The Channel Company

Sponsored by:

Visit AMD Visit Supermicro

Performance Intensive Computing

Capture the full potential of IT

AMD Instinct MI300A blends GPU, CPU for super-speedy AI/HPC

Featured content

AMD Instinct MI300A blends GPU, CPU for super-speedy AI/HPC

CPU or GPU for AI and HPC? You can get the best of both with the AMD Instinct MI300A.

Learn More about this topic
  • Applications:
  • Featured Technologies:

The AMD Instinct MI300A is the world’s first data center accelerated processing unit for high-performance computing and AI. It does this by integrating both CPU and GPU cores on a single package.

That makes the AMD Instinct MI300A highly efficient at running both HPC and AI workloads. It also makes the MI300A powerful enough to accelerate training the latest AI models.

Introduced about a year ago, the AMD Instinct MI300A accelerator is shipping soon. So are two Supermicro servers—one a liquid-cooled 2U system, the other an air-cooled 4U—each powered by four MI300A units.

Under the Hood

The technology of the AMD Instinct MI300A is impressive. Each MI300A integrates 24 AMD ‘Zen 4’ x86 CPU cores with 228 AMD CDNA 3 high-throughput GPU compute units.

You also get 128GB of unified HBM3 memory. This presents a single shared address space to CPU and GPU, all of which are interconnected into the coherent 4th Gen AMD Infinity architecture.

Also, the AMD Instinct MI300A is designed to be used in a multi-unit configuration. This means you can connect up to four of them in a single server.

To make this work, each APU has 1 TB/sec. of bidirectional connectivity through eight 128 GB/sec. AMD Infinity Fabric interfaces. Four of the interfaces are dedicated Infinity Fabric links. The other four can be flexibly assigned to deliver either Infinity Fabric or PCIe Gen 5 connectivity.

In a typical four-APU configuration, six interfaces are dedicated to inter-GPU Infinity Fabric connectivity. That supplies a total of 384 GB/sec. of peer-to-peer connectivity per APU. One interface is assigned to support x16 PCIe Gen 5 connectivity to external I/O devices. In addition, each MI300A includes two x4 interfaces to storage, such as M.2 boot drives, plus two USB Gen 2 or 3 interfaces.

Converged Computing

There’s more. The AMD Instinct MI300A was designed to handle today’s convergence of HPC and AI applications at scale.

To meet the increasing demands of AI applications, the APU is optimized for widely used data types. These include FP64, FP32, FP16, BF16, TF32, FP8 and INT8.

The MI300A also supports native hardware sparsity for efficiently gathering data from sparse matrices. This saves power and compute cycles, and it also lowers memory use.

Another element of the design aims at high efficiency by eliminating time-consuming data copy operations. The MI300A can easily offload tasks easily between the CPU and GPU. And it’s all supported by AMD’s ROCm 6 open software platform, built for HPC, AI and machine learning workloads.

Finally, virtualized environments are supported on the MI300A through SR-IOV to share resources with up to three partitions per APU. SR-IOV—short for single-root, input/output virtualization—is an extension of the PCIe spec. It allows a device to separate access to its resources among various PCIe functions. The goal: improved manageability and performance.

Fun fact: The AMD Instinct MI300A is a key design component of the El Capitan supercomputer recently dedicated by Lawrence Livermore Labs. This system can process over two quintillion (1018) calculations per second.

Supermicro Servers

As mentioned above, Supermicro now offers two server systems based on the AMD Instinct MI300A APU. They’re 2U and 4U systems.

These servers both take advantage of AMD’s integration features by combining four MI300A units in a single system. That gives you a total of 912 GPUs, 96 CPUs, and 512GB of HBM3 memory.

Supermicro says these systems can push HPC processing to Exascale levels, meaning they’re very, very fast. “Flop” is short for floating point operations per second, and “exa” indicates a 1 with 18 zeros after it. That’s fast.

Supermicro’s 2U server (model number AS -2145GH-TNMR-LCC) is liquid-cooled and aimed at HPC workloads. Supermicro says direct-to-chip liquid-cooling technology enables a nice TCO with over 51% data center energy cost savings. The company also cites a 70% reduction in fan power usage, compared with air-cooled solutions.

If you’re looking for big HPC horsepower, Supermicro’s got your back with this 2U system. The company’s rack-scale integration is optimized with dual AIOM (advanced I/O modules) and 400G networking. This means you can create a high-density supercomputing cluster with as many as 21 of Supermicro’s 2U systems in a 48U rack. With each system combining four MI300A units, that would give you a total of 84 APUs.

The other Supermicro server (model number AS -4145GH-TNMR) is an air-cooled 4U system, also equipped with four AMD Instinct MI300A accelerators, and it’s intended for converged HPC-AI workloads. The system’s mechanical airflow design keeps thermal throttling at bay; if that’s not enough, the system also has 10 heavy-duty 80mm fans.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

Tech Explainer: CPUs and GPUs for AI training and inferencing

Featured content

Tech Explainer: CPUs and GPUs for AI training and inferencing

Which is best for AI – a CPU or a GPU? Like much in life, it depends.

Learn More about this topic
  • Applications:
  • Featured Technologies:

While central processing units and graphics processing units serve different roles in AI training and inferencing, both roles are vital to AI workloads.

CPUs and GPUs were both invented long before the AI era. But each has found new purpose as the robots conduct more of our day-to-day business.

Each has its tradeoffs. Most CPUs are less expensive than GPUs, and they typically require less electric power. But that doesn’t mean CPUs are always the best choice for AI workloads. Like lots of things in life, it depends.

Two Steps to AI

A typical AI application involves a two-step process. First training. Then inferencing.

Before an AI model can be deployed, it must be trained. That could include suggesting which movie to watch next on Netflix or detecting fake currency in a retail environment.

Once the AI model has been deployed, it can begin the inferencing process. In this stage, the AI application interfaces with users, devices and other models. Then it autonomously makes predictions and decisions based on new input.

For example, Netflix’s recommendation engine is powered by an AI model. The AI was first trained to consider your watching history and stated preferences, as well as to review newly available content. Then the AI employs inferencing—what we might call reasoning—to suggest a new movie or TV show you’re likely to enjoy.

AI Training

GPU architectures like those found in the AMD Instinct MI325X accelerator offers highly parallel processing. In other words, a GPU can perform many calculations simultaneously.

The AMD Instinct MI325X has more than 300 GPU compute units. They make the accelerator faster and more adept at both processing large datasets and handling the repetitious numerical operations common to the training process.

These capabilities also mean GPUs can accelerate the training process. That’s especially true for large models, such as those that underpin the networks used for deep learning.

CPUs, by contrast, excel at general-purpose tasks. Compared with a GPU, a CPU will be better at completing sequential tasks that require logic or decision-making. For this reason, a CPU’s role in AI training is mostly limited to data preprocessing and coordinating GPU tasks.

AI Inferencing

However, when it comes to AI inferencing, CPUs play a much more significant role. Often, inferencing can be a relatively lightweight workload, because it’s not highly parallel. A good example is the AI capability present in modern edge devices such as the latest iOS and Android smartphones.

As mentioned above, the average CPU also consumes less power than a GPU. That makes a CPU a better choice in situations where heat and battery life are important.

However, not all inferencing applications are lightweight, and such workloads may not be appropriate for CPUs. One example is autonomous vehicles. They will require massive parallel processing in real-time to ensure safety and optimum efficiency.

In these cases, GPUs will play a bigger role in the AI inferencing process, despite their higher cost and power requirements.

Powerful GPUs are already used for AI inferencing at the core. Examples include large-scale cloud services such as AWS, Google Cloud and Microsoft Azure.

Enterprise Grade

Enterprises often conduct AI training and inferencing on a scale so massive, it eclipses those found in edge environments. In these cases, IT engineers must rely on hugely powerful systems.

One example is the Supermicro AS -8125GS-TNMR2 server. This 8U behemoth—weighing in at 225 pounds—can operate up to eight AMD Instinct MI300X accelerators. And it’s equipped with dual AMD EPYC processors, the customer’s choice of either the 9004 or 9005 series.

To handle some of the world’s most demanding AI workloads, Supermicro’s server is packed with an astonishing amount of tech. In addition to its eight GPUs, the server also has room for a pair of AMD EPYC 9005-series processors, 6TB of ECC DDR5 memory, and 18 hot-swap 2.5-inch NVMe and SATA drives.

That makes the Supermicro system one of the most capable and powerful servers now available. And as AI evolves, tech leaders including AMD and Supermicro will undoubtedly produce more powerful CPUs, GPUs and servers to meet the growing demand.

What will the next generation of AI training and inferencing technology look like? To find out, you won’t have to wait long.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

AMD’s new ROCm 6.3 makes GPU programming even better

Featured content

AMD’s new ROCm 6.3 makes GPU programming even better

AMD recently introduced version 6.3 of ROCm, its open software stack for GPU programming. New features included expanded OS support and other optimizations.

Learn More about this topic
  • Applications:
  • Featured Technologies:

There’s a new version of AMD ROCm, the open software stack designed to enable GPU programming from low-level kernel all the way up to end-user applications.  

The latest version, ROCm 6.3, adds features that include expanded operating system support, an open-source toolkit and more.

Rock On

AMD ROCm provides the tools for HIP (the heterogeneous-computing interface for portability), OpenCL and OpenMP. These include compilers, APIs, libraries for high-level functions, debuggers, profilers and runtimes.

ROCm is optimized for Generative AI and HPC applications, and it’s easy to migrate existing code into. Developers can use ROCm to fine-tune workloads, while partners and OEMs can integrate seamlessly with AMD to create innovative solutions.

The latest release builds on ROCm 6, which AMD introduced last year. Version 6 added expanded support for AMD Instinct MI300A and MI300X accelerators, key AI support features, optimized performance, and an expanded support ecosystem.

The senior VP of AMD’s AI group, Vamsi Boppana, wrote in a recent blog post: “Our vision is for AMD ROCm to be the industry’s premier open AI stack, enabling choice and rapid innovation.”

New Features

Here’s some of what’s new in AMD ROCm 6.3:

  • rocJPEG: A high-performance JPEG decode SDK for AMD GPUs.
  • ROCm compute profiler and system profiler: Previously known as Omniperf and Omnitrace, these have been renamed to reflect their new direction as part of the ROCm software stack.
  • Shark AI toolkit: This open-source toolkit is for high-performance serving of GenAI and  LLMs. Initial release includes support for the AMD Instinct MI300.
  • PyTorch 2.4 support: PyTorch is a machine learning library used for applications such as computer vision and natural language processing. Originally developed by Meta AI, it’s now part of the Linux Foundation umbrella.
  • Expanded OS support: This includes added support for Ubuntu 24.04.2 and 22.04.5; RHEL 9.5; and Oracle Linux 8.10. In addition, ROCm 6.3.1 includes support for both Debian 12 and the AMD Instinct MI325X accelerator.
  • Documentation updates: ROCm 6.3 offers clearer, more comprehensive guidance for a wider variety of use cases and user needs.

Super for Supermicro

Developers can use ROCm 6.3 to create tune workloads and create solutions for Supermicro GPU systems based on AMD Instinct MI300 accelerators.

Supermicro offers three such systems:

Are your customers building AI and HPC systems? Then tell them about the new features offered by AMD ROCm 6.3.

Do More:

 

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

2024: A look back at the year’s best

Featured content

2024: A look back at the year’s best

Let's look back at 2024, a year when AI was everywhere, AMD introduced its 5th Gen EPYC processors, and Supermicro led with liquid cooling.

Learn More about this topic
  • Applications:
  • Featured Technologies:

You couldn't call 2024 boring.

If anything, the year was almost too exciting, too packed with important events, and moving much too fast.

Looking back, a handful of 2024’s technology events stand out. Here are a few of our favorite things.

AI Everywhere

In March AMD’s chief technology officer, Mark Papermaster, made some startling predictions that turned out to be absolutely true.

Speaking at an investors’ event sponsored by Arete Research, Papermaster said, “We’re thrilled to bring AI across our entire product portfolio.” AMD has indeed done that, offering AI capabilities from PCs to servers to high-performance GPU accelerators.

Papermaster also said the buildout of AI is an event as big as the launch of the internet. That certainly sounds right.

He also said AMD believes the total addressable market for AI through 2027 to be $400 billion. If anything, that was too conservative. More recently, consultants Bain & Co. predicted that figure will reach $780 billion to $990 billion.

Back in March, Papermaster said AMD had increased its projection for full-year AI sales from $2 billion to $3.5 billion. That’s probably too low, too.

AMD recently reported revenue of $3.5 billion for its data-center group for just the third quarter alone. The company attributed at least some of the group’s 122% year-on-year increase to the strong ramp of AMD Instinct GPU shipments.

5th Gen AMD EPYC Processors

October saw AMD introduce the fifth generation of its powerful line of EPYC server processors.

The 5th Gen AMD EPYC processors use the company’s new ‘Zen 5’ core architecture. It includes over 25 SKUs offering anywhere from 8 to 192 cores. And the line includes a model—the AMD EPYC 9575F—designed specifically to work with GPU-powered AI solutions.

The market has taken notice. During the October event, AMD CEO Lisa Su told the audience that nearly one in three servers worldwide (34%) are now powered by AMD EPYC processors. And Supermicro launched its new H14 line of servers that will use the new EPYC processors.

Supermicro Liquid Cooling

As servers gain power to add AI and other compute-intensive capabilities, they also run hotter. For data-center operators, that presents multiple challenges. One big one is cost: air conditioning is expensive. What’s more, AC may be unable to cool the new generation of servers.

Supermicro has a solution: liquid cooling. For some time, the company has offered liquid cooling as a data-center option.

In November the company took a new step in this direction. It announced a server that comes with liquid cooling only.

The server in question is the Supermicro 2U 4-node FlexTwin, model number AS -2126FT-HE-LCC. It’s a high-performance, hot-swappable, high-density compute system designed for HPC workloads.

Each 2U system comprises 4 nodes, and each node is powered by dual AMD EPYC 9005 processors. (The previous-gen AMD EPYC 9004s are supported, too.)

To keep cool, the FlexTwin server uses a direct-to-chip (D2C) cold plate liquid cooling setup. Each system also runs 16 counter-rotating fans. Supermicro says this cooling arrangement can remove up to 90% of server-generated heat.

AMD Instinct MI325X Accelerator

A big piece of AMD’s product portfolio for AI is its Instinct line of accelerators. This year the company promised to maintain a yearly cadence of new Instinct models.

Sure enough, in October the company introduced the AMD Instinct MI325X Accelerator. It’s designed for Generative AI performance and working with large language models (LLMs). The system offers 256GB of HBM3E memory and up to 6TB/sec. of memory bandwidth.

Looking ahead, AMD expects to formally introduce the line’s next member, the AMD Instinct MI350, in the second half of next year. AMD has said the new accelerator will be powered by a new AMD CDNA 4 architecture, and will improve AI inferencing performance by up to 35x compared with the older Instinct MI300.

Supermicro Edge Server

A lot of computing now happens at the edge, far beyond either the office or corporate data center.

Even more edge computing is on tap. Market watcher IDC predicts double-digit growth in edge-computing spending through 2028, when it believes worldwide sales will hit $378 billion.

Supermicro is on it. At the 2024 MWC, held in February in Barcelona, the company introduced an edge server designed for the kind of edge data centers run by telcos.

Known officially as the Supermicro A+ Server AS -1115SV-WTNRT, it’s a 1U short-depth server powered by a single AMD EPYC 8004 processor with up to 64 cores. That’s edgy.

Happy Holidays from all of us at Performance Intensive Computing. We look forward to serving you in 2025.

Check out these related blog posts:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

Tech Explainer: Why does PCIe 5.0 matter? And what’s coming next?

Featured content

Tech Explainer: Why does PCIe 5.0 matter? And what’s coming next?

PCIe 5.0 connects high-speed components to servers and PCs. Versions 6 & 7, coming soon, will deliver even higher speeds for tomorrow’s AI workloads.

Learn More about this topic
  • Applications:
  • Featured Technologies:

You’ve no doubt heard of PCIe 5.0. But what is it exactly? And why does it matter?

As the name and number imply, PCIe 5.0 is the fifth generation of the Peripheral Component Interconnect Express interface standard. PCIe essentially sets the rules for connecting high-speed components such as GPUs, networking cards and storage devices to servers, desktop PCs and other devices.

To be sure, these components could be connected via a number of other interface standards, such as USB-C and SATA.

But PCIe 5.0 alone offers extremely high bandwidth and low latency. That makes it a better choice for mission-critical enterprise IT operations and resource-intensive AI workloads.

Left in the Dust

The 5th generation of PCIe was released in May 2019, bringing significant improvements over PCIe 4.0. These include:

  • Increased Bandwidth. PCIe 5.0 has a maximum throughput of 32 giga-transfers per second (GT/s)—effectively double the bandwidth of its predecessor. In terms of data transfer, 32 GT/s translates to around 4 GB of data throughput per lane in each direction. That allows for a total of 64 GB/s in a 16-lane PCIe-based GPU. That’s perfect for modern GPU-dependent workflows such as AI-inferencing.
  • Lower Latency. Keeping latency as low as possible is crucial for applications like gaming, high-performance computing (HPC) and AI workloads. High latency can inhibit data retrieval and processing, which in turn hurts both application performance and the user experience. The latency of PCIe 5.0 varies depending on multiple factors, including network connectivity, attached devices and workloads. But it’s safe to assume an average latency of around 100 nanoseconds (ns) — roughly 50% less than PCIe 4.0. And again, with latency, lower is better.
  • Enhanced Data-Center Features. Modern data-center operations are among the most demanding. That’s especially true for IT operations focused on GenAI, machine learning and telecom. So it’s no surprise that PCIe 5.0 includes several features focused on enhanced operations for data centers. Among the most notable is increased bandwidth and faster data access for NVMe storage devices. PCIe 5.0 also includes features that enhance power management and efficiency.

Leveraging PCIe 5

AMD is a front-runner in the race to help enterprises cope with modern AI workloads. And the company has been quick to take advantage of PCIe 5.0’s performance improvements. Take, for example, the AMD Instinct MI325X Accelerator.

This system is a leading-edge accelerator module for generative AI, inference, training and HPC. Each discrete AMD Instinct MI325X offers a 16-lane PCIe Gen 5 host interface and seven AMD Infinity Fabric links for full connectivity between eight GPUs in a ring.

By leveraging a PCIe 5.0 connection, AMD’s accelerator can offer I/O-to-host-CPU and scale-out network bandwidths of 128 GB/sec.

AMD is also using PCIe on its server processors. The new 5th generation AMD EPYC server processors take advantage of PCIe 5.0’s impressive facility. Specifically, the AMD EPYC 9005 Series processors support 128 PCIe 5 I/O lanes in a single-socket server. For dual-socket servers, support increases to 160 lanes.

Supermicro is another powerful force in enterprise IT operations. The company’s behemoth H14 8-GPU system (model number AS-8126GS-TNMR2) leverages AMD EPYC processors and AMD Instinct accelerators to help enterprises deploy the largest AI and large language models (LLMs).

The H14’s standard configuration includes eight PCIe 5.0 x16 low-profile slots and two full-height slots. Users can also opt for a PCIe expansion kit, which adds two additional PCIe 5.0 slots. That brings the grand total to an impressive 12 PCIe 5.0 16-lane expansion slots.

PCIe 6.0 and Beyond

PCIe 5.0 is now entering its sixth year of service. That’s not a long time in the grand scheme of things. But the current version might feel ancient to IT staff who need to eke out every shred of bandwidth to support modern AI workloads.

Fortunately, a new PCIe generation is in the works. The PCIe 6.0 specification, currently undergoing testing and development, will offer still more performance gains over its predecessor.

PCI-SIG, an organization committed to developing and enhancing the PCI standard, says the 6.0 platform’s upgrades will include:

  • A data rate of up to 64 GT/sec., double the current rate and providing a maximum bidirectional bandwidth of up to 256 GB/sec for x16 lanes
  • Pulse Amplitude Modulation with 4 levels (PAM4)
  • Lightweight Forward Error Correct (FEC) and Cyclic Redundancy Check (CRC) to mitigate the bit error rate increase associated with PAM4 signaling
  • Backwards compatibility with all previous generations of PCIe technology

There’s even a next generation after that, PCIe 7.0. This version could be released as soon as 2027, according to the PCI-SIG. That kind of speed makes sense considering the feverish rate at which new technology is being developed to enable and expand AI operations.

It’s not yet clear how accurate those release dates are. But one thing’s for sure: You won’t have to wait long to find out.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

Tech Explainer: What is the AMD “Zen” core architecture?

Featured content

Tech Explainer: What is the AMD “Zen” core architecture?

Originally launched in 2017, this CPU architecture now delivers high performance and efficiency with ever-thinner processes.

Learn More about this topic
  • Applications:
  • Featured Technologies:

The recent release of AMD’s 5th generation processors—formerly codenamed Turin—also heralded the introduction of the company’s “Zen 5” core architecture.

“Zen” is AMD’s name for a design ethos that prioritizes performance, scalability and efficiency. As any CTO will tell you, these 3 aspects are crucial for success in today’s AI era.

AMD originally introduced its “Zen” architecture in 2017 as part of a broader campaign to steal market share and establish dominance in the all-important enterprise IT space.

Subsequent generations of the “Zen” design have markedly increased performance and efficiency while delivering ever-thinner manufacturing processes.

Now and Zen

Since the “Zen” core’s original appearance in AMD Ryzen 1000-series processors, the architecture’s design philosophy has maintained its focus on a handful of vital aspects. They include:

  • A modular design. Known as Infinity Fabric, it facilitates efficient connectivity among multiple CPU cores and other components. This modular architecture enhances scalability and performance, both of which are vital for modern enterprise IT infrastructure.
  • High core counts and multithreading. Both are common to EPYC and Ryzen CPUs built using the AMD “Zen” core architecture. Simultaneous multithreading enables each core to process 2 threads. In the case of EPYC processors, this makes AMD’s CPUs ideal for multithreaded workloads that include Generative AI, machine learning, HPC and Big Data.
  • Advanced manufacturing processes. These allow faster, more efficient communication among individual CPU components, including multithreaded cores and multilevel caches. Back in 2017, the original “Zen” architecture was manufactured using a 14-nanometer (nm) process. Today’s new “Zen 5” and “Zen 5c” architectures (more on these below) reduce the lithography to just 4nm and 3nm, respectively.
  • Enhanced efficiency. This enables IT staff to better manage complex enterprise IT infrastructure. Reducing heat and power consumption is crucial, too, both in data centers and at the edge. The AMD “Zen” architecture makes this possible by offering enterprise-grade EPYC processors that offer up to 192 cores, yet require a maximum thermal design power (TDP) of only 500W.

The Two-Fold Path

The latest, fifth generation “Zen” architecture is divided into two segments: “Zen 5” and “Zen 5c.”

“Zen 5” employs a 4-nanometer (nm) manufacturing process to deliver up to 128 cores operating at up to 4.1GHz. It’s optimized for high per-core performance.

“Zen 5c,” by contrast, offers a 3nm lithography that’s reserved for AMD EPYC 96xx, 97xx, 98xx, and 99xx series processors. It’s optimized for high density and power efficiency.

The most powerful of these CPUs—the AMD EPYC 9965—includes an astonishing 192 cores, a maximum boost clock speed of 3.7GHz, and an L3 cache of 384MB.

Both “Zen 5” and “Zen 5c” are key components of the 5th gen AMD EPYC processors introduced earlier this month. Both have also been designed to achieve double-digit increases in instructions per clock cycle (IPC) and equip the core with the kinds of data handling and processing power required by new AI workloads.

Supermicro’s Satori

AMD isn’t the only brand offering bold, new tech to harried enterprise IT managers.

Supermicro recently introduced its new H14 servers, GPU-accelerated systems and storage servers powered by AMD EPYC 9005 Series processors and AMD Instinct MI325X Accelerators. A number of these servers also support the new AMD “Turin” CPUs.

The new product line features updated versions of Supermicro’s vaunted Hyper system, Twin multinode servers, and AI-inferencing GPU systems. All are now available with the user’s choice of either air or liquid cooling.

Supermicro says its collection of purpose-built powerhouses represents one of the industry’s most extensive server families. That should be welcome news for organizations intent on building a fleet of machines to meet the highly resource-intensive demands of modern AI workloads.

By designing its next-generation infrastructure around AMD 5th Generation components, Supermicro says it can dramatically increase efficiency by reducing customers’ total data-center footprints by at least two-thirds.

Enlightened IT for the AI Era

While AMD and Supermicro’s advances represent today’s cutting-edge technology, tomorrow is another story entirely.

Keeping up with customer demand and the dizzying pace of AI-based innovation means these tech giants will soon return with more announcements, tools and design methodologies. AMD has already promised a new accelerator, the AMD Instinct MI350, will be formally announced in the second half of 2025.

As far as enterprise CTOs are concerned, the sooner, the better. To survive and thrive amid heavy competition, they’ll need an evolving array of next-generation technology. That will help them reduce their bottom lines even as they increase their product offerings—a kind of technological nirvana.

Do More:

Watch a related video: 

Featured videos

Events

Find AMD & Supermicro Elsewhere

Do your customers need more room for AI? AMD has an answer

Featured content

Do your customers need more room for AI? AMD has an answer

If your customers are looking to add AI to already-crowded, power-strapped data centers, AMD is here to help. 

Learn More about this topic
  • Applications:
  • Featured Technologies:

How can your customers make room for AI in data centers that are already full?

It’s a question that’s far from academic. Nine in 10 tech vendors surveyed recently by the Uptime Institute expect AI to be widely used in data centers in the next 5 years.

Yet data center space is both hard to find and costly to rent. Vacancy rates have hit new lows, according to real-estate services firm CBRE Group.

Worse, this combination of supply shortages and high demand is driving up data center pricing and rents. Across North America, CRBE says, pricing is up by 20% year-on-year.

Getting enough electric power is an issue, too. Some utilities have told prospective data-center customers they won’t get the power they requested until the next decade, reports The Wall Street Journal. In other cases, strapped utilities are simply giving customers less power than they asked for.

So how to help your customers get their data centers ready for AI? AMD has some answers. And a free software tool to help.

The AMD Solution

AMD’s solution is simple, with just 2 points:

  • Make the most of existing data-center real estate and power by consolidating existing workloads.
  • Replace the low-density compute of older, inefficient and out-of-warranty systems with compute that’s newer, denser and more efficient.

AMD is making the case that your customers can do both by moving from older Intel-based systems to newer ones that are AMD-based.

For example, the company says, replacing servers based on Intel Xeon 6143 Sky Lake processors with those based on AMD EPYC 9334 CPUs can result in the need for 73% fewer servers, 70% fewer racks and 69% less power.

That could include Supermicro servers powered by AMD EPYC processors. Supermicro H13 servers using AMD EPYC 9004 Series processors offer capabilities for high-performance data centers.

AMD hasn’t yet done comparisons with either its new 5th gen EPYC processors (introduced last week) or Intel’s 86xx CPUs. But the company says the results should be similar.

Consolidating processor-based servers can also make room in your customers’ racks for AMD Instinct MI300 Series accelerators designed specifically for AI and HPC workloads.

For example, if your customer has older servers based on Intel Xeon Cascade Lake processors, migrating them to servers based on AMD EPYC 9754 processors instead can gain them as much as a 5-to-1 consolidation.

The result? Enough power and room to accommodate a new AI platform.

Questions Answered

Simple doesn’t always mean easy. And you and your customers may have concerns.

For example, isn’t switching from one vendor to another difficult?

No, says AMD. The company cross-licenses the X86 instruction set, so on its processors, most workloads and applications will just work.

What about all those cores on AMD processors? Won’t they raise a customer’s failure domain too high?

No, says AMD. Its CPUs are scalable enough to handle any failure domain from 8 to 256 cores per server.

Wouldn’t moving require a cold migration? And if so, wouldn’t that disrupt the customer’s business?

Again, AMD says no. While moving virtual machines (VMs) to a new architecture does require a cold migration, the job can be done without any application downtime.

That’s especially true if you use AMD’s free open-source tool known as VAMT, short for VMware Architecture Migration Tool. VAMT automates cold migration. In one AMD test, it migrated hundreds of VMs in just an hour.

So if your customers among those struggling to find room for AI systems in their already-crowded and power-strapped data centers, tell them consider a move to AMD.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

AMD intros CPUs, accelerators, networking for end-to-end AI infrastructure -- and Supermicro supports

Featured content

AMD intros CPUs, accelerators, networking for end-to-end AI infrastructure -- and Supermicro supports

AMD expanded its end-to-end AI infrastructure products for data centers with new CPUs, accelerators and network controllers. And Supermicro is already offering supporting servers. 

Learn More about this topic
  • Applications:
  • Featured Technologies:

AMD today held a roughly two-hour conference in San Francisco during which CEO Lisa Su and other executives introduced a new generation of server processors, the next model in the Instinct MI300 Accelerator family, and new data-center networking devices.

As CEO Su told the audience the live and online audience, AMD is committed to offering end-to-end AI infrastructure products and solutions in an open, partner-dependent ecosystem.

Su further explained that AMD’s new AI strategy has 4 main goals:

  • Become the leader in end-to-end AI
  • Create an open AI software platform of libraries and models
  • Co-innovate with partners including cloud providers, OEMs and software creators
  • Offer all the pieces needed for a total AI solution, all the way from chips to racks to clusters and even entire data centers.

And here’s a look at the new data-center hardware AMD announced today.

5th Gen AMD EPYC CPUs

The EPYC line, originally launched in 2017, has become a big success for AMD. As Su told the event audience, there are now more than 950 EPYC instances at the largest cloud providers; also, AMD hardware partners now offer EPYC processors on more than 350 platforms. Market share is up, too: Nearly one in three servers worldwide (34%) now run on EPYC, Su said.

The new EPYC processors, formerly codenamed Turin and now known as the AMD EPYC 9005 Series, are now available for data center, AI and cloud customers.

The new CPUs also have a new core architecture known as Zen5. AMD says Zen5 outperforms the previous Zen4 generation by 17% on enterprise instructions-per-clock and up to 37% on AI and HPC workloads.

The new 5th Gen line has over 25 SKUs, and core count ranges widely, from as few as 8 to as many as 192. For example, the new AMD EPYC 9575F is a 65-core, 5GHz CPU designed specifically for GPU-powered AI solutions.

AMD Instinct MI325X Accelerator

About a year ago, AMD introduced the Instinct MI300 Accelerators, and since then the company committed itself to introducing new models on a yearly cadence. Sure enough, today Lisa Su introduced the newest model, the AMD Instinct MI325X Accelerator.

Designed for Generative AI performance and built on the AMD CDNA3 architecture, the new accelerator offers up to 256GB of HBM3E memory, and bandwidth up to 6TB/sec.

Shipments of the MI325X are set to begin in this year’s fourth quarter. Partner systems with the new AMD accelerator are expected to start shipping in next year’s first quarter.

Su also mentioned the next model in the line, the AMD Instinct MI350, which will offer up to 288GB of HBM3E memory. It’s set to be formally announced in the second half of next year.

Networking Devices

Forrest Norrod, AMD’s head of data-center solutions, introduced two networking devices designed for data centers running AI workloads.

The AMD Pensando Salina DPU is designed for front-end connectivity. It supports thruput of up to 400 Gbps.

The AMD Pensando Pollara 400, designed for back-end networks connecting multiple GPUs, is the industry’s first Ultra-Ethernet Consortium-ready AI NIC.

Both parts are sampling with customers now, and AMD expects to start general shipments in next year’s first half.

Both devices are needed, Norrod said, because AI dramatically raises networking demands. He cited studies showing that connectivity currently accounts for 40% to 75% of the time needed to run certain AI training and inference models.

Supermicro Support

Supermicro is among the AMD partners already ready with systems based on the new AMD processors and accelerator.

Wasting no time, Supermicro today announced new H14 series servers, including both Hyper and FlexTwin systems, that support the 5th gen AMD 9005 EPYC processors and AMD Instinct MI325X Accelerators.

The Supermicro H14 family includes three systems for AI training and inference workloads. Supermicro says the systems can also accommodate the higher thermal requirements of the new AMD EPYC processors, which are rated at up to 500W. Liquid cooling is an option, too.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

The AMD Instinct MI300X Accelerator draws top marks from leading AI benchmark

Featured content

The AMD Instinct MI300X Accelerator draws top marks from leading AI benchmark

In the latest MLPerf testing, the AMD Instinct MI300X Accelerator with ROCm software stack beat the competition with strong GenAI inference performance. 

Learn More about this topic
  • Applications:
  • Featured Technologies:

New benchmarks using the AMD Instinct MI300X Accelerator show impressive performance that surpasses the competition.

This is great news for customers operating demanding AI workloads, especially those underpinned by large language models (LLMs) that require super-low latency.

Initial platform tests using MLPerf Inference v4.1 measured AMD’s flagship accelerator against the Llama 2 70B benchmark. This test is an indication for real-world applications, including natural language processing (NLP) and large-scale inferencing.

MLPerf is the industry’s leading benchmarking suite for measuring the performance of machine learning and AI workloads from domains that include vision, speech and NLP. It offers a set of open-source AI benchmarks, including rigorous tests focused on Generative AI and LLMs.

Gaining high marks from the MLPerf Inference benchmarking suite represents a significant milestone for AMD. It positions the AMD Instinct MI300X accelerator as a go-to solution for enterprise-level AI workloads.

Superior Instincts

The results of the LLaMA2-70B test are particularly significant. That’s due to the benchmark’s ability to produce an apples-to-apples comparison of competitive solutions.

In this benchmark, the AMD Instinct MI300X was compared with NVIDIA’s H100 Tensor Core GPU. The test concluded that AMD’s full-stack inference platform was better than the H100 at achieving high-performance LLMs, a workload that requires both robust parallel computing and a well-optimized software stack.

The testing also showed that because the AMD Instinct MI300X offers the largest GPU memory available—192GB of HBM3 memory—it was able to fit the entire LLaMA2-70B model into memory. Doing so helped to avoid network overhead by preventing model splitting. This, in turn, maximized inference throughput, producing superior results.

Software also played a big part in the success of the AMD Instinct series. The AMD ROCm software platform accompanies the AMD Instinct MI300X. This open software stack includes programming models, tools, compilers, libraries and runtimes for AI solution development on the AMD Instinct MI300 accelerator series and other AMD GPUs.

The testing showed that the scaling efficiency from a single AMD Instinct MI300X, combined with the ROCm software stack, to a complement of eight AMD Instinct accelerators was nearly linear. In other words, the system’s performance improved proportionally by adding more GPUs.

That test demonstrated the AMD Instinct MI300X’s ability to handle the largest MLPerf inference models to date, containing over 70 billion parameters.

Thinking Inside the Box

Benchmarking the AMD Instinct MI300X required AMD to create a complete hardware platform capable of addressing strenuous AI workloads. For this task, AMD engineers chose as their testbed the Supermicro AS -8125GS-TNMR2, a massive 8U complete system.

Supermicro’s GPU A+ Client Systems are designed for both versatility and redundancy. Designers can outfit the system with an impressive array of hardware, starting with two AMD EPYC 9004-series processors and up to 6TB of ECC DDR5 main memory.

Because AI workloads consume massive amounts of storage, Supermicro has also outfitted this 8U server with 12 front hot-swap 2.5-inch NVMe drive bays. There’s also the option to add four more drives via an additional storage controller.

The Supermicro AS -8125GS-TNMR2 also includes room for two hot-swap 2.5-inch SATA bays and two M.2 drives, each with a capacity of up to 3.84TB.

Power for all those components is delivered courtesy of six 3,000-watt redundant titanium-level power supplies.

Coming Soon: Even More AI power

AMD engineers continually push the limits of silicon and human ingenuity to expand the capabilities of their hardware. So it should come as little surprise that new iterations of the AMD Instinct series are expected to be released in the coming months. This past May, AMD officials said they plan to introduce AMD Instinct MI325, MI350 and MI400 accelerators.

Forthcoming Instinct accelerators, AMD says, will deliver advances including additional memory, support for lower-precision data types, and increased compute power.

New features are also coming to the AMD ROCm software stack. Those changes should include software enhancements including kernel improvements and advanced quantization support.

Are you customers looking for a high-powered, low-latency system to run their most demanding HPC and AI workloads? Tell them about these benchmarks and the AMD Instinct MI300X accelerators.

Do More:

 

Featured videos

Events

Find AMD & Supermicro Elsewhere

Developing AI and HPC solutions? Check out the new AMD ROCm 6.2 release

Featured content

Developing AI and HPC solutions? Check out the new AMD ROCm 6.2 release

The latest release of AMD’s free and open software stack for developing AI and HPC solutions delivers 5 important enhancements. 

Learn More about this topic
  • Applications:
  • Featured Technologies:

If you develop AI and HPC solutions, you’ll want to know about the most recent release of AMD ROCm software, version 6.2.

ROCm, in case you’re unfamiliar with it, is AMD’s free and open software stack. It’s aimed at developers of artificial intelligence and high-performance computing (HPC) solutions on AMD Instinct accelerators. It's also great for developing AI and HPC solutions on AMD Instinct-powered servers from Supermicro. 

First introduced in 2016, ROCm open software now includes programming models, tools, compilers, libraries, runtimes and APIs for GPU programming.

ROCm version 6.2, announced recently by AMD, delivers 5 key enhancements:

  • Improved vLLM support 
  • Boosted memory efficiency & performance with Bitsandbytes
  • New Offline Installer Creator
  • New Omnitrace & Omniperf Profiler Tools (beta)
  • Broader FP8 support

Let’s look at each separately and in more detail.

LLM support

To enhance the efficiency and scalability of its Instinct accelerators, AMD is expanding vLLM support. vLLM is an easy-to-use library for the large language models (LLMs) that power Generative AI.

ROCm 6.2 lets AMD Instinct developers integrate vLLM into their AI pipelines. The benefits include improved performance and efficiency.

Bitsandbytes

Developers can now integrate Bitsandbytes with ROCm for AI model training and inference, reducing their memory and hardware requirements on AMD Instinct accelerators. 

Bitsandbytes is an open source Python library that enables LLMs while boosting memory efficiency and performance. AMD says this will let AI developers work with larger models on limited hardware, broadening access, saving costs and expanding opportunities for innovation.

Offline Installer Creator

The new ROCm Offline Installer Creator aims to simplify the installation process. This tool creates a single installer file that includes all necessary dependencies.

That makes deployment straightforward with a user-friendly GUI that allows easy selection of ROCm components and versions.

As the name implies, the Offline Installer Creator can be used on developer systems that lack internet access.

Omnitrace and Omniperf Profiler

The new Omnitrace and Omniperf Profiler Tools, both now in beta release, provide comprehensive performance analysis and a streamlined development workflow.

Omnitrace offers a holistic view of system performance across CPUs, GPUs, NICs and network fabrics. This helps developers ID and address bottlenecks.

Omniperf delivers detailed GPU kernel analysis for fine-tuning.

Together, these tools help to ensure efficient use of developer resources, leading to faster AI training, AI inference and HPC simulations.

FP8 Support

Broader FP8 support can improve the performance of AI inferencing.

FP8 is an 8-bit floating point format that provides a common, interchangeable format for both AI training and inference. It lets AI models operate and perform consistently across hardware platforms.

In ROCm, FP8 support improves the process of running AI models, particularly in inferencing. It does this by addressing key challenges such as the memory bottlenecks and high latency associated with higher-precision formats. In addition, FP8's reduced precision calculations can decrease the latency involved in data transfers and computations, losing little to no accuracy.  

ROCm 6.2 expands FP8 support across its ecosystem, from frameworks to libraries and more, enhancing performance and efficiency.

Do More:

Watch the related video podcast:

Featured videos

Events

Find AMD & Supermicro Elsewhere

Pages