AI Server GPUs Compared: VRAM, Bandwidth, and Power per Accelerator

Server-grade NVIDIA and AMD accelerators with a computed matrix of how many GPUs each open model needs. Every spec comes from the vendor datasheet linked in the table.

Which server GPU has the most memory?

The AMD Instinct MI325X has the most accelerator memory available today: 256 GB of HBM3e per GPU with 6,000 GB/s of bandwidth, according to AMD's product datasheet. The NVIDIA B200 follows with 192 GB of HBM3e at 8,000 GB/s, and the AMD Instinct MI300X also provides 192 GB at 5,300 GB/s. For transformer inference, a model must fit its weights plus KV cache inside the memory of the GPUs allocated to it, so VRAM capacity — not raw FLOPS — usually decides how many GPUs a deployment needs.

Server GPU comparison table

AcceleratorYearVRAMTypeBandwidthTDPGB per kWInterconnect
NVIDIA B300 (Blackwell Ultra) 2025 288 GB HBM3e 8,000 GB/s 1400 W 205.7 GB
AMD Instinct MI355X 2025 288 GB HBM3E 8,000 GB/s 1400 W 205.7 GB
AMD Instinct MI325X 2024 256 GB HBM3e 6,000 GB/s 1000 W 256 GB
NVIDIA B200 2025 192 GB HBM3e 8,000 GB/s 1000 W 192 GB NVLink 5 (1.8 TB/s)
AMD Instinct MI300X 2023 192 GB HBM3 5,300 GB/s 750 W 256 GB
H200 SXM 2024 141 GB HBM3e 4,800 GB/s 700 W 201.4 GB
H100 SXM 2023 80 GB HBM3 3,350 GB/s 700 W 114.3 GB
NVIDIA A100 80GB SXM 2021 80 GB HBM2e 2,039 GB/s 400 W 200 GB
NVIDIA H100 PCIe 80GB 2023 80 GB HBM2e 2,039 GB/s 350 W 228.6 GB PCIe 5.0 x16
NVIDIA A100 40GB SXM (EOL) 2020 40 GB HBM2 1,555 GB/s 400 W 100 GB

Specs from vendor datasheets (source link on each accelerator's detail page). GB per kW = VRAM ÷ board power, computed from the same datasheet values — a rough proxy for memory capacity per rack watt.

How many GPUs does each model need?

The matrix below computes GPUs needed = required VRAM ÷ accelerator VRAM, rounded up. Required VRAM is Q4_K_M weights (4.8 bits per weight) plus a 10% loading overhead and 4k-context KV cache, using the same ModelFitService math as our model pages. Green = fits on one GPU.

Model (Q4_K_M, 4k ctx)Required VRAMB300 (Blackwell Ultra)
288 GB
MI355X
288 GB
MI325X
256 GB
B200
192 GB
MI300X
192 GB
H200 SXM
141 GB
H100 SXM
80 GB
A100 80GB SXM
80 GB
H100 PCIe 80GB
80 GB
A100 40GB SXM
40 GB
DeepSeek V4 & V4-Flash 187.4 GB 1 1 1 1 1 2 3 3 3 5
Llama 3.1 70B 47.9 GB 1 1 1 1 1 1 1 1 1 2
Llama 3.3 70B 47.9 GB 1 1 1 1 1 1 1 1 1 2
Qwen 3 32B 22.7 GB 1 1 1 1 1 1 1 1 1 1
Gemma 3 27B 20.2 GB 1 1 1 1 1 1 1 1 1 1

Weights-only math. Tensor-parallel inference also needs headroom for activations and NCCL buffers; production clusters typically reserve ~20% more VRAM than this table shows.

What is the minimum server GPU for 70B-class models?

A single 80 GB accelerator runs a 70B-parameter model like Llama 3.3 70B at Q4_K_M: the weights are about 48 GB, which fits with room for KV cache on one A100 80GB, H100 80GB, or H200. The A100 40GB SXM cannot hold it and needs two GPUs. The cheapest realistic path to a 70B model today is a used dual-A100 80GB server; for new hardware, one H100 PCIe 80GB (350 W, 80 GB HBM2e) is the entry point.

What does it take to run DeepSeek V4 locally?

DeepSeek V4 is a 284B-parameter mixture-of-experts model, so at Q4_K_M its weights alone are roughly 170 GB and the full requirement is about 187.4 GB. A single B200 (192 GB), MI300X (192 GB), or MI325X (256 GB) holds it; H200 needs two GPUs and H100 needs three. Its 13B active-parameter count keeps token generation fast relative to dense models of the same total size, but all weights must still be resident in VRAM.

SXM versus PCIe for AI servers

SXM modules (B200, H100 SXM, H200, MI300X, MI325X) mount directly on the server board and exchange tensor-parallel traffic over NVLink 5 at up to 1.8 TB/s per GPU, which matters when a model's layers are split across GPUs. PCIe cards (H100 PCIe, A100 PCIe) cap GPU-to-GPU traffic at PCIe 5.0 x16 speeds and suit single-GPU or loosely coupled deployments. If your model fits on one GPU, PCIe is simpler and cheaper; if you split models across GPUs, SXM plus NVLink pays for itself in inter-GPU bandwidth.

Power and rack planning quick facts

Methodology: specs are pulled live from our product database, where each accelerator was verified against the vendor datasheet (NVIDIA or AMD product pages). GPU-count math uses Q4_K_M (4.8 bits/weight) + 10% overhead + 4k-context KV cache — the same formulas documented on our methodology page. No benchmark numbers on this page are estimates; this page contains no throughput benchmarks at all.

Related