AMD Challenges Goliath with MI355, Doubles Down on Open Innovation
With over $500 billion forecasted in the AI accelerator market, all eyes were focused in San Jose this week to hear from Lisa Su and her team on how AMD was progressing on its strategy to take the leadership mantle for AI infrastructure delivery.
The theme of the day? Open Innovation. With a keynote filled with heady announcements, the throughline weaved across discussions showcased that AMD fully intends to be both the partner and customer choice for collaborative innovation as the world’s data centers pivot to broadscale deployments of generative and agentic AI at scale. And while many feel that competitive NVIDIA platforms have a stronghold grip on the market, one thing I kept thinking about as Lisa discussed the commanding progress of her team is that she knows more than a little about relentless pursuit of a Goliath and is pretty good at playing the giant slayer. Let’s break down the news of the day.
AI Leadership Starts with Performance – Across CPU, GPU, and Network
Data center-scale performance will define success with AI system proliferation, and AMD is the only company offering a full suite of CPU, GPU, and network solutions needed to deliver AI clusters. Today, AMD delivered its Instinct MI355 to the market, the 4th generation architecture built on 3nm process and packing 185 billion transistors. What does that mean for real-world performance? With up to 35X gen-over-gen performance, AMD has closed a lot of the gap vs. NVIDIA B200, delivering an average of 3.5X performance uplift across training and inference. Some examples of competitive head-to-head metrics showed performance parity with different Llama configurations running pre-training workloads and 1.1X+ performance in equivalent fine-tuning environments. While it’s too early and frankly naïve to declare AMD the performance winner, these gains certainly will open doors to deeper collaboration opportunities with customers with this generation of products.
Deeper customer collaborations are essential for the enormous sunshine of what comes next: Helios. In 2026, AMD will unify their entire silicon portfolio to deliver Helios, an AI-optimized rack-scale system that integrates next-generation Venice EPYC processors, next-generation Instinct MI400 GPUs, and next-generation Pensando Vulcano DPUs, tapping UAL and Ultra Ethernet connectivity. This behemoth will be delivered as an OCP-compliant, open-standards based solution, which is expected to turn a lot of heads given claimed performance parity with NVIDIA Vera Rubin across GPU domain, scale-up bandwidth, FP4/FP8 FLOPS, and a 1.5x advantage in HBM4 memory capacity, memory bandwidth, and scale-out networking. At the heart of this configuration, of course, is the upcoming Instinct MI400 GPU, which did a bit of stage stealing from the MI355 introduction today as Lisa claimed a targeted 10X performance improvement gen over gen. Even in the era of 2x Moore’s Law progression, this is a stunning performance improvement target that places the competition on its heels.
Powering the AI software stack behind Helios is ROCm, AMD’s open compute platform, which provides the common programming foundation across the Instinct GPU lineup. With support for major AI frameworks and optimized libraries, ROCm enables portability, scale, and high performance across AMD hardware – cementing AMD’s strategy of combining open software with open systems to meet enterprise and hyperscaler needs.
It's Open Everywhere
While the Instinct MI355 introduction and promise of Helios commanded attention, the message of open innovation was delivered in every word from AMD and partner executives, reflecting a strong customer desire for choice in AI solution alternatives. This open innovation starts with advancement of the ROCm software platform, AMD’s alternative to CUDA with ROCm 7 introduced today. Historically, software has not been AMD’s strong suit, and it’s obvious that the company is investing to change this through internal development and acquisition of talent including a high-profile addition of Lamini, who was on hand to share how a full suite of developer training will soon be delivered to accelerate developer activation of the software in AI development.
Open was also supported through AMD’s strong commitment to the Open Compute Project Foundation and leadership within the Universal Accelerator Link (UAL) and Ultra Ethernet standards efforts, with Ultra Ethernet hitting 1.0 this week. This commitment to standards-based networking and significant keynote time delivered to standards-based advancement underscored the importance of network scale to AI compute delivery as well as a key differentiator from competition who relies on proprietary solutions. It was fantastic to see the industry support of Astera Labs and Marvell highlighted as leading examples of a vibrant networking industry assembling to deliver the next generation of AI data center connectivity, and I expect to see a lot more about solution delivery from a host of vendors in the coming year as standards mature and customers begin deploying these standards based solutions.
The Developer is the Center of Innovation
AMD placed a central spotlight on developers in attendance and its developer sessions as it continues to advance ROCm. To help tell this story, leaders from (FILL IN) emerged to share the progress of advancing AI together. This demonstrated that real developer cycles are being spent today on optimizing on AMD Instinct GPU-based infrastructure, and there is a groundswell of activation in this space. With the announcement of a new developer cloud and free access to all developer attendees, AMD is providing the access and tools required to support community advancement. While CUDA has a tremendous lead in this space, AMD is taking the right steps to at least get developers to take their tools and platforms for a test drive, and while doing so gain traction with developer loyalty.
AI is Everywhere, and Access for All is Essential
With AI becoming a central aspect of geo-political policy, and as nation states race for AI sovereignty and supremacy, Lisa shared how important AI advancement for all was to realize the true vision of this historic technology. Tariq Amin, CEO of Humain, a leading AI operator in Saudi Arabia, came onstage to share his vision for AI advancement in the kingdom, noting that Saudi Arabia was a young nation full of innovation potential. The collaboration announced earlier this spring will bring AI platforms featuring AMD Instinct GPUs, EPYC CPUs, Pensando DPUs, RyzenAI and ROCm software to Humain data centers to deliver 500 megawatts of compute capacity over the next 5 years. This reflects a joint commitment to democratize AI access and foster innovation through scalable compute.
The Customer is King
One thing that was striking about this event was the customer-centricity in everything that the AMD team designed. Through walk-ons with OpenAI CEO Sam Altman, Meta, Oracle Cloud, Cohere and more, Lisa and team continually discussed how deep collaborations with customers directly fuel design targets for AMD silicon advancement. This laser focus played out in the progress of collaboration advancement from science project sized deployments of first gen Instinct GPU clusters to scale deployments running significant customer workloads today. It also reflects in the open innovation focus with a notion that many voices, centered on customer requirements, will define ultimate AI advancement – and underscores the danger of so much global innovation in the hands of a single company.
Altman's appearance was particularly notable as he announced that OpenAI will be using AMD's upcoming MI400 chips, telling the audience, “It's gonna be an amazing thing” – a significant endorsement that lends considerable credibility to AMD's ambitious 10x performance improvement targets for the MI400.
The TechArena Take: The Slingshot’s Aimed at You, Goliath
So what’s the TechArena take? While NVIDIA certainly holds pole position on both market share and AI zeitgeist, the industry is collectively hungry for alternatives and worried that a single vendor will squeeze industry opportunity from this important moment. Frankly, the TAM for AI accelerators is too vast not to expect competitive alternatives to earn a segment of upcoming deployments, and AMD has put together the portfolio and partnerships to be that leading competitor. What’s striking to me is the advancement that the company is delivering gen-over-gen to bridge the gap and deliver a credible alternative to market. Investments and leadership in UAL and Ultra Ethernet will pay off as InfiniBand, a proprietary solution that was developed decades ago, is showing some tarnish.
I want to see more from ROCm software advancement and developer traction to get true solution parity with CUDA-fueled solutions, and AMD has work to deepen relationships with this essential element of the ecosystem. But even given that, the advancement is truly eye-opening, welcome, and inspiring us to anticipate what comes next. Well done, AMD.
With over $500 billion forecasted in the AI accelerator market, all eyes were focused in San Jose this week to hear from Lisa Su and her team on how AMD was progressing on its strategy to take the leadership mantle for AI infrastructure delivery.
The theme of the day? Open Innovation. With a keynote filled with heady announcements, the throughline weaved across discussions showcased that AMD fully intends to be both the partner and customer choice for collaborative innovation as the world’s data centers pivot to broadscale deployments of generative and agentic AI at scale. And while many feel that competitive NVIDIA platforms have a stronghold grip on the market, one thing I kept thinking about as Lisa discussed the commanding progress of her team is that she knows more than a little about relentless pursuit of a Goliath and is pretty good at playing the giant slayer. Let’s break down the news of the day.
AI Leadership Starts with Performance – Across CPU, GPU, and Network
Data center-scale performance will define success with AI system proliferation, and AMD is the only company offering a full suite of CPU, GPU, and network solutions needed to deliver AI clusters. Today, AMD delivered its Instinct MI355 to the market, the 4th generation architecture built on 3nm process and packing 185 billion transistors. What does that mean for real-world performance? With up to 35X gen-over-gen performance, AMD has closed a lot of the gap vs. NVIDIA B200, delivering an average of 3.5X performance uplift across training and inference. Some examples of competitive head-to-head metrics showed performance parity with different Llama configurations running pre-training workloads and 1.1X+ performance in equivalent fine-tuning environments. While it’s too early and frankly naïve to declare AMD the performance winner, these gains certainly will open doors to deeper collaboration opportunities with customers with this generation of products.
Deeper customer collaborations are essential for the enormous sunshine of what comes next: Helios. In 2026, AMD will unify their entire silicon portfolio to deliver Helios, an AI-optimized rack-scale system that integrates next-generation Venice EPYC processors, next-generation Instinct MI400 GPUs, and next-generation Pensando Vulcano DPUs, tapping UAL and Ultra Ethernet connectivity. This behemoth will be delivered as an OCP-compliant, open-standards based solution, which is expected to turn a lot of heads given claimed performance parity with NVIDIA Vera Rubin across GPU domain, scale-up bandwidth, FP4/FP8 FLOPS, and a 1.5x advantage in HBM4 memory capacity, memory bandwidth, and scale-out networking. At the heart of this configuration, of course, is the upcoming Instinct MI400 GPU, which did a bit of stage stealing from the MI355 introduction today as Lisa claimed a targeted 10X performance improvement gen over gen. Even in the era of 2x Moore’s Law progression, this is a stunning performance improvement target that places the competition on its heels.
Powering the AI software stack behind Helios is ROCm, AMD’s open compute platform, which provides the common programming foundation across the Instinct GPU lineup. With support for major AI frameworks and optimized libraries, ROCm enables portability, scale, and high performance across AMD hardware – cementing AMD’s strategy of combining open software with open systems to meet enterprise and hyperscaler needs.
It's Open Everywhere
While the Instinct MI355 introduction and promise of Helios commanded attention, the message of open innovation was delivered in every word from AMD and partner executives, reflecting a strong customer desire for choice in AI solution alternatives. This open innovation starts with advancement of the ROCm software platform, AMD’s alternative to CUDA with ROCm 7 introduced today. Historically, software has not been AMD’s strong suit, and it’s obvious that the company is investing to change this through internal development and acquisition of talent including a high-profile addition of Lamini, who was on hand to share how a full suite of developer training will soon be delivered to accelerate developer activation of the software in AI development.
Open was also supported through AMD’s strong commitment to the Open Compute Project Foundation and leadership within the Universal Accelerator Link (UAL) and Ultra Ethernet standards efforts, with Ultra Ethernet hitting 1.0 this week. This commitment to standards-based networking and significant keynote time delivered to standards-based advancement underscored the importance of network scale to AI compute delivery as well as a key differentiator from competition who relies on proprietary solutions. It was fantastic to see the industry support of Astera Labs and Marvell highlighted as leading examples of a vibrant networking industry assembling to deliver the next generation of AI data center connectivity, and I expect to see a lot more about solution delivery from a host of vendors in the coming year as standards mature and customers begin deploying these standards based solutions.
The Developer is the Center of Innovation
AMD placed a central spotlight on developers in attendance and its developer sessions as it continues to advance ROCm. To help tell this story, leaders from (FILL IN) emerged to share the progress of advancing AI together. This demonstrated that real developer cycles are being spent today on optimizing on AMD Instinct GPU-based infrastructure, and there is a groundswell of activation in this space. With the announcement of a new developer cloud and free access to all developer attendees, AMD is providing the access and tools required to support community advancement. While CUDA has a tremendous lead in this space, AMD is taking the right steps to at least get developers to take their tools and platforms for a test drive, and while doing so gain traction with developer loyalty.
AI is Everywhere, and Access for All is Essential
With AI becoming a central aspect of geo-political policy, and as nation states race for AI sovereignty and supremacy, Lisa shared how important AI advancement for all was to realize the true vision of this historic technology. Tariq Amin, CEO of Humain, a leading AI operator in Saudi Arabia, came onstage to share his vision for AI advancement in the kingdom, noting that Saudi Arabia was a young nation full of innovation potential. The collaboration announced earlier this spring will bring AI platforms featuring AMD Instinct GPUs, EPYC CPUs, Pensando DPUs, RyzenAI and ROCm software to Humain data centers to deliver 500 megawatts of compute capacity over the next 5 years. This reflects a joint commitment to democratize AI access and foster innovation through scalable compute.
The Customer is King
One thing that was striking about this event was the customer-centricity in everything that the AMD team designed. Through walk-ons with OpenAI CEO Sam Altman, Meta, Oracle Cloud, Cohere and more, Lisa and team continually discussed how deep collaborations with customers directly fuel design targets for AMD silicon advancement. This laser focus played out in the progress of collaboration advancement from science project sized deployments of first gen Instinct GPU clusters to scale deployments running significant customer workloads today. It also reflects in the open innovation focus with a notion that many voices, centered on customer requirements, will define ultimate AI advancement – and underscores the danger of so much global innovation in the hands of a single company.
Altman's appearance was particularly notable as he announced that OpenAI will be using AMD's upcoming MI400 chips, telling the audience, “It's gonna be an amazing thing” – a significant endorsement that lends considerable credibility to AMD's ambitious 10x performance improvement targets for the MI400.
The TechArena Take: The Slingshot’s Aimed at You, Goliath
So what’s the TechArena take? While NVIDIA certainly holds pole position on both market share and AI zeitgeist, the industry is collectively hungry for alternatives and worried that a single vendor will squeeze industry opportunity from this important moment. Frankly, the TAM for AI accelerators is too vast not to expect competitive alternatives to earn a segment of upcoming deployments, and AMD has put together the portfolio and partnerships to be that leading competitor. What’s striking to me is the advancement that the company is delivering gen-over-gen to bridge the gap and deliver a credible alternative to market. Investments and leadership in UAL and Ultra Ethernet will pay off as InfiniBand, a proprietary solution that was developed decades ago, is showing some tarnish.
I want to see more from ROCm software advancement and developer traction to get true solution parity with CUDA-fueled solutions, and AMD has work to deepen relationships with this essential element of the ecosystem. But even given that, the advancement is truly eye-opening, welcome, and inspiring us to anticipate what comes next. Well done, AMD.
.png)
