Connexion
When Russia began attacking Ukrainian data centers and telecommunication infrastructure, it apparently failed to prepare for possible retaliation. Now Russia’s digital services are facing widespread outages after Ukrainian drone strikes knocked out three of the largest data centers owned by Russian tech giant Yandex within several days.
Pro-Kremlin bloggers are warning that Russia’s centralized digital economy could collapse because the country is not prepared for the “data center war” it started, according to posts translated by the pseudonymous independent researcher ChrisO_wiki. The Russian commentators also suggested that Ukraine’s decentralized digital services could prove more resilient than Russia’s system if the countries continue trading such blows.
Russia initially launched a drone strike campaign targeting Ukrainian data centers and mobile phone networks in late September. The attacks temporarily disrupted Internet and cell phone service for up to 100,000 households in the Ukrainian capital of Kyiv and came as part of a broader long-range strike campaign that has involved Russian missiles and jet-powered drones hitting Ukraine’s power grid infrastructure, transportation hubs, and even residential apartment building blocks.
The Museo de Historia de la Computación in Spain has unveiled a 1:1 scale visual replica of the famous Cray-1 supercomputer from 1976. Instead of sourcing expensive, cutting-edge tech to run the system, the creative team powered the project with 30 2012-era Mac Minis. Developed over 18 months, the project serves as a 50th-anniversary tribute to the Cray-1 and Apple's founding.
Mac Minis from over a decade ago might not seem like the obvious first choice for building a cluster supercomputer, but when networked together, the developers can orchestrate the 64 active processing cores into a combined 1.3 teraflops of performance in their most recent tests. The museum reportedly achieved 1.5 teraflops previously using all 70 available cores, prior to one node going offline. Benchmarks were obtained using a custom matrix-multiplication Python script. Five PCs are quad-core i7s, and the remaining 25 are dual-core i5 models. Except for two units from 2014, the rest of the array dates to 2012.
The replica is a valuable tool for museum-goers to understand the shifts in computer architecture since 1976. The original Cray-1 used vector processing, applying a single instruction to multiple data elements simultaneously, whereas the replica distributes work across independent processors in a parallel cluster. The peak theoretical performance of the original Cray-1 was approximately 160 megaflops. What was originally a purpose-built research machine has been re-created with consumer hardware that's vastly more capable than the machine that inspired it.
Nvidia commits $1 billion over five years to support US science, quantum computing and AI research, expanding its role in the federal Genesis Mission.
The post Nvidia’s $1 Billion Research Plan Targets AI, Quantum Computing and Medicine appeared first on TechRepublic.
China’s LineShine debuted as the world’s fastest supercomputer on TOP500, but mixed-precision benchmark results complicate its AI performance claim.
The post China’s LineShine Tops Supercomputer List Despite AI Benchmark Gap appeared first on TechRepublic.
CoreWeave, an NVIDIA-backed cloud service provider specializing in GPU-accelerated services, has secured a debt facility worth $2.3 billion using NVIDIA's H100-based hardware as collateral. The company intends to use the funds to procure more compute GPUs and systems from NVIDIA, construct new data centers, and hire additional personnel to meet the growing needs for AI and HPC workloads.
CoreWeave has reaped enormous benefits from the rise in generative AI due to its large-scale cloud infrastructure as well as an exclusive relationship with NVIDIA, and its ability to procure the company's H100 compute GPUs as well as HGX H100 supercomputing platforms amid shortages of AI and HPC hardware. Since many AI and HPC applications used nowadays were developed for NVIDIA's CUDA platform and API, they require NVIDIA's GPUs. Therefore, access to H100 gives CoreWeave a competitive edge over traditional CSPs like AWS, Google, and Microsoft.
In addition to offering its customers access to advanced hardware, CoreWeave collaborates with AI startups and major CSPs — which are essentially its competitors — to build clusters that power AI workloads. These rivals — AWS and Google — have their own processors for AI workloads, and they continue to develop new ones. Still, given the dominance of CUDA, they have to offer NVIDIA-powered instances to their clients and are currently grappling with NVIDIA GPU supply limitations.
CoreWeave's competitive advantage, facilitated by access to NVIDIA's latest hardware, is a key factor in the company's ability to secure such substantial credit lines from companies like Magnetar Capital, Blackstone, Coatue, DigitalBridge, BlackRock, PIMCO, and Carlyle. Meanwhile, CoreWeave has already gotten $421 million from Magnetar at a valuation exceeding $2 billion.
Notably, this is not the first example of an NVIDIA-supported startup reaping substantial benefits from its association with the tech giant. Last month, Inflection AI built a supercomputer worth hundreds of millions of dollars powered by 22,000 NVIDIA H100 compute GPUs.
Meanwhile, this is the first time NVIDIA's H100-based hardware was used as collateral, emphasizing these processors' importance in the capital-intensive AI and HPC cloud business. Moreover, this massive loan indicates the growing market for private asset-based financing secured by actual physical assets.
"We negotiated with them to find a schedule for how much collateral to go into it, what the depreciation schedule was going to be versus the payoff schedule," said Michael Intrator, CoreWeave's CEO. "For us to go out and to borrow money against the asset base is a very cost-effective way to access the debt markets."
The company recently announced a $1.6 billion data center in Texas and plans to expand its presence to 14 locations within the U.S. by the end of 2023.
The Texas Advanced Computing Center (TACC) unveiled its latest Stampede supercomputer for open science research projects, Stampede3. TACC anticipates that Stampede3 will come online this fall and will deliver its full performance in early 2024. The supercomputer will be a crucial component of the U.S. National Science Foundation’s (NSF) ACCESS scientific supercomputing ecosystem, and it is projected to serve the open science community from 2024 until 2029.
The third-generation Stampede cluster, which will be built by Dell, will incorporate 560 nodes equipped with Intel's Sapphire Rapids generation Xeon CPU Max processors, each offering 56 CPU cores and 64GB of on-package HBM2E memory. Surprisingly, TACC is going to be operating these nodes in HBM-only mode, so no additional DRAM will be attached to the CPU nodes – all of their memory will come from the on-chip HBM stacks.
With these specifications, Stampede3 is expected to have a peak performance of approximately 4 FP64 PetaFLOPS, while offering nearly 63,000 general-purpose cores. In addition, TACC also plans to install 10 Dell PowerEdge XE9640 servers with 40 Intel Data Center GPU Max compute GPUs for artificial intelligence and machine learning workloads.
Given this layout, the bulk of Stampede3's compute performance will be supplied by CPUs. This makes Stampede3 a bit of a rarity in this day and age, as most high-performance systems are GPU driven, leaving Stampede3 as one of the last supercomputers that relies almost solely on general-purpose CPUs.
And while the current cluster is primarily focused on CPU performance, TACC is also going to use the Intel GPUs in the latest Stampede revamp to investigate on how to incorporate larger numbers of GPUs into future versions of the system. For now, most of TACC's AI tasks are run on its Lone Star systems, which is powered by hundreds Nvidia A100 compute GPUs. So the organization's aim is to explore whether a portion of this workload can be transferred to Intel's Ponte Vecchio.
We are going to put in a small system with exploratory capability using Intel Ponte Vecchio," said Dan Stanzione, executive director of TACC. "We are still negotiating exactly how much of that will have, but I would say a minimum of 40 nodes and maximum of a hundred or so. […] We are just putting a couple of racks of Ponte Vecchio out there to see how people work with it."
Stampede3 will leverage 400 Gb/s Omni-Path Fabric technology that will enable a backplane bandwidth of 24TB/s. This setup will allow the machine to efficiently scale and minimize latencies, making it well-suited for various applications requiring simulations.
TACC also plans to reincorporate nodes from the previous version, Stampede2, which were based on older-generation Xeon Scalable CPUs. This integration will enhance the capacity of Stampede3 for high-memory applications, high-throughput computing, interactive workloads, and other previous-generation applications. In total, the new supercomputer system will feature 1,858 compute nodes with over 140,000 cores, more than 330 TBs of RAM, new storage capacity of 13 PBs, and a peak performance close to 10 PetaFLOPS.
Cerebras Systems and G42, a tech holding group, have unveiled their Condor Galaxy project, a network of nine interlinked supercomputers for AI model training with aggregated performance of 36 FP16 ExaFLOPs. The first supercomputer, named Condor Galaxy 1 (CG-1), boasts 4 ExaFLOPs of FP16 performance and 54 million cores. CG-2 and CG-3 will be located in the U.S. and will follow in 2024. The remaining systems will be located across the globe and the total cost of the project will be over $900 million.
The CG-1 supercomputer, situated in Santa Clara, California, combines 64 Cerebras CS-2 systems into a single user-friendly AI supercomputer, capable of providing 4 ExaFLOPs of dense, systolic FP16 compute for AI training. Based around Cerebras's 2.6 trillion transistor second-generation wafer scale engine processors, the machine is designed specifically for Large Language Models and Generative AI. It supports up to 600 billion parameter models, with configurations that can be expanded to support up to 100 trillion parameter models. Its 54 million AI-optimized compute cores and massivefabric network bandwidth of 388 Tb/s allow for nearly linear performance scaling from 1 to 64 CS-2 systems, according to Cerebras.
The CG-1 supercomputer also offers inherent support for long sequence length training (up to 50,000 tokens) and does not require any complex distributed programming languages, which is common in case of GPU clusters.
“Delivering 4 exaFLOPs of AI compute at FP16, CG-1 dramatically reduces AI training timelines while eliminating the pain of distributed compute,” said Andrew Feldman, CEO of Cerebras Systems. “Many cloud companies have announced massive GPU clusters that cost billions of dollars to build, but that are extremely difficult to use. Distributing a single model over thousands of tiny GPUs takes months of time from dozens of people with rare expertise. CG-1 eliminates this challenge. Setting up a generative AI model takes minutes, not months and can be done by a single person. CG-1 is the first of three 4 ExaFLOP AI supercomputers to be deployed across the U.S. Over the next year, together with G42, we plan to expand this deployment and stand up a staggering 36 exaFLOPs of efficient, purpose-built AI compute.”
This supercomputer is provided as a cloud service by Cerebras and G42 and since it is located in the U.S., Cerebras and G42 assert that it will not be used by hostile states.
CG-1 is the first of three 4 FP16 ExaFLOP AI supercomputers (CG-1, CG-2, and CG-3) created by Cerebras and G42 in collaboration and located in the U.S. Once connected, these three AI supercomputers will form a 12 FP16 ExaFLOP, 162 million core distributed AI supercomputer, though it remains to be seen how efficient this network will be.
In 2024, G42 and Cerebras plan to launch six additional Condor Galaxy supercomputers across the world, which will increase the total compute power to 36 FP16 ExaFLOPs delivered by 576 CS-2 systems.
The Condor Galaxy project aims to democratize AI by offering sophisticated AI compute technology in the cloud.
Lawrence Livermore National Laboratory (LLNL) started to install its El Capitan supercomputer that promises to achieve computational performance of over 2 FP64 ExaFLOPS for classified national security research. Parallel to this, the LLNL also plans to introduce a less potent but still extremely fast supercomputer named Tuolumne. This companion model will be dedicated to unclassified research and will offer 10% to 15% of El Capitan's computational prowess, which is still enough to beat most of the systems in the current Top500 supercomputer list.
While LLNL's El Capitan will not be the first exascale system and will not even be the first to break 2 FP64 ExaFLOPS performance record, it is still a very special system as it will be the first exascale supercomputer to use AMD's Instinct MI300A accelerated processing units comprising of Zen 4 x86 cores and CDNA 3-based compute GPUs. When El Capitan goes online, it will substantially outpace Frontier, the world's fastest PC, which is rated for an Rpeak performance of 1.679 FP64 ExaFLOPS.
Tuolumne is expected to use similar architecture to El Capitan, so it will also use Instinct MI300A APU. Even with just a tenth of El Capitan's performance, Tuolumne will join the upper echelons of the global supercomputers list, capable of delivering 200 PetaFLOPS. This potential performance could place it among the Top 10 supercomputers listed in the current Top 500 list. Still, if it manages to deliver 15% of El Capitan's performance, it could rival Leonardo, a supercomputer powered by Intel Xeon Platinum 8358 and Nvidia A100 with an Rpeak performance of 304.47 PetaFLOPS, which is currently ranked as the fourth fastest supercomputer.
"We are planning to get an unclassified system that will be called Tuolumne," said Bronis R. de Supinski in an interview with ExaScaleProject.org. De Supinski is the chief technology officer for Livermore Computing at LLNL. "It will be roughly between 10% to 15% the size of El Capitan."
Supercomputers with classified status, like the existing Sierra and the planned El Capitan at LLNL, primarily serve national security needs. According to Bronis R. de Supinski, El Capitan will be mainly used for the stockpile stewardship program, ensuring the reliability of nuclear weapons without resorting to real testing. Conversely, unclassified supercomputers, like the upcoming Tuolumne, serve diverse computational requirements spanning various scientific fields, from research and engineering simulations to data analysis and weather prediction.
"Tuolumne will be contributing more to the wider range of scientific areas," de Supinski added. "There's a lot of materials modeling. We have typically had a wide range of molecular dynamics. Some QCD get run on the system, seismic modeling. What will probably happen is that you know, those sorts of applications, climate, and that sort of thing will run on Tuolumne. And if there is a particular case to be made, we can occasionally provide for briefer runs on the big system."
Source: ExaScaleProject.org
Lawrence Livermore National Laboratory had received the first components of its upcoming El Capitan supercomputer and begun to install them, the laboratory announced on Wednesday. The system is set to come online in mid-2024 and is expected to deliver performance of over 2 ExaFLOPS.
LLML's El Capitan is based on Cray's Shasta supercomputer architecture and will be built by HPE, just like two other exascale systems in the U.S., Frontier and Aurora. Unlike the first two exascale machines, which use a traditional discrete CPU plus discrete GPU configuration, the El Capitan supercomputer will be the first one based on AMD server-grade APUs that integrate both processor types in to a single, highly connected package.
AMD's Instinct MI300A APU incorporates both CPU and GPU chiplets, offering 24 general-purpose Zen 4 cores, compute GPUs powered by the CDNA 3 architecture, and 128 GB of unified on-package HBM3 memory. AMD has been internally evaluating its Instinct MI300A APU for months, and it appears that AMD and HPE are now ready to start installing the first pieces of hardware that make up El Capitan.
According to pictures released by the Lawrence Livermore National Laboratory, its engineers have already put a substantial number of servers into racks. Though LLNL's announcement leaves it unclear whether these are "completed" servers with production-quality silicon, or pre-production servers that will be filled out with production silicon at a later date. Notably, parts of Aurora were initially assembled with pre-production CPUs, which were only swapped out for Xeon CPU Max chips over the past couple of months. Given the amount of validation work required to stand-up a world-class supercomputer, AMD and HPE may be employing a similar strategy here.
"We have begun receiving & installing components for El Capitan, first #exascale #supercomputer," a Tweet by LLNL reads. "While we are still a ways from deploying it for national security purposes in 2024, it is exciting to see years of work becoming reality."
When it comes online in 2024, LLNL is expecting El Capitan to be the fastest supercomputer in the world. Though with its full specifications still being held back, it's not clear how much faster it is on paper compared to the 2 EFLOPS Aurora – let alone real-world performance. Part of the design goal of AMD's MI300A APU is to exploit additional performance efficiency gains that come from placing CPU and GPU blocks so close together, so it will be interesting to see what the software development teams programming for El Capitan can achieve, especially as they get their software further optimized.
LLNL's El Capitan is expected to cost $600 million. The system will be used nuclear weapons simulations and will be crucial for the U.S. national security. It replaces Sierra, a supercomputer based on IBM Power 9 and NVIDIA Volta accelerators, and promises to offer performance that is 16 times higher.
Argonne National Laboratory and Intel said on Thursday that they had installed all 10,624 blades for the Aurora supercomputer, a machine announced back in 2015 with a particularly bumpy history. The system promises to deliver a peak theoretical compute performance over 2 FP64 ExaFLOPS using its array of tens of thousands of Xeon Max 'Sapphire Rapids' CPUs with on-package HBM2E memory as well as Data Center GPU Max 'Ponte Vecchio' compute GPUs. The system will come online later this year.
"Aurora is the first deployment of Intel's Max Series GPU, the biggest Xeon Max CPU-based system, and the largest GPU cluster in the world," said Jeff McVeigh, Intel corporate vice president and general manager of the Super Compute Group.
The Aurora supercomputer looks quite impressive, even by the numbers. The machine is powered by 21,248 general-purpose processors with over 1.1 million cores for workloads that require traditional CPU horsepower and 63,744 compute GPUs that will serve AI and HPC workloads. On the memory side of matters, Aurora has 1.36 PB of on-package HBM2E memory and 19.9 PB of DDR5 memory that is used by the CPUs as well as 8.16 PB of HBM2E carried by the Ponte Vecchi compute GPUs.
The Aurora machine uses 166 racks that house 64 blades each. It spans eight rows and occupies a space equivalent to two basketball courts. Meanwhile, that does not count the storage subsystem of Aurora, which employs 1,024 all-flash storage nodes offering 220PB of storage capacity and a total bandwidth of 31 TB/s. For now, Argonne National Laboratory does not publish official power consumption numbers for Aurora or its storage subsystem.
The supercomputer, which will be used for a wide variety of workloads from nuclear fusion simulations to whether prediction and from aerodynamics to medical research, uses HPE's Shasta supercomputer architecture with Slingshot interconnects. Meanwhile, before the system passes ANL's acceptance tests, it will be used for large-scale scientific generative AI models.
"While we work toward acceptance testing, we are going to be using Aurora to train some large-scale open-source generative AI models for science," said Rick Stevens, Argonne National Laboratory associate laboratory director. "Aurora, with over 60,000 Intel Max GPUs, a very fast I/O system, and an all-solid-state mass storage system, is the perfect environment to train these models."
Even though Aurora blades have been installed, the supercomputer still has to undergo and pass a series of acceptance tests, a common procedure for supercomputers. Once it successfully clears these and comes online later in the year, it is projected to attain a theoretical performance exceeding 2 ExaFLOPS (two billion billion floating point operations per second). With vast performance, it is expected to secure the top position in the Top500 list.
The installation of the Aurora supercomputer marks several milestones: it is the industry's first supercomputer with performance higher than 2 ExaFLOPS and the first Intel'-based ExaFLOPS-class machine. Finally, it marks the conclusion of the Aurora saga that began eight years ago as the supercomputer's journey has seen its fair share of bumps.
Originally unveiled in 2015, Aurora was initially intended to be powered by Intel's Xeon Phi co-processors and was projected to deliver approximately 180 PetaFLOPS in 2018. However, Intel decided to abandon the Xeon Phi in favor of compute GPUs, resulting in the need to renegotiate the agreement with Argonne National Laboratory to provide an ExaFLOPS system by 2021.
The delivery of the system was further delayed due to complications with compute tile of Ponte Vecchio due to the delay of Intel's 7 nm (now known as Intel 4) production node and the necessity to redesign the tile for TSMC's N5 (5 nm-class) process technology. Intel finally introduced its Data Center GPU Max products late last year and has now shipped over 60,000 of these compute GPUs to ANL.
Although it’s a very flexible term, supercomputing generally refers to the idea of running multiple computers as one, dividing up the work between them so that they process in parallel.
In theory, every time you double the amount of processing power available, you half the time needed to complete your task. This concept of ‘clusters’ of computers has been implemented heavily in large data processing operations, including intensive graphics work such as Pixar’s famous ‘render farm’. Normally the domain of large organisations, supercomputing is now in the hands of the masses in the form of education projects and makes from the cluster-curious, but there have also been some impressive real-world applications. Here, we’ll look at some amazing projects and get you started on your own supercomputing adventure.
One of the first high-profile cluster projects surprisingly came from the boffins at GCHQ (Government Communications Headquarters) in the UK. Created as part of their educational outreach programme, the OctaPi used eight Raspberry Pi 3B computers to create a cluster. Kits were loaned out to schools with multiple coding projects to engage young minds. The first demonstrated how supercomputing could speed up difficult equations by calculating pi. A more advanced, and very appropriate, task showed how these eight machines could work together to crack a wartime Enigma code in a fraction of the time it would have taken Bletchley Park.
As we’ve already said, most Raspberry Pi cluster projects are for education or fun, but there are those who take it seriously. The Raspberry Pi Compute Module form factor is perfect for building industrial-grade supercomputers, and that’s exactly what Turing Pi has done. Their custom Turing Pi 1 PCB can accept up to seven Raspberry Pi 3+ Compute Modules and takes care of networking, power, and USB connectivity. Although claiming a wide range of uses, it appears to have found a niche in the Kubernetes world, being a surprisingly powerful device for its price. Future plans have been announced for the Turing Pi 2, based on the more powerful Raspberry Pi 4.
Multiple machines are one thing, but there’s also the individual speed of those machines. The faster they go, the faster the cluster operates exponentially. Overclocking is common in supercomputing, and that means heat. This water-cooled cluster, which maker Michael Klements freely admits is one of those ‘just because’ undertakings, uses the kind of water cooling usually found on high-end gaming PCs and applies it to a Raspberry Pi cluster. This beautiful build, complete with laser-cut mounts and elegant wiring, has been extensively documented by Klements in his blog posts. We can’t wait to see what he does with it!
So how far can we take this? Who has built the largest Raspberry Pi cluster? A strong contender seems to be Oracle, who showed off their efforts at Oracle OpenWorld in 2019. No fewer than 1060 Raspberry Pi 3B+ computers were used in its construction (that’s 4240 cores). Why 1060? That’s as much as they could physically fit in the frame! The creation has no particular purpose bar a demonstration of what is possible in a small space, cramming in several network switches, arrays of USB power supplies, and a NAS (network-attached storage) for boot images.
We’re thinking you probably don’t fancy trying to beat Oracle’s record on your first attempt, and would like to start with something a bit simpler. Our sister magazine, The MagPi, has published a cluster project you can make at home with any number of Raspberry Pi devices (although just one might be a little pointless). In this case, four Raspberry Pi 4B computers were assigned the job of searching for prime numbers. Each is assigned a different starting number, and then each adds four before testing again. This is handled by an open-source cluster manager, MPI (Message Passing Interface). A solid introduction to what is possible.
Each month, HackSpace magazine brings you the best projects, tips, tricks and tutorials from the makersphere. You can get it from the Raspberry Pi Press online store or your local newsagents. As always, every issue is free to download from the HackSpace magazine website.
The post Supercomputing with Raspberry Pi | HackSpace 41 appeared first on Raspberry Pi.
Lateo.net - Flux RSS en pagaille (pour en ajouter : @ moi) 