Hardware & InferenceBusiness & Market 🇷🇺 07.08.2026 12:01

NVIDIA RTX PRO 5000 Blackwell with 72 GB VRAM: Is It Worth the Premium Over the Flagship?

NVIDIANVIDIA DeepSeekDeepSeek OpenAIOpenAI Alibaba/QwenAlibaba/Qwen
The RTX PRO 5000 Blackwell in a special 72 GB version is tested and compared to the RTX PRO 6000 (96 GB), the previous RTX A5000, and AMD Radeon AI PRO R9700 setups. Despite lower specs, it performs close to the flagship in inference, typically within 15-20% at half the power draw, but pricing and availability make it a 'golden middle' for local AI workloads.
This article from Habr examines the NVIDIA RTX PRO 5000 Blackwell with an increased 72 GB of GDDR7 VRAM, a high-memory variant released for larger AI models. Compared to the standard 48 GB version, it doubles the price (around $9999 vs. $5000) and supports MIG splits of 2x36 GB. Benchmarks against the RTX PRO 6000 Blackwell (96 GB) show only 15-20% slower generation speed, with both handling large contexts (up to 128k tokens for mid-size models). Against the previous-gen RTX A5000, it is nearly twice as fast and can load 10x more context, while outperforming multi-GPU AMD setups by 2-5x in tokens/sec, though AMD cards are cheaper per GB. The card was also tested with Qwen3-Coder-Next via Ollama, delivering 100-120 tok/s on a real coding task, and with gpt-oss-120b reaching 40-50 tok/s. The author concludes that due to current prices and memory shortages, the RTX PRO 5000 Blackwell offers the best balance for enthusiasts and professionals, consuming half the power of the 6000 and excelling in inference tasks.
Abbreviations
VRAM = Video Random Access Memory — видеопамять
MIG = Multi-Instance GPU — многоинстансный GPU
GDDR7 = Graphics Double Data Rate 7 — графическая память седьмого поколения
CUDA = Compute Unified Device Architecture — вычислительная архитектура NVIDIA
FP32 = Floating Point 32-bit — числа с плавающей запятой одинарной точности
TDP = Thermal Design Power — тепловыделение
FHFL = Full Height Full Length — полная высота и длина
ECC = Error Correction Code — код коррекции ошибок
QoS = Quality of Service — качество обслуживания
NVFP4 = NVIDIA Floating Point 4-bit — 4-битные числа с плавающей запятой NVIDIA
Source: Habr — хаб ИИ — original
Our earlier posts on this topic ↓
Fresh news