News AMD announces MI350X and MI355X AI GPUs, claims up to 4X generational gain, up to 35X faster inference performance

[Admin]

AMD unveiled its new MI350X and MI355X GPUs for AI workloads here at its Advancing AI 2025 event in San Jose, California, claiming the new accelerators offer a 3X performance boost over the prior-gen MI300X, positioning the company to improve its competitive footing against its market-leading rival, Nvidia.

AMD announces MI350X and MI355X AI GPUs, claims up to 4X generational gain, up to 35X faster inference performance : Read more

 

[anemusek]

AMD had a huge advantage thanks to the efficiency in Matrix FP64, so it is getting rid of this advantage. And this at a time when current AI models based on the ability to high quantization to parameters with low precision have shown their limitations and lack of development prospects, and the most promising revolution in the approach to AI are models based on simulating analog neurons, which require... High precision, preferably in Matrix FP64.
It's a bit like deciding that selling a kidney is a good business idea ;P

 

[Pemalite]

AMD had a huge advantage thanks to the efficiency in Matrix FP64, so it is getting rid of this advantage. And this at a time when current AI models based on the ability to high quantization to parameters with low precision have shown their limitations and lack of development prospects, and the most promising revolution in the approach to AI are models based on simulating analog neurons, which require... High precision, preferably in Matrix FP64.
It's a bit like deciding that selling a kidney is a good business idea ;P

AMD is just going where the money is at this stage.

 

[redgarl]

Your title is misleading, TCO means Total Cost of Ownership, which means higher TDP == HIGHER TCO.

What you means is Higher TBP than previous gen, but better TCO.

 

[redgarl]

AMD is just going where the money is at this stage.

This is what they receive from their partners feedback. What would have been problematic is if they stuck to their guns while doing their own thing and asking their partner to adapt to their ecosystem.

 

[bit_user]

the most promising revolution in the approach to AI are models based on simulating analog neurons, which require... High precision, preferably in Matrix FP64.

I'm not sure exactly what you're referring to, in that first part of your assertion. However, to go from the super low-precision formats that are currently in vogue to the other end of the spectrum seems quite extreme and I think probably unjustifiable. Even if neither FP16 nor BF16 were up to the task, FP32 is much more energy-efficent and area-efficient than FP64. IIRC, it's something like a 6:1 advantage, on each axis.

FP64 is mainly useful just for scientific and financial applications. It has some extremely limited applications in graphics, which is why client GPUs require at least token support for it.