Networking is the third pillar of AMD's Advancing AI 2026 rack-scale strategy, and today the company detailed the AMD Pensando silicon and fabric technology that connects the Helios rack. As Helios is AMD's first rack-scale design, the company encountered challenges in interconnecting many GPUs, memories, and data movement speeds within a single system, both for scaling up and scaling out for multi-rack connections inside the data center. Current AI infrastructure must closely follow model development as AI model sizes have grown exponentially, from the original 65-million-parameter Transformer in 2017 to trillion-parameter GPT-4-class systems and now 10-trillion-plus parameter agentic and reasoning models. To run a model with more than 10 trillion parameters, you need to connect dozens of GPUs with high-speed VRAM, as well as a high-speed interconnect to split model weights effectively so latency isn't an issue. The modern interconnect of choice is again the plain-old Ethernet, not proprietary interconnects, which have become the standard for AI networking.

AMD's solution spans three layers: a third-generation Pensando "Salina" DPU on the front end, a first-generation UALink-over-Ethernet (UALoE) scale-up fabric inside the rack, and a second-generation Pensando "Vulcano" AI NIC for scale-out. All of this is managed by AMD's Helios management software and Fabric Manager. AMD outlined challenges that bandwidth needs are doubling in less than two years driven by massive AI models growing at a rapid pace, and that the hardware iteration cycles need to account for intelligence in the data fabric layer, not only compute layer.

AMD used its Advancing AI 2026 keynote today to launch the Instinct MI400 Series, its next-generation data-center GPU family, alongside the Helios rack-scale solution and the CDNA 5 architecture that underpins both. CDNA 5 is AMD's first GPU architecture to reach the 2 nm node, with the flagship Instinct MI455X carrying 320 billion transistors across a multi-chiplet package and 432 GB of HBM4. This is a data-center compute architecture, not a Radeon gaming architecture, though some of the design choices in CDNA 5 point to where AMD's future unified gaming/compute architecture could be headed.

The MI400 Series splits into two parts: the Instinct MI455X for frontier AI and AI-factory deployments, and the Instinct MI430X for sovereign AI and HPC. Together with the Helios rack, 6th Gen EPYC "Venice" CPUs and AMD Pensando networking, they form AMD's open answer to NVIDIA's Vera Rubin NVL72. This post also covers the new ROCm.ai software stack and AMD's "Gorgon Halo" client teaser.

The rapid growth of GenAI and LLM workloads, from agentic workflows and multi-step tool use to retrieval-augmented reasoning, demands inference infrastructure that is fast, adaptable, and highly optimized. While AMD continues to invest in general-purpose inference frameworks such as vLLM and SGLang, ATOM provides the most direct path to peak Instinct MI355X GPU performance for modern reasoning and MoE-heavy workloads, which increasingly dominate frontier LLM architectures.

Over the past months, AMD have implemented numerous optimizations to improve both single-node performance and multi-node distributed inference for DeepSeek-R1 on the MI355X GPU.

Super Micro Computer, Inc. (SMCI), a Total IT Solution Provider for AI, Cloud, Storage, and 5G/Edge, is announcing the latest addition to the AMD-based Instinct MI350 series GPU optimized solutions, delivering unparalleled performance, maximum scalability, and power efficiency. Supermicro has designed this new system for organizations that need the high-end performance of AMD Instinct MI355X GPUs but require an air-cooled environment.

"Supermicro leads the industry with the most experience in delivering high-performance solutions in AI and HPC for our customers," said Vik Malyala, SVP Technology and AI at Supermicro. "Supermicro's DCBBS enables us to quickly incorporate AMD solutions by deploying advanced technology with proven data center solutions to market. The addition of the new air-cooled AMD Instinct MI355X GPU systems expands and strengthens our portfolio of AI solutions and gives customers more choices as they build next-generation data centers."

The latest MLPerf 5.1 Training results mark an important milestone—the first MLPerf training submission using AMD Instinct MI350 Series GPUs. These new benchmarks demonstrate breakthrough generational performance gains and highlight broad ecosystem participation across some of today's most demanding AI training workloads.

Building on the strong MLPerf 5.1 Inference results released earlier this year, this submission represents the first time the AMD Instinct MI350 Series—including both the MI355X and MI350X GPUs—have been publicly benchmarked for AI training. The results show clear progress in scalability, efficiency, and compute performance, underscoring how the AMD Instinct MI350 Series is accelerating the development of next-generation generative AI models.

At its Financial Analyst Day 2025, AMD shared new details about its upcoming Instinct MI400 and MI500 AI accelerators. The Instinct MI400 series set to launch in 2026, will be based on the new CDNA 5 architecture. AMD claims it will deliver up to 40 FP4 and 20 FP8 PFLOPs, roughly twice the compute performance of the current MI350. The GPUs also move to HBM4 memory from HMB3e, increasing capacity from 288 GB to 432 GB and raising total bandwidth from 8 TB/s to 19.6 TB/s. Each GPU provides 300 GB/s of scale-out bandwidth and adds broader AI data format support along with expanded AI pipelines. Two models are planned: the Instinct MI455X for large-scale AI training and inference, and the MI430X for HPC and sovereign AI. The MI430X features FP64 hardware based support and hybrid CPU+GPU compute capability, while retaining the same HBM4 memory subsystem as the MI455X. The MI400 series also moves away from the CoWoS-S to the CoWoS-L (Local Silicon Interconnect) advanced packaging technology.

AMD compares the MI400 lineup to NVIDIA's upcoming Vera Rubin series, claiming similar compute performance and memory bandwidth, but 1.5× higher memory capacity and scale-out bandwidth. During her presentation, Dr. Lisa Su said that AMD's Instinct MI400 accelerators will be available to clients and partners from day one, while the next-generation MI500 is already in advanced design stages. AMD is also now moving to an annual refresh cycle for its Instinct products, confirming that its next-gen MI500 series will be launched in 2027.

AMD today unveiled ROCm 7.0, a massive update to its open GPU software platform for AI workloads across datacenter racks and even client devices. The release focuses on support for the new Instinct MI350 accelerator series, built on CDNA 4. It pairs hardware-level refinements, including advanced HBM utilization and expanded datatype handling, with software improvements designed to enhance throughput for both training and inference. AMD classifies this release as a generational step that widens deployment choices for enterprises, from single-node servers to multi-rack clusters. Company benchmarks shared with partners indicate multi-fold gains over previous ROCm 6.0 versions on select large-model tasks, and AMD shows tighter integration with open-source serving stacks to help teams transition from prototype to production more quickly.

Under the hood, ROCm 7.0 delivers algorithmic and systems advances intended to reduce cost and improve utilization: enhanced attention and reasoning primitives, sparse mixture-of-experts techniques, and support for FP4, FP6, and FP8 precision formats that trade a bit of accuracy for massive memory savings. ROCm 7.0 also focuses on scale and manageability by enabling GPU-direct communications, distributed inference across nodes with vLLM-d and DeepEP compatibility, and cluster tools such as AMD Resource Manager and a new AI Workbench for orchestration and monitoring. To smooth adoption, AMD provides prebuilt Docker images with quantized model examples and turnkey stacks for common frameworks. AMD is also expanding endpoint support to Ryzen AI processors and Radeon RX series GPUs, so developers can deploy consistent pipelines from edge devices to cloud racks. ROCm 7 is available now through AMD's developer resources, with documentation, examples, and enterprise support to help teams get started.

Micron Technology, Inc. today announced the integration of its HBM3E 36 GB 12-high offering into the upcoming AMD Instinct MI350 Series solutions. This collaboration highlights the critical role of power efficiency and performance in training large AI models, delivering high-throughput inference and handling complex HPC workloads such as data processing and computational modeling. Furthermore, it represents another significant milestone in HBM industry leadership for Micron, showcasing its robust execution and the value of its strong customer relationships.

Micron HBM3E 36 GB 12-high solution brings industry-leading memory technology to AMD Instinct MI350 Series GPU platforms, providing outstanding bandwidth and lower power consumption. The AMD Instinct MI350 Series GPU platforms, built on AMD advanced CDNA 4 architecture, integrate 288 GB of high-bandwidth HBM3E memory capacity, delivering up to 8 TB/s bandwidth for exceptional throughput. This immense memory capacity allows Instinct MI350 series GPUs to efficiently support AI models with up to 520 billion parameters—on a single GPU. In a full platform configuration, Instinct MI350 Series GPUs offers up to 2.3 TB of HBM3E memory and achieves peak theoretical performance of up to 161 PFLOPS at FP4 precision, with leadership energy efficiency and scalability for high-density AI workloads. This tightly integrated architecture, combined with Micron's power-efficient HBM3E, enables exceptional throughput for large language model training, inference and scientific simulation tasks—empowering data centers to scale seamlessly while maximizing compute performance per watt. This joint effort between Micron and AMD has enabled faster time to market for AI solutions.

Tomorrow evening, AMD will host its "Advancing AI" livestream to introduce the Instinct MI350 series, a new line of GPU accelerators designed for large-scale AI training and inference. First shown in prototype form at ISC 2025 in Hamburg just a day ago, each MI350 card features 288 GB of HBM3E memory, delivering up to 8 TB/s of sustained bandwidth. Customers can choose between the single-card MI350X and the higher-clocked MI355X or opt for a full eight-GPU platform that aggregates to over 2.3 TB of memory. Both chips are built on the CDNA 4 architecture, which now supports four different precision formats: FP16, FP8, FP6, and FP4. The addition of FP6 and FP4 is designed to boost throughput in modern AI workloads, where models of tomorrow with tens of trillions of parameters are trained on FP6 and FP4.

In half-precision tests, the MI350X achieves 4.6 PetaFLOPS on its own and 36.8 PetaFLOPS in eight-GPU platform form, while the MI355X surpasses those numbers, reaching 5.03 PetaFLOPS and just over 40 PetaFLOPS. AMD is also aiming to improve energy efficiency by a factor of thirty compared with its previous generation. The MI350X card runs within a 1,000 Watt power envelope and relies on air cooling, whereas the MI355X steps up to 1,400 Watts and is intended for direct-liquid cooling setups. That 400 Watt increase puts it right at NVIDIA's upcoming GB300 "Grace Blackwell Ultra" superchip, which is also a 1,400 W design. With memory capacity, raw computing, and power efficiency all pushed to new heights, the question remains whether real-world benchmarks will match these ambitious specifications. AMD now only lacks platform scaling beyond eight GPUs, which the Instinct MI400 series will address.

Last September, AMD leadership publicly revealed UDNA—an "unforking" of previously separate enterprise and commercial GPU branches. Not long after this announcement, TechPowerUp's resident Serbian correspondent—AleksandarK—sat down with Team Red's Andrej Zdravkovic. The Chief Software Officer (and SVP) stated that a fair chunk of UDNA-related development work would be done by local engineers. Zdravkovic discussed this technology's eventual deployment in futuristic "AI PCs," but gamers have been salivating at the prospect of a proper successor to RDNA 4. A next-gen graphics architecture seeker—MrMPFR—has combed through official documents for any sign of UDNA preview material. The noted /Hardware subreddit member managed to distill their initial (very long) set of findings into an "easily digestible overview." They stated that this was just a small case of: "reporting and a little analysis on AMD's publicly available US patents filings," and other public-facing resources/archives.

Gleaned information included: "finalized architectural characteristics in future RDNA generations, AMD DXR IHV stacks (driver agnostic), and AMD sponsored titles. But please take everything with a grain of salt given my lack of professional expertise and experience with Real-time ray tracing (RTRT)". MrMPFR believes that Team Red started picking up former NVIDIA and Intel engineering talent, back in 2022/2023. In addition, a lot of new hires were apparently sourced from academic institutions. In theory, these newer team members have not had the time to make major inroads—in terms of getting finalized products out into the wild. MrMPFR reckons that noticeable contributions will accelerate AMD's making of "RDNA 6+/UDNA 2+," and beyond. Early 2025 leaks have pointed to the company collaborating with Sony; their "PlayStation 6" console is tipped to be powered by some fork of Team Red's "UDNA" graphics technology.

AMD's Instinct MI355X accelerators for AI workloads are gaining traction, and Oracle just became one of the bigger customers. According to Oracle's latest financial results, the company noted that it had acquired 30,000 AMD Instinct MI355X accelerators. "In Q3, we signed a multi billion dollar contract with AMD to build a cluster of 30,000 of their latest MI355X GPUs," noted Larry Ellison, adding that "And all four of the leading cloud security companies, CrowdStrike, Cyber Reason, Newfold Digital and Palo Alto, they all decided to move to the Oracle Cloud. But perhaps most importantly, Oracle has developed a new product called the AI data platform that enables our huge install base of database customers to use the latest AI models from OpenAI, XAI and Meta to analyze all of the data they have stored in their millions of existing Oracle databases. By using Oracle version 23 AI's vector capabilities, customers can automatically put all of their existing data into the vector format that is understood by AI models. This allows those AI models to learn, understand and analyze every aspect of your company or government agency, instantly unlocking the value in your data while keeping your data private and secure."

AMD's Instinct MI355X accelerator introduces the CDNA4 architecture on TSMC's N3 process node with a focus on AI workload acceleration. The chiplet-based GPU delivers 2.3 petaflops of FP16 compute and 4.6 petaflops of FP8 compute, marking a 77% performance increase over the MI300X series. The MI355X's key advancement comes through support for reduced-precision FP4 and FP6 numerical formats, enabling up to 9.2 petaflops of FP4 compute. Memory specifications include 288 GB of HBM3E across eight stacks, providing 8 TB/s of total bandwidth. Production timelines place the MI355X's market entry in the second half of 2025, continuing AMD's annual cadence for data center GPU launches. By second half, Oracle will likely prepare data center space for these GPUs and just power them on once AMD ships these accelerators.

While the current generation of AMD graphics cards employs RDNA 3 at its core, and the upcoming RX 8000 series will feature RDNA 4, the latest leaks suggest RDNA 5 is not in development. Instead, UDNA will succeed RDNA 4, simplifying AMD's GPU roadmap. A credible source on the Chiphell forums, zhangzhonghao, reports that the UDNA-based RX 9000 series and Instinct MI400 AI accelerator will incorporate the same advanced Arithmetic Logic Unit (ALU) designs in both products, reminiscent of AMD's earlier GCN architectures before the CDNA and RDNA split. Sony's next-generation PlayStation 6 is also rumored to adopt UDNA technology. The PS5 and PS5 Pro currently utilize RDNA 2, while the Pro variant integrates elements of RDNA 4 for enhanced ray tracing. The PS6's CPU configuration remains unclear, but speculation revolves around Zen 4 or Zen 5 architectures.

The first UDNA gaming GPUs are expected to enter production by Q2 2026. Interestingly, AMD's RDNA 4 GPUs are anticipated to focus on entry-level to mid-range markets, potentially leaving high-end offerings until the UDNA generation. This strategic pause may allow AMD to refine AI-accelerated technologies like FidelityFX Super Resolution (FSR) 4, aiming to compete with NVIDIA's DLSS. This unification is inspired by NVIDIA's CUDA ecosystem, which supports cross-platform compatibility from laptops to high-performance servers. As AMD sees it, the decision addresses the challenges posed by maintaining separate architectures, which complicate memory subsystem optimizations and hinder forward and backward compatibility. Putting developer resources into RDNA 5 is not economically or strategically wise, given that UDNA is about to take over. Additionally, the company is enabling ROCm software support across all products ranging from consumer Radeon to enterprise Instinct MI. Accelerating software for one platform will translate to the entire product stack.

Giga Computing, a subsidiary of GIGABYTE and an industry leader in generative AI servers and advanced cooling technologies, today announced support for AMD EPYC 9005 Series processors with the release of new GIGABYTE servers alongside BIOS updates for some existing GIGABYTE servers using the SP5 platform. This first wave of updates supports over 60 servers and motherboards that customers can choose from that deliver exceptional performance for 5th Generation AMD EPYC processors. In addition, with the launch of the AMD Instinct MI325X accelerator, a newly designed GIGABYTE server was created, and it will be showcased at SC24 (Nov. 19-21) in Atlanta.

New GIGABYTE Servers and Updates
To fill in all possible workload scenarios, using modular design servers to edge servers to enterprise-grade motherboards, these new solutions will ship already supporting AMD EPYC 9005 Series processors. The XV23-ZX0 is one of the many new solutions and it is notable for its modularized server design using two AMD EPYC 9005 processors and supporting up to four GPUs and three additional FHFL slots. It also has 2+2 redundant power supplies on the front-side for ease of access.

During its "Advancing AI" conference today, AMD has updated its AI accelerator portfolio with the Instinct MI325X accelerator, designed to succeed its MI300X predecessor. Built on the CDNA 3 architecture, Instinct MI325X brings a suite of improvements over the old SKU. Now, the MI325X features 256 GB of HBM3E memory running at 6 TB/s bandwidth. The capacity memory alone is a 1.8x improvement over the old MI300 SKU, which features 192 GB of regular HBM3 memory. Providing more memory capacity is crucial as upcoming AI workloads are training models with parameter counts measured in trillions, as opposed to billions with current models we have today. When it comes to compute resources, the Instinct MI325X provides 1.3 PetaFLOPS at FP16 and 2.6 PetaFLOPS at FP8 training and inference. This represents a 1.3x improvement over the Instinct MI300.

A chip alone is worthless without a good platform, and AMD decided to make the Instinct MI325X OAM modules a drop-in replacement for the current platform designed for MI300X, as they are both pin-compatible. In systems packing eight MI325X accelerators, there are 2 TB of HBM3E memory running at 48 TB/s memory bandwidth. Such a system achieves 10.4 PetaFLOPS of FP16 and 20.8 PetaFLOPS of FP8 compute performance. The company uses NVIDIA's H200 HGX as reference claims for its performance competitiveness, where the company claims that the Instinct MI325X outperforms NVIDIA H200 HGX system by 1.3x across the board in memory bandwidth, FP16 / FP8 compute performance and 1.8x in memory capacity.

A few days ago, we reported on AMD's newest expansion plans for Serbia. The company opened two new engineering design centers with offices in Belgrade and Nis. We were invited to join the opening ceremony and got an exclusive interview with one of AMD's top executives, Andrej Zdravkovic, who is the senior vice president and Chief Software Officer. Previously, we reported on AMD's transition to become a software company. The company has recently tripled its software engineering workforce and is moving some of its best people to support these teams. AMD's plan is spread over a three to five-year timeframe to improve its software ecosystem, accelerating hardware development to launch new products more frequently and to react to changes in software demand. AMD found that to help these expansion efforts, opening new design centers in Serbia would be very advantageous.

We sat down with Andrej Zdravkovic to discuss the purpose of AMD's establishment in Serbia and the future of some products. Zdravkovic is actually an engineer from Serbia, where he completed his Bachelor's and Master's degrees in electrical engineering from Belgrade University. In 1998, Zdravkovic joined ATI and quickly rose through the ranks, eventually becoming a senior director. During his decade-long tenure, Zdravkovic witnessed a significant industry shift as AMD acquired ATI in 2006. After a brief stint at another company, Zdravkovic returned to AMD in 2015, bringing with him a wealth of experience and a unique perspective on the evolution of the graphics and computing industry.Here is the full interview:

According to new information from Tom's Hardware, AMD has announced plans to unify its consumer-focused gaming RDNA and data center CDNA graphics architectures into a single, unified design called "UDNA." The announcement was made by AMD's Jack Huynh, Senior Vice President and General Manager of the Computing and Graphics Business Group, at IFA 2024 in Berlin. The goal of the new UDNA architecture is to provide a single focus point for developers so that each optimized application can run on consumer-grade GPU like Radeon RX 7900XTX as well as high-end data center GPU like Instinct MI300. This will create a unification similar to NVIDIA's CUDA, which enables CUDA-focused developers to run applications on everything ranging from laptops to data centers.

Jack Huynh, AMDSo, part of a big change at AMD is today we have a CDNA architecture for our Instinct data center GPUs and RDNA for the consumer stuff. It's forked. Going forward, we will call it UDNA. There'll be one unified architecture, both Instinct and client [consumer]. We'll unify it so that it will be so much easier for developers versus today, where they have to choose and value is not improving.

NVIDIA has reportedly been hit with a US Department of Justice (DOJ) antitrust probe over the tactics the company allegedly employs to sell or lease its AI GPUs and data-center networking equipment, "The Information" reported. Shares of the NVIDIA stock fell 3.6% in the pre-market trading on Friday (08/02). The main complainants behind the probe appear to be a special interest group among the customers of AI GPUs, and not NVIDIA's competitors in the AI GPU industry per se. US Senator Elizabeth Warren and US progressives have been most vocal about calling upon the DOJ to investigate antitrust allegations against NVIDIA.

Meanwhile, US officials are reportedly reaching out to NVIDIA's competitors, including AMD and Intel, to gather information about the complaints. NVIDIA holds 80% of the AI GPU market, while AMD, and to a much lesser extent, Intel, have received spillover demand for AI GPUs. "The Information" report says that the complaint alleges NVIDIA pressured cloud customers to buy "multiple products". We don't know what this means, one theory holds that NVIDIA is getting them to commit to buying multiple generations of products (eg: Ampere, Hopper, and over to Blackwell); while another holds that it's getting them to buy multiple kinds of products, which include not just the AI GPUs, but also NVIDIA's first-party server systems and networking equipment. Yet another theory holds that it is bundle first-party software and services to go with the hardware, far beyond the basic software needed to get the hardware to work.

Just a few weeks ago, AMD invited us to Barcelona as part of a roundtable, to share their vision for the future of the company, and to get our feedback. On site, were prominent AMD leadership, including Phil Guido, Executive Vice President & Chief Commercial Officer and Jack Huynh, Senior VP & GM, Computing and Graphics Business Group. AMD is making changes in a big way to how they are approaching technology, shifting their focus from hardware development to emphasizing software, APIs, and AI experiences. Software is no longer just a complement to hardware; it's the core of modern technological ecosystems, and AMD is finally aligning its strategy accordingly.

The major difference between AMD and NVIDIA is that AMD is a hardware company that makes software on the side to support its hardware; while NVIDIA is a software company that designs hardware on the side to accelerate its software. This is about to change, as AMD is making a pivot toward software. They believe that they now have the full stack of computing hardware—all the way from CPUs, to AI accelerators, to GPUs, to FPGAs, to data-processing and even server architecture. The only frontier left for AMD is software.

A curiously described graphics card was detected by Olrak29 as it was making it through international shipping. The shipment description for the card reads "GRAPHIC CARD NAVI48 G28201 DT XTX REVB-PRE-CORRELATION AO PLATSI TT(SAMSUNG)-Q2 2024-3A-102-G28201." This can be decoded as a graphics card with the board number "G28201," for the desktop platform. It features a maxed out version of the "Navi 48" silicon, and is based on the B revision of the PCB. It features Samsung-made memory chips, and is dated Q2-2024.

AMD is planning to retreat from the enthusiast segment of gaming graphics cards with the RDNA 4 generation. The company originally entered this segment with the RX 6800 series and RX 6900 series RDNA 2 generation, where it saw unexpected success with the crypto-mining market boom, besides being competitive with the RTX 3080 and RTX 3090. This bust by the time RDNA 3 and the RX 7900 series arrived, and the chip wasn't competitive with NVIDIA's top-end. Around this time, the AI acceleration boom squeezed foundry allocation of all major chipmakers, including AMD, making large chips based on the latest process nodes even less viable for a market such as enthusiast graphics—the company would rather make CDNA AI accelerators with its allocation. Given all this, the company's fastest GPUs from the RDNA 4 generation could be the ones that succeed the current RX 7800 XT and RX 7700 XT, so AMD could capture a slice of the performance segment.

AMD has quietly added some interesting codenames to its ROCm hardware support list. The biggest surprise is the appearance of "RDNA 4" and "Navi 44" codenames, hinting at a potential successor to the current RDNA 3 GPU architecture powering AMD's Radeon RX 7000 series graphics cards. The upcoming Radeon RX 8000 series could see Navi 44 SKU with a codename "gfx1200". While details are scarce, the inclusion of RDNA 4 and Navi 44 in the ROCm list suggests AMD is working on a new GPU microarchitecture that could bring significant performance and efficiency gains. While RDNA 4 may be destined for future Radeon gaming GPUs, in the data center GPU compute market, AMD is preparing a CDNA 4 based successors to the MI300 series. However, it appears that we haven't seen all the MI300 variants first. Equally intriguing is the "MI300X1" codename, which appears to reference an upcoming AI-focused accelerator from AMD.

While we wait for more information, we can't decipher whether the Navi 44 GPU SKU is for the high-end or low-end segment. If previous generations are for reference, then the Navi 44 SKU would target the low end of the GPU performance spectrum. The previous generation RDNA 3 had Navi 33 as an entry-level model, whereas the RDNA 2 had a Navi 24 SKU for entry-level GPUs. We have reported on RDNA 4 merely being a "bug correction" generation to fix the perf/Watt curve and offer better efficiency overall. What happens finally, we have to wait and see. AMD could announce more details in its upcoming Computex keynote.

The TinyBox AI server system attracted plenty of media attention last week—its creator, George Hotz, decided to build with AMD RDNA 3.0 GPU hardware rather than the expected/traditional choice of CDNA 3.0. Tiny Corp. is a startup firm dealing in neural network frameworks—they currently "write and maintain tinygrad." Hotz & Co. are in the process of assembling rack-mounted 12U TinyBox systems for customers—an individual server houses an AMD EPYC 7532 processor and six XFX Speedster MERC310 Radeon RX 7900 XTX graphics cards. The Tiny Corp. social media account has engaged in numerous NVIDIA vs. AMD AI hardware debates/tirades—Hotz appears to favor the latter, as evidenced in his latest choice of components. ROCm support on Team Red AI Instinct accelerators is fairly mature at this point in time, but a much newer prospect on gaming-oriented graphics cards.

Tiny Corporation's unusual leveraging of Radeon RX 7900 XTX GPUs in a data center configuration has already hit a developmental roadblock. Yesterday, the company's social media account expressed driver-related frustrations in a public forum: "If AMD open sources their firmware, I'll fix their LLVM spilling bug and write a fuzzer for HSA. Otherwise, it's not worth putting tons of effort into fixing bugs on a platform you don't own." Hotz's latest complaint was taken onboard by AMD's top brass—Dr. Lisa Su responded with the following message: "Thanks for the collaboration and feedback. We are all in to get you a good solution. Team is on it." Her software engineers—within a few hours—managed to fling out a set of fixes in Tiny Corporation's direction. Hotz appreciated the quick turnaround, and proceeded to run a model without encountering major stability issues: "AMD sent me an updated set of firmware blobs to try. They are responsive, and there have been big strides in the driver in the last year. It will be good! This training run is almost 5 hours in, hasn't crashed yet." Tiny Corp. drummed up speculation about AMD open sourcing GPU MES firmware—Hotz disclosed that he will be talking (on the phone) to Team Red leadership.

Today, AMD announced the availability of the AMD Instinct MI300X accelerators - with industry leading memory bandwidth for generative AI and leadership performance for large language model (LLM) training and inferencing - as well as the AMD Instinct MI300A accelerated processing unit (APU) - combining the latest AMD CDNA 3 architecture and "Zen 4" CPUs to deliver breakthrough performance for HPC and AI workloads.

"AMD Instinct MI300 Series accelerators are designed with our most advanced technologies, delivering leadership performance, and will be in large scale cloud and enterprise deployments," said Victor Peng, president, AMD. "By leveraging our leadership hardware, software and open ecosystem approach, cloud providers, OEMs and ODMs are bringing to market technologies that empower enterprises to adopt and deploy AI-powered solutions."

GIGABYTE Technology: Giga Computing, a subsidiary of GIGABYTE and an industry leader in high-performance servers, and IT infrastructure, today announced the GIGABYTE G383-R80 for the AMD Instinct MI300A APU and two GIGABYTE G593 series servers for the AMD Instinct MI300X GPU and AMD EPYC 9004 Series processor. As a testament to the performance of AMD Instinct MI300 Series family of products, the El Capitan supercomputer at Lawrence Livermore National Laboratory uses the MI300A APU to power exascale computing. And these new GIGABYTE servers are the ideal platform to propel discoveries in HPC & AI at exascale.

Marrying of a CPU & GPU: G383-R80
For incredible advancements in HPC there is the GIGABYTE G383-R80 that houses four LGA6096 sockets for MI300A APUs. This chip integrates a CPU that has twenty-four AMD Zen 4 cores with a powerful GPU built with AMD CDNA 3 GPU cores. And the chiplet design shares 128 GB of unified HBM3 memory for impressive performance for large AI models. The G383 server has lots of expansion slots for networking, storage, or other accelerators, with a total of twelve PCIe Gen 5 slots. And in the front of the chassis are eight 2.5" Gen 5 NVMe bays to handle heavy workloads such as real-time big data analytics and latency-sensitive workloads in finance and telecom.

When Lawrence Livermore National Laboratory (LLNL) announced the creation of a two-ExaFLOP supercomputer named El Capitan, we heard that AMD would power it with its Instinct MI300 accelerator. Today, LNLL published a Tweet that states, "We've begun receiving & installing components for El Capitan, @NNSANews' first #exascale #supercomputer. While we're still a ways from deploying it for national security purposes in 2024, it's exciting to see years of work becoming reality." As published images show, HPE racks filled with AMD Instinct MI300 are showing up now at LNLL's facility, and the supercomputer is expected to go operational in 2024. This could mean that November 2023 TOP500 list update wouldn't feature El Capitan, as system enablement would be very hard to achieve in four months until then.

The El Capitan supercomputer is expected to run on AMD Instinct MI300A accelerator, which features 24 Zen4 cores, CDNA3 architecture, and 128 GB of HBM3 memory. All paired together in a four-accelerator configuration goes inside each node from HPE, also getting water cooling treatment. While we don't have many further details on the memory and storage of El Capitan, we know that the system will exceed two ExFLOPS at peak and will consume close to 40 MW of power.

AMD recently revealed its Instinct MI300X GPU at their Data Center and AI Technology Premiere event on Tuesday (June 15). The keynote presentation did not provide any details about the new accelerator model's power consumption, but that did not stop one tipster - Hoang Anh Phu - from obtaining this information from Team Red's post-event footnotes. A comparative observation was made: "MI300X (192 GB HBM3, OAM Module) TBP is 750 W, compared to last gen, MI250X TBP is only 500-560 W." A leaked Giga Computing roadmap from last month anticipated server-grade GPUs hitting the 700 W mark.

NVIDIA's Hopper H100 took the crown - with its demand for a maximum of 700 W - as the most power-hungry data center enterprise GPU until now. The MI300X's OCP Accelerator Module-based design now surpasses Team Green's flagship with a slightly greater rating. AMD's new "leadership generative AI accelerator" sports 304 CDNA 3 compute units, which is a clear upgrade over the MI250X's 220 (CDNA 2) CUs. Engineers have also introduced new 24G B HBM3 stacks, so the MI300X can be specced with 192 GB of memory (as a maximum), the MI250X is limited to a 128 GB memory capacity with its slower HBM2E stacks. We hope to see sample units producing benchmark results very soon, with the MI300X pitted against H100.