AMD, OpenAI Jointly Release Ultra-Powerful AI Chip with 35x Improvement in Inference, Surpassing NVIDIA B200!

Early this morning.AMDorganized2025the (whole) worldAIDevelopment Conference, which featured the release and presentation of the latestAIChips and cloud infrastructure hardware devices.

OpenAICo-Founder and Chief Executive OfficerSam AltmanAttended the conference as a special guest with theADMJointly releasedInstinct MI400,Instinct MI350Series UltraAIChip. In particular, the development of theOpenAIalways been a good idea forAMDProvide technical feedback to help them optimizeGPUs.

At the launch event, whenSam Altmanhear from an individualMI400outfit with432G(used form a nominal expression)HBM4Memory was also stunned and exclaimed that it was impossible.

AMDThis release of the latestAIThe chip is primarily linked to NVIDIA'sBlackwell chip competition, NVIDIA is currently the AMD existAIdata centerGPUsThe only competitor in the field.

AMD Instinct™ MI350rangeGPUsis based onAMD CDNA™ 4The latest in a series of architectures designed for modernAIdesigned for the needs of the infrastructure. The range includesMI350Xcap (a poem)MI355Xtwo modelsGPUs.

Compared to the previous generationMI350equipped with288GB(used form a nominal expression)HBM3Ememory and up to8TB/sThe memory bandwidth of theAIIncreased computing power on the4times and improved inference performance by a factor of35Times.

AMDsaid that because the chip consumes less power than competitors, theMI355XEach dollar provides more chips than NVIDIA's40%(used form a nominal expression)tokens.

MI355XPlatforms inFP4Performance is up to161 PFLOPSbut (not)MI350XThe platforms, on the other hand, are inFP16Performance is up to36.8 PFLOPS. TheseGPUsNot only does it excel in performance, but it also offers flexible cooling configurations, including air-cooled and direct liquid cooling, to support large-scale deployments, such as up to64classifier for individual things or people, general, catch-all classifierGPUsor in direct liquid-cooled environments to support up to128classifier for individual things or people, general, catch-all classifierGPUs.

To further enhance GPU performance, AMD has also open sourced an AI acceleration platform, ROCm7. over the past year, ROCm has matured rapidly, delivering leading inference performance, expanded training capabilities, and deep integration with the open source community. rocm now supports some of the world's largest AI platforms, such as LLaMA and DeepSeek, and will provide more than a doubling of inference performance improvements in the upcoming release of ROCm version 7 provides over 3.5x inference performance improvements.

ROCm Enterprise AIbecause ofAIDeployment provides a completeMLOpsplatform that supports secure, scalableAIdevelopment and provides a wealth of tools for fine-tuning, compliance, deployment and integration.

Instinct MI400thenAMDnext-generation flagshipAIChip, too.AIAll-in-one"Helios"Core Components. In terms of memory configuration, theMI400 The series is expected to carry up to 432GB(used form a nominal expression)HBM4 High-speed video memory, compared to the previous generationMI350collection of 36TB HBM3EMemory has been significantly increased, and the high-bandwidth memory architecture can be used for large-scale AI The model provides sufficient data throughput to meet the needs of model parameter loading and fast computation.

In terms of computing performance, the MI400 series can reach 40 petaflops of computing power at FP4 precision, which is optimized for low-precision computation in AI training, and can effectively accelerate the training efficiency of mainstream models such as Transformer. Meanwhile, the MI400 series is equipped with 300GB/s scale-out bandwidth, and the UALink open standard technology realizes seamless interconnection of 72 GPUs, enabling GPUs in the entire rack to work together as a unified computing unit, breaking through the communication bottleneck of traditional architectures.

MI400series with6th Gen AMD EPYC "Venice" CPUs up to Pensando "Vulcano" AI NIC Formation of technological synergies. Among other things, based on Zen 6 structured Venice CPU Offers up to 256 Core and 1.6 TB/s The memory bandwidth of the GPUs Efficient task scheduling and resource management for clusters;

(indicates contrast) Vulcano AI NIC be in favor of 800G Network throughput, with its UALink cap (a poem) PCIe The dual interface design realizes GPUs together with CPU Low-latency data transfer between the two, an improvement over its predecessor 8 increase or multiply scale-out bandwidth, effectively solving the communication congestion problem in high-density clusters.

In terms of architecture, the MI400 series adopts the open-standard UALink technology, which is different from NVIDIA's proprietary interconnect solution. The technology enables high-speed connectivity between GPUs through Ethernet tunnels, supports unified compute resource pooling at the rack level, and works in conjunction with the open architecture of the OCP and the Ultra Ethernet Consortium to ensure compatibility with the existing data center infrastructure. The open architecture of OCP and Ultra Ethernet Consortium ensures compatibility with existing data center infrastructures. ExpectedThe MI400 will be available in 2026.

apart fromOpenAIBeyond that, Microsoft, Oracle (Oracle),Meta,xAIet al. (and other authors)7oldestAIThe development platform is working withAMDCooperating in the use of theirAIChip.

Oraclewill be among the first to adoptInstinct MI355XOne of the industry leaders in rack-level solutions for drive, highlighting theOracleOffering the widest range ofAIInfrastructure Commitment. Oracle Cloud Infrastructure supports a wide range of mission-critical enterprise workloads with stringent requirements for scalability, reliability, security and performance.

OracleExecutive Vice President, Cloud InfrastructureMahesh Thiagarajansaid.OracleCloud infrastructures continue to evolve from a relationship withAMDbenefit from the strategic partnership. We will be among the first to offer access toEPYC,Instinctcap (a poem)PensandoJoint Powers of theMI355XOne of the companies with rack-level infrastructure.

We've seen our customers' interest inAMDThe impressive adoption of bare-metal instances driven by this highlights how easy it is for customers to adopt and scale theirAIworkloads. In addition.OracleWidely relied upon internally for its own workloads and externally in customer-facing applicationsAMDTechnology. We plan to continue to work on multipleAMDDeep cooperation in the product generation and theAMDroadmap and its ability to consistently meet expectations with confidence.

Launch Replay:
https://www.youtube.com/watch?v=5dmFa9iXPWI