What is the best GPU for running local LLMs?

For the most memory per dollar, a 128 GB DGX Spark or Ryzen AI Max+ 395 desktop; for speed with large models, an RTX PRO 6000 with 96 GB. The RTX 5090's 32 GB limits it to smaller models.

How much VRAM do I need for a 70B model?

A quantised 70-billion-parameter model needs roughly 40 to 50 GB, so a 96 GB RTX PRO 6000 or a 128 GB desktop runs it in one device.

Is DGX Spark faster than an RTX 5090?

Not for generating text on models that fit both: the RTX 5090 has about 1,792 GB/s of bandwidth against DGX Spark's 273 GB/s. DGX Spark's advantage is 128 GB of memory, four times the 5090's.

Buyer guide

Best GPU for AI and local LLMs in Australia (November 2026)

By Nicholas Dodemaide, Director · Updated 15 November 2026 · 6 min read

For local AI, memory decides what you can run and bandwidth decides how fast it runs. On Australian prices of 15 November 2026, 128 GB AI desktops cost about A$70 to A$73 per gigabyte of GPU-usable memory, against A$270 for an RTX PRO 6000 and A$406 for an RTX 5090. Buy a 128 GB desktop to run big models cheaply, an RTX PRO 6000 for speed with 96 GB, and an H200 NVL or HGX server for production.

The short answer

Pick by the largest model you need in memory, then by speed. Up to about 30 billion parameters: RTX 5090. Up to about 70 billion with fast responses: RTX PRO 6000 (96 GB). Large models on a budget: a 128 GB DGX Spark or Ryzen AI Max+ 395 desktop. Many users or production serving: H200 NVL cards or an HGX server.

Price per gigabyte of model memory in Australia

Option	GPU-usable memory	Memory bandwidth	Lowest Australian price	A$ per GB
Ryzen AI Max+ 395 desktop (128 GB)	up to 96 GB	256 GB/s	A$6,999	about A$73
NVIDIA DGX Spark (GB10)	128 GB	273 GB/s	A$8,999	about A$70
RTX PRO 6000 Blackwell	96 GB	1,792 GB/s	A$25,900	about A$270
H200 NVL	141 GB	4,800 GB/s	A$52,999	about A$376
GeForce RTX 5090	32 GB	1,792 GB/s	A$12,999	about A$406

Lowest Australian listings incl. GST, 15 November 2026 (Parkways Compute Price Index). A$ per GB = lowest price divided by GPU-usable memory; desktops include CPU, storage and case, cards do not.

Why bandwidth matters as much as memory

When a model generates text it reads its weights from memory for every token, so tokens per second scale roughly with memory bandwidth. An RTX PRO 6000 moves data about seven times faster than a 128 GB desktop (1,792 GB/s against 256 to 273 GB/s), so the same model answers several times faster. The desktops win on how big a model fits, the cards on how quickly it replies.

Which to buy

Developer experimenting with open models: Ryzen AI Max+ 395 desktop, the lowest cost per gigabyte.
Team that needs CUDA and NVIDIA's software stack on a desk: DGX Spark or another GB10 system; two can be linked for larger models.
Fast private inference for a team, 70B-class models: RTX PRO 6000, or two in one workstation for 192 GB.
Gaming and AI on one card, smaller models: RTX 5090.
Production serving for many users: H200 NVL cards in a standard server, or an HGX H200 system.

Methodology: how we compared

Prices are the lowest Australian retail listings including GST recorded in the Parkways Compute Australian AI Hardware Price Index on 15 November 2026. Memory and bandwidth figures are manufacturer specifications. Price per gigabyte divides the lowest price by memory the GPU can use; it is a buying shortcut, not a full cost of ownership.

Common questions

What is the best GPU for running local LLMs?: For the most memory per dollar, a 128 GB DGX Spark or Ryzen AI Max+ 395 desktop; for speed with large models, an RTX PRO 6000 with 96 GB. The RTX 5090's 32 GB limits it to smaller models.
How much VRAM do I need for a 70B model?: A quantised 70-billion-parameter model needs roughly 40 to 50 GB, so a 96 GB RTX PRO 6000 or a 128 GB desktop runs it in one device.
Is DGX Spark faster than an RTX 5090?: Not for generating text on models that fit both: the RTX 5090 has about 1,792 GB/s of bandwidth against DGX Spark's 273 GB/s. DGX Spark's advantage is 128 GB of memory, four times the 5090's.