Gaudi 3 vs MI300X

MI300X has 1.5× the memory; MI300X has 1.4× the bandwidth; MI300X rents for 1.2× less.

The Intel Gaudi 3 (Gaudi, 2024) and the AMD Instinct MI300X (CDNA 3, 2023) come from different vendors and different design goals, but they get cross-shopped for a reason. On capacity, the MI300X carries 1.5x the memory (192 vs 128 GB), which sets what each can hold at all. On memory bandwidth, the number that governs LLM serving speed, the MI300X leads at 5.3 vs 3.7 TB/s. Raw compute favors the MI300X by 1.4x, which matters for training, prefill-heavy traffic, and diffusion work more than for chat-style decoding.

Price is where it settles: the MI300X rents from $1.85/hr against $2.30/hr for the Gaudi 3, a 1.2x gap that the performance numbers above only partly close. Per gigabyte of VRAM, the MI300X is the cheaper rental, worth knowing if your model is capacity-bound rather than speed-bound. The full case for and against Intel's part is in Gaudi 3: the accelerator nobody is fighting over. Everything below is the underlying data.

Intel · Gaudi · 2024

Intel Gaudi 3

Intel's price-performance play, with Ethernet-native scale-out instead of proprietary interconnect.

From $2.30/hr at FluidStack

AMD · CDNA 3 · 2023

AMD Instinct MI300X

AMD's answer to Hopper: 192 GB on one GPU, a 70B model in FP16 fits with room to spare.

From $1.85/hr at Vast.ai

Head to head

■ Gaudi 3   ■ MI300X, bars share one scale across the whole catalog.

Memory 128 GB 192 GB
Memory bandwidth 3.7 TB/s 5.3 TB/s
FP16 dense 918 TFLOPS 1,307 TFLOPS
FP8 dense 1,835 TFLOPS 2,614 TFLOPS
Power (TDP) 900 W 750 W
Gaudi 3MI300X
Memory128 GB HBM2e192 GB HBM3
Bandwidth3.7 TB/s5.3 TB/s
FP16 dense918 TF1,307 TF
FP8 dense1,835 TF2,614 TF
TDP900 W750 W
Interconnect24×200GbE RoCE on-chipInfinity Fabric · 896 GB/s
Cheapest rental $2.30/hr $1.85/hr
$/hr per GB VRAM $1.8¢ $1.0¢

Rule of thumb: for LLM serving, prefer the GPU with more memory bandwidth per dollar; for training and prefill-heavy work, prefer FLOPS per dollar; and if the model doesn't fit in VRAM, none of the other numbers matter. Sanity-check with the VRAM calculator.