AMD Announces MI350 GPU And Future Roadmap Details

Forbes

2 days ago

Dr. Lisa Su, Chairman and CEO of AMD kicked off the event.
AMD held their now-annual Advancing AI event today in Silicon Valley, with new GPUs, new networking, new software, and even a rack-scale architecture for 2026/27 to better compete with the Nvidia NVL72 that is taking the AI world by storm. Let's dive in!
While AMD g has yet to achieve investor expectations, and its products remain a distant second to Nvidia, AMD continues to keep to its commitment to an annual accelerator roadmap, delivering nearly four times better performance gen-on-gen with the MI350. That pace could help it catch up to Nvidia on GPU performance, and keeps it ahead of Nvidia regarding memory capacity and bandwidth, although Nvidia's lead in networking, system design, AI software, and ecosystem remains intact.
However, AMD has stepped up its networking game with support for UltraEthernet this year and UALink next year for scale-out and scale-up, respectively. And, for the first time, AMD showed a 2026/27 roadmap with the 'Helios' rack-scale AI system that helps somewhat versus Nvidia NVL72 and the upcoming Kyber rack-scale system. At least AMD is now on the playing field.
Oracle said they are standing up a 27,000 GPU cluster using AMD Instinct GPUs on Oracle Cloud Compute Infrastructure, so AMD is definitely gaining traction. AMD also unveiled ROCm 7.0 and the AMD Developer Cloud Access Program, helping it build a larger and stronger AI ecosystem.
While the AMD Instinct GPU portfolio has struggled to catch up with Nvidia, customers value the price/performance and openness of AMD. In fact, AMD claims to offer 40% more tokens per dollar, and that 7 of the 10 largest AI companies have adopted AMD GPUs, among over 60 named customers.
The biggest claim to fame AMD touts is the larger memory footprint it supports, now at 288 GB of HBM3 memory with the MI350. Thats enough memory to hold today's larger models, up to 520B parameters, on a single node, and 60% more than the competition. That translates to lower TCO for many models. The MI350 also has twice the 64-bit floating point performance versus Nvidia, important for HPC workloads.
The MI350 and 355X GPUs represent a step up in memory and performance over their predecessors.
The MI355 is the same silicon as the MI300 but is selected to run faster and hotter, and is AMD's flagship data center GPU. Both GPUs are available on the UBB8 industry standard boards in both air- and liquid cooled versions.
AMD MI350 supports 288GB of HBM3 mempory and UBB8 baseboards.
AMD claims, and has finally demonstrated through MLPerf benchmarks, that the MI355 is roughly three times faster than the MI300, and even on par with the Nvidia B200 GPU from Nvidia. But keep in mind that Nvidia NVLink, InfiniBand, system design, ecosystem, and software keep it in a leadership position for AI, while the B300 will begin shipment soon.
AMD claims that the MI355X with FP4 support is a bit faster than the B200 and GB200 using ... More TensorRT-LLM.
AMD added some detail on next year's MI400 series as well. Sam Altman himself appeared on stage and gave the MI450 some serious love. His company has been instrumental in laying out the market requirements to the AMD engineering teams.
Sam Altman, CEO Of OpenAI, gave AMD some serious love.
The MI400 will use HBM4 at 423GB per GPU, as well as supporting 300GB/s UltraEthernet through Pensando NICs.
AMD announced details about next year's MI400 GPU. Looks good!
To put the MI400 performance into perspective, check out the hockey stick performance they are expecting in the graph below. This reminds us of a similar slide Jensen Huang used at GTC. Clearly, AMD is on the right path.
The MI400 is a huge step forward for AMD.
While a lot of attention in the AMD Advancing AI event surrounded the MI350/355 GPUs and the roadmap, the networking section was more exciting and important.
The Pensando Pollara 400 AI NIC will support UltraEthernet for massive cluster scaling.
More important to large-scale AI, AMD is an original member of the UALink consortium, and will support UALink with the MI400 series. While the slide below makes it look amazing, keep in mind that Nvidia will likely be shipping NVLink 6.0 in the same timeframe, or earlier.
AMD will support UALink for Scale-up and UltraEthernet for Scale-out.
Finally, let's give ROCm some credit. The development team has been hard at work since the Silicon Analysis crushed the AI software stack late last year, and they have some good performance results to show for it as well as ecosystem adoption.
AMD ROCm has improved significantly over the last 2 years and has seen broad ecosystem collaboration
To demonstrate the performance point, AMD showed over three times the performance for inference processing using ROCm 7. This is in part due to the ever-improving state of the open AI stack such as Triton from OpenAI, and is a developing trend that will keep Nvidia on its toes.
AMD has improved ROCm performance by over 3-fold,