweet spot. In BF16 a 32B model needs about 64 GB of weights — use an a100-80 or h100.

70B models

A 70B model in FP8 needs about 70 GB, so use an 80 GB or larger Hopper/Blackwell GPU: h100-sxm, h100-pcie, h200 (141 GB, much more KV-cache room) or b200 (192 GB). This is what the llama-3.3-70b template targets. Hopper and Blackwell run FP8 natively, which is why they are preferred over a100-80 for this template.

Image generation

SDXL and FLUX.1-schnell run well on 16–24 GB cards. The rtx4090 is the usual choice for interactive ComfyUI work; rtx4080s is a cheaper option for SDXL at standard resolutions; l40s or rtx5090 are better for large batches or high resolutions.

Speech and notebooks

Whisper large-v3 needs about 10 GB — an rtx4080s or rtx3090 is plenty. JupyterLab runs on anything; choose the GPU that matches the experiments you plan to run.

Summary

WorkloadMinimumGood valueHigh throughput
7–8B LLM (BF16)24 GBrtx4090, rtx3090l40s, h100-pcie
32B LLM (AWQ)24 GBrtx4090, rtx5090l40s, rtx6000-ada
70B LLM (FP8)80 GBh100-pcieh100-sxm, h200, b200
SDXL / FLUX.1-schnell16 GBrtx4080s, rtx4090rtx5090, l40s
Whisper large-v310 GBrtx4080s, rtx3090rtx4090
Fine-tuning / researchdependsa100-80h100-sxm, h200