NVIDIA L40S GPU Overview: Characteristics, Performance, AI Use Cases
- By Gcore
- March 21, 2024
- 6 min read
The new NVIDIA L40S is a formidable contender in the AI and graphics server marketplace. It’s a popular choice across industries and use cases for its power and performance, evidenced by long purchase wait times due to high demand. This high-performance computing solution is designed to handle generative AI, LLM training, and inference tasks with a moderate workload and low-precision arithmetic calculations. Additionally, its support for ray tracing means the L40S GPU can tackle tasks involving graphics-intensive workloads. In this article, we’ll take a deep dive into how the L40S GPU performs on AI tasks compared to other NVIDIA options and identify the use cases for which it’s the optimal choice.
Why Performance Matters in AI
Training and inference, the most computationally demanding steps in an AI workflow, both require high-performance GPUs due to their complex calculations and large dataset handling. AI training involves teaching an AI system to understand and learn from the data provided to it. AI inference uses a trained model to analyze new data and produce new outputs. The superior parallel processing abilities of GPUs accelerate these tasks, significantly reducing training times and enabling rapid inference, essential for real-time applications.
As a high-performance GPU, the NVIDIA L40S delivers high throughput and low latency with maximum efficiency across use cases. We need to measure the performance of a GPU to see how it fits a particular use case. This means assessing its training time, power consumption, temperature, and GPU utilization, as well as the throughput and latency of the AI application.
With this understanding, let’s benchmark the L40S against two other high-performing NVIDIA GPUs, the NVIDIA A100 and H100, to explore the L40S’s efficiency and effectiveness across various demanding tasks.
L40S Performance Benchmarking
Benchmarking, a method for comparing the performance of one product with another, is helpful because it provides quick results and is simple to execute. In contrast to other evaluation methods, such as user reviews or theoretical analysis, it provides empirical data. It allows direct comparisons between NVIDIA’s L40S, A100, and H100. The MLPerf inference benchmark suite is a popular tool for measuring the speed at which a system can run models across diverse deployment scenarios.
How the L40S Stacks Up
In tests using this tool, the L40S relative performance is slightly weaker than the A100 on average, but when we compare the cost of each chip relative to its performance, the results for the L40S are noticeably stronger.
However, the H100 GPU outperforms the A100 and L40S by a factor of 2.3 to 2.5. This makes the H100 GPU best suited for AI tasks that involve building complex models from scratch, requiring high precision and extensive computational effort.
The NVIDIA H100 and A100 chips are priced at 4x and 2.6x the cost of the L40S, respectively—a significant price disparity, reflecting not just raw performance but also relative efficiency in handling complex AI workloads. The choice between these GPUs should thus consider both immediate budget constraints and the specific computational demands of the project.
In summary, if budget concerns are of primary importance, the NVIDIA L40S GPU is a solid choice for certain training and inference projects, including tasks involving the customization of existing trained models like OpenAI GPT, where the L40S offers sufficient performance at a more accessible price point. If absolute performance is the most important factor, the H100 is a superior choice.
L40S Use Cases
With this benchmarking in mind, let’s explore five use cases and see what the L40S’s potential is relative to other NVIDIA chip options.
AI Inference
The NVIDIA L40S GPU, explicitly designed for AI inference tasks, achieves up to 1.5x greater inference performance than the A100 GPU. This enhancement is due to the use of Ada Lovelace Tensor Cores, which support fine-grained structured sparsity to boost inference speeds and employ 8-bit floating point (FP8) precision. FP8 precision allows for up to 4X higher inference performance compared to the previous generation GPUs. By opting for FP8 calculations over the more memory-intensive FP32 and FP64 formats, the Ada Lovelace Tensor Cores significantly reduce memory demands and enhance AI processing speeds.
The following performance results showcase how the L40S compares to the A100 across various deep learning models:
But the NVIDIA H100 GPU emerges as the top performer, delivering approximately twice the performance of the A100 for inference tasks using lower precision calculations (FP8 and FP16). This superior performance is attributed to the H100’s integration of a new FP8 data type, which significantly enhances calculation rates, quadrupling those of FP16 seen in the A100.
If you want the absolute top-performer GPU to run AI inference, the NVIDIA H100 is the one to go for. However, if you’re looking for a chip that provides a solid performance for AI inference tasks, comes with a more affordable price tag, and offers accessibility, the NVIDIA L40S is a better option.
High-Performance Computing
High-performance computing (HPC) refers to the technology capable of processing complex calculations at high speeds using clusters of powerful processors working in parallel. Regarding HPC processing improvement, the NVIDIA L40S GPU delivers up to 2.2 times the performance of the A100 GPU and the H100 GPU in simulation tasks using Altair nanoFluidX software, designed for complex workflow simulations in sectors such as automotive or aerospace.
Deep learning is a subset of machine learning that uses multi-layer neural networks for tasks using HPC technology. In deep learning tasks with FP32 (32-bit floating point in data representation where ultra-high precision isn’t crucial, the L40S chip often outperforms the A100 chip in computational performance.
Generative AI
In generative AI model training, the L40S GPU demonstrates 1.2 times the performance of the A100 GPU when running Stable Diffusion—a text-to-image modeling technique developed by Stability AI that has been optimized for efficiency, allowing users to create diverse and artistic images based on text prompts. However, the H100 GPU enhances performance by a factor of 1.25 compared to the L40S for the same task.
While both the L40S and the H100 GPUs deliver solid performance for generative AI training tasks, if the goal is to maximize performance for these tasks then the H100 GPU is the preferred choice. However, when considering the relative cost of the chips, the L40S delivers superior value.
AI Graphics
The NVIDIA L40S is designed to support graphics-related AI applications, such as interactive videos and AI-driven design and automation. It boasts advanced relevant inference capabilities, complemented by NVIDIA RTX™-accelerated ray tracing and specialized engines for encoding and decoding, enhancing a wide range of AI-powered applications in audio, speech, and video generation, in both 2D and 3D formats.
For professional visualization workflows demanding high fidelity—such as real-time rendering, product design, and 3D content creation—the L40S GPU is equipped with 142 third-generation Ray Tracing (RT) Cores and a substantial 48GB memory capacity, allowing it to deliver up to twice the real-time ray-tracing performance of the A100 and H100. Ray tracing, a technique for simulating scene lighting to produce physically accurate reflections, refractions, shadows, and indirect lighting, benefits from RT cores’ efficient operation, allowing rapid graphics rendering. Artists and engineers can craft immersive visual experiences and photorealistic content with remarkable speed. This, combined with its cost-effectiveness and reduced lead times, positions the L40S as the preferred choice for building AI models focused on image generation.
Scientific Simulations
The NVIDIA L40S is equipped with 18,176 CUDA® cores, giving it ample computational power for demanding workflows like engineering and scientific simulations. The L40S achieves single-precision floating-point (FP32) performance nearly five times that of the NVIDIA A100 GPU. This means the L40S can perform complex calculations and handle data-intensive tasks quickly, making it an excellent choice for researchers and engineers who require quick, accurate results from their simulations.
When Should You Not Use the L40S?
For certain use cases such as training complex AI models, the L40S might not be the best option. In this section, we’ll discuss scenarios where the L40S is potentially not the best choice, and what you should consider as an alternative.
Building Original Models
For building and training ML models from scratch, the H100 is the preferred GPU. It’s a high-end GPU designed for AI and machine learning workloads, featuring more CUDA cores, additional memory, and higher bandwidth than the L40S. These advantages make the H100 more capable than the L40S for these specific tasks.
GPU Clusters
If your objective is to create a GPU cluster, combining multiple GPUs to form a powerful resource for AI inference and modeling, you should consider the NVIDIA A100 or H100 GPUs instead of the L40S. These models support InfiniBand and NVLink technology, enabling effective GPU clustering for more demanding tasks. InfiniBand is a powerful new architecture designed to support I/O connectivity for internet infrastructure, thereby enabling high-speed communication between interconnected nodes. NVLink is a rapid GPU interconnect that connects two NVIDIA graphics cards.
While it is feasible to create GPU clusters using L40S GPUs, such clusters may not achieve the efficiency levels of A100 or H100 GPU clusters. This discrepancy is due to potential high latency and lower throughput, resulting from slower data transfer rates in L40S GPU clusters.
Conclusion
The NVIDIA L40S is ideally suited for training and inference for models that require specialized image processing and 3D graphics capabilities, thanks to its strong ray-tracing performance. It presents a cost-effective alternative, requiring a lower initial investment than its counterparts, the A100 and H100 GPUs. Despite the advantages of the NVIDIA L40S GPU, purchasing one is a significant investment and requires expertise to set up and maintain. It can also be challenging to assess whether the L40S aligns with the specific demands of your workload, especially when relying on general benchmarks.
We offer access to the industry-leading NVIDIA A100s and H100s through Gcore Edge AI GPU Infrastructure. Gcore Inference at the Edge, slated for launch in Q2 of 2024, will be powered by L40S chips, allowing you to benefit from the L40S capabilities in a simple and cost-effective way. We can help you determine the most suitable GPU for your AI or HPC workload. Just get in touch for a consultation.
Related articles
What are virtual machines?
A virtual machine (VM), also called a virtual instance, is a software-based version of a physical computer. Instead of running directly on hardware, a VM operates inside a program that emulates a complete computer system, including a processor, memory, storage, and network connections. This allows multiple VMs to run on a single physical machine, each with its own operating system and applications, as if they were independent computers.VMS are useful because they provide flexibility, isolation, and scalability. Since each VM is self-contained, it can run different operating systems (like Windows, Linux, or macOS) on the same hardware without affecting other VMs or the host machine. This makes them ideal for testing software, running legacy applications, or efficiently using server resources in data centers. Because VMs exist as software, they can be easily copied, moved, or backed up, making them a powerful tool for both individuals and businesses.Read on to learn about types of VMs, their benefits, common use cases, and how to choose the right VM provider for your needs.How do VMs work?A virtual machine (VM) runs inside a program called a hypervisor, which acts as an intermediary between the VM and the actual computer hardware. Every time a VM needs to perform an action—such as running software, accessing storage, or using the processor—the hypervisor intercepts these requests and decides how to allocate resources like CPU power, memory, and disk space. You can think of a hypervisor as an operating system for VMs, managing multiple virtual machines on a single physical computer. Popular hypervisors like VirtualBox and VMware enable users to run multiple operating systems simultaneously while providing strong isolation.Modern hypervisors optimize performance by giving VMs direct access to certain hardware components when possible, reducing the need for constant intervention. However, some level of overhead remains because the hypervisor still needs to manage and coordinate resources efficiently. This means that while VMs can leverage most of the system’s hardware, they can’t use 100% of it, as some processing power is always reserved for managing virtualization itself. This small trade-off is often worth it, as hypervisors keep each VM isolated and secure, preventing one VM from interfering with another.VM layersFigure 1 illustrates the layers of a system virtual machine setup. The layer model can vary depending on the hypervisor. Some hypervisors include a built-in host operating system, while modern hardware offers native virtualization support. Many hypervisors can also manage multiple physical machines and VMs efficiently.VM snapshots are an essential feature in cloud computing, allowing users to quickly restore a virtual machine to a previous state.Figure 1: Layers of system virtual machinesHypervisors that emulate hardware architectures different from what the guest OS expects have a bigger overhead, as they can’t relay commands directly to the hardware without first translating them.VM snapshotsVM snapshots are an essential feature in cloud computing, allowing users to quickly restore a virtual machine to a previous state. The hypervisor can save the complete state of the VM and restore it at a later time to skip the boot process of the guest OS. The hypervisor can also move these snapshots between different physical machines, making the software running in the VM completely independent from the underlying hardware.What are the benefits of using VMs?Virtual machines offer benefits including resource efficiency, isolation, simplified operations, easy migration, faster deployment, cost savings, and security. Let’s look at these one by one.Multiple VMs can run on a single physical machine, making sharing resources between various guest operating systems easier. This is especially important when each guest OS needs to be isolated from the others, such as when they belong to different customers of a cloud service provider. Sharing resources through VMs makes running a server cheaper because you don’t have to buy or rent a whole physical machine, but only parts of it.Since VMs abstract the underlying hardware, they also improve resilience. If the physical machine fails, the hypervisor can perform a quick recovery by moving the snapshots to another machine without changing the guest OS installations to minimize downtime. This abstraction also allows operations teams to focus their deployment efforts on a standardized VM instead of considering different physical implementations.Migrations become easier with snapshots as you can simply move them to a faster machine without modifying the software running inside the VM.Faster deployments are possible because starting a VM is just a software execution instead of setting up a physical server in a data center. While you had to buy a server or rent it for months, with fast deployments, you can now rent a machine for hours, minutes, or even seconds, which allows for quite some savings.Modern CPUs have built-in virtualization features that enable easy resource sharing and enforce the isolation at the hardware layer. This prevents the services of one VM from accessing the resources of the others, improving security compared to running multiple apps inside one OS.Common use cases for VMsVMs have a range of use cases. Let’s look at the most popular ones.Cloud computingThe most popular use case is cloud computing, where VMs allow the secure sharing of the cloud provider’s resources, enabling their customers to rent only the resources they need for the period their workload will run.Software development and testingSoftware development often requires specific tools and libraries that aren’t available on a production machine, so having a development VM with all these tools preinstalled can be helpful. An example is cloud IDEs, which look and feel like regular IDEs but run on a cloud VM. A developer can have one for each project with the required dev tools installed.VMs also allow a developer to set up a machine for software testing that looks exactly like the production environment. Here, the opposite of the development VM is required; it should not have any development tools installed because they would also be missing from production.Cross-platform developmentA special case of the software development use case is cross-platform development. When you implement an app for Android or iOS, for example, you usually don’t do this on a mobile device but on your computer. With VMs, developers can simulate different hardware environments, enabling cross-platform testing without requiring physical devices.Legacy system supportIf the hardware your application requires is no longer in production, a VM might be the only way to keep running your software without reimplementing it. This is similar to the cross-platform development use case, as the VM emulates different hardware, but the difference is that the hardware no longer exists.How to choose the right VM providerTo find the right provider for your workload, the most important factor to assess is your own workload requirements. Ask the following questions and compare the answers to what providers offer.Is your workload compute or I/O-bound?Many workloads, like web servers, are I/O-bound. They don’t make complex calculations but rather simply load data and send it over the network. If you need a VM for an I/O-bound workload, you care more about disk and memory size, as well as network speed.However, compute-heavy workloads, such as AI inference or Kubernetes clusters, require careful resource allocation. If you’re evaluating whether to run Kubernetes on bare metal or VMs, check out our white paper on Bare Metal vs. VM-based Kubernetes Clusters for an in-depth comparison.If your workload is compute-bound instead, you need a high-performance CPU or a GPU and loads of memory. An AI inference engine, for example, only sends a bit of text to a client, but it does many calculations to generate this text.How long will your workload run?Web servers usually run indefinitely, but some workloads only run a few hours or minutes. If you’re doing AI training, you don’t want to pay for your huge VM cluster 24/7 if it only runs a few hours or days a week. In such cases, looking for a provider that allows renting your desired VM type hourly on a pay-as-you-go model might be worthwhile.Certain cloud providers offer cost-effective spot instances, which provide lower prices for non-critical workloads that can tolerate interruptions. These cheap VMs can get shut down at any time with minimal notice, but if your calculations aren’t time-critical, you might save quite a bit of money here.How does your workload scale?Scaling in the cloud is usually done horizontally. That is, by adding more VMs and distributing the work between them. Workloads can have different requirements for when and how fast they must be added and removed.In the AI training example, you might know in advance that one training takes more resources than the other, so you can provision enough VMs when starting. However, a web server workload might change its requirements constantly. Hence, you need a load balancer that automatically scales the instances up and down depending on the number of clients that want to access your service.Do you handle sensitive data?You might have to comply with specific laws and regulations depending on your jurisdiction(s) and industry. This means you must check whether the cloud provider also complies. How secure are their data centers? Where are they located? Do they support encryption in transit, at rest, and in process?What are your reliability requirements?Reliability is a question of costs and, again, of compliance. You might get into financial or regulatory troubles if your workload can’t run. Cloud providers often boast about their guaranteed uptimes, but remember that 99% uptime a year still means over three days of potential downtime. Check your needs and then seek a provider that can meet them reliably.Do you need customer support?If your organization doesn’t have the know-how for operating VMs in the cloud, you might need technical support from the provider. Most cloud providers are self-service, offering you a GUI and an API to manage resources. If your business lacks the resources to operate VMs, seek out a provider that can manage VMs on your behalf.SummaryVMs are a core technology for cloud computing and software development alike. They enable efficient resource sharing, improve security with hardware-enforced guest isolation, and simplify migration and disaster recovery. Choosing the right VM provider starts with understanding your workload requirements, from resource allocation to security and scalability.Maximize cloud efficiency with Gcore Virtual Machines—engineered for high performance, seamless scalability, and enterprise-grade security at competitive pricing. Whether you need to run workloads at scale or deploy applications in seconds, our VMs provide enterprise-grade security, built-in resilience, and optimized resource allocation, all powered by cutting-edge infrastructure. With global reach, fast provisioning, egress traffic included, and pay-as-you-go pricing, you get the scalability and reliability your business needs without overspending. Start your journey with Gcore VMs today and experience cloud computing that’s built for speed, security, and savings.Discover Gcore VMs
How to deploy DeepSeek 70B with Ollama and a Web UI on Gcore Everywhere Inference
Large language models (LLMs) like DeepSeek 70B are revolutionizing industries by enabling more advanced and dynamic conversational AI solutions. Whether you’re looking to build intelligent customer support systems, enhance content generation, or create data-driven applications, deploying and interacting with LLMs has never been more accessible.In this tutorial, we’ll show you exactly how to set up DeepSeek 70B using Ollama and a Web UI on Gcore Everywhere Inference. By the end, you’ll have a fully functional environment where you can easily interact with your custom LLM via a user-friendly interface. This process involves three simple steps: deploying Ollama, deploying the web UI, and configuring the web UI and connecting to Ollama.Let’s get started!Step 1: Deploy OllamaLog in to Gcore Everywhere Inference and select Deploy Custom Model.In the model image field, enter ollama/ollama.Set the Port to 11434.Under Pod Configuration, configure the following:Select GPU-Optimized.Choose a GPU type, such as 1×A100 or 1×H100.Choose a region (e.g., Luxembourg-3).Set an autoscaling policy or use the default settings.Name your deployment (e.g., ollama).Click Deploy model on the right side of the screen.Once deployed, you’ll have an Ollama endpoint ready to serve your model.Step 2: Deploy the Web UI for OllamaGo back to the Gcore Everywhere Inference console and select Deploy Custom Model again.In the Model Image field, enter ghcr.io/open-webui/open-webui:main.Set the Port to 8080.Under Pod Configuration, set:CPU-Optimized.Choose 4 vCPU / 16 GiB RAM.Select the same region as before (e.g., Luxembourg-3).Configure an autoscaling policy or use the default settings.Name your deployment (e.g., webui).Click Deploy model on the right side of the screen.Once deployed, navigate to the Web UI endpoint from the Gcore Customer Portal.Step 3: Configure the Web UIFrom the Web UI endpoint and set up a username and password when prompted.Log in and navigate to the admin panel.Go to Settings → Connections → Disable the OpenAI API integration.In the Ollama API field, enter the endpoint for your Ollama deployment. You can find this in the Gcore Customer Portal. It will look similar to this: https://<your-ollama-deployment>.ai.gcore.dev/.Click Save to confirm your changes.Step 4: Pull and Use DeepSeek 70BOpen the chat section in the Web UI.In the Select a model field, type deepseek-r1:70b.Click Pull to download the model.Wait for the download to complete.Once downloaded, select the model and start chatting!Your AI environment is ready to exploreBy following these steps, you’ve successfully deployed DeepSeek 70B on Gcore Everywhere Inference with Ollama. This setup provides a powerful and user-friendly environment for experimenting with LLMs, prototyping AI-driven features, or integrating advanced conversational AI into your applications.Ready to unlock the full potential of AI? Gcore Everywhere Inference offers outstanding scalability, performance, and support, making it the perfect solution for developers and businesses working with advanced AI models. Dive deeper into our powerful tools and resources by exploring our AI blog and docs.Discover Gcore Everywhere Inference
What is AI inference and how does it work?
Artificial intelligence (AI) inference is what happens when a trained AI model is used to predict outcomes from new, unseen data. While training focuses on learning from historical datasets, inference is about putting that learned knowledge into action—such as identifying production bottlenecks before they happen, converting speech to text, or guiding self-driving cars in real time. This article walks you through the basics of AI inference and shows how to get started.What is AI inference?AI inference is the application phase of artificial intelligence. Once a model has been trained on large datasets, it shifts from “learning mode” to “doing mode”—providing predictions or decisions from new data inputs.For example, an e-commerce platform with a model trained on purchasing behavior uses inference to personalize recommendations for each site visitor. Without re-training from scratch, the model quickly adapts to new browsing patterns and purchasing signals, offering instant, relevant suggestions.By enabling actionable insights, inference is transforming how businesses and technologies function, empowering relevance and instant responsiveness in an increasingly data-driven world.How does AI inference work? A practical guideAI inference has four steps: data preparation, model loading, processing and prediction, and output generation.#1 Data preparationThe first step involves transforming raw input—such as text, images, or numerical data—into a format that the AI model can process. For instance, customer feedback might be converted into numerical representations of words and patterns, or an image could be resized and normalized. Proper data preparation ensures that the AI model can effectively understand and analyze the input. For businesses, this means making sure that input data is clean, well-structured, and formatted according to the model’s requirements.#2 Model loadingOnce the input data is ready, the trained AI model is loaded into memory. This model, equipped with patterns and relationships learned during training, acts as the foundation for predictions and decisions.Businesses must make sure that their infrastructure is capable of quickly loading and deploying AI models, especially during high-demand periods. We simplify this process by providing a high-performance platform with global scalability. Your models are loaded and operational in seconds, whether you’re using a custom model or an open-source one.#3 Processing and predictionIn this step, the prepared data is passed through the model’s neural networks, which apply learned patterns to generate insights or predictions. For example, a customer service AI might analyze incoming messages to determine if they express satisfaction or frustration.The speed and accuracy of this stage depend on access to low-latency infrastructure capable of handling complex calculations. Our edge inference solution means data processing happens close to the source, reducing latency and enabling real-time decision making.#4 Output generationThe final stage translates the model’s mathematical outputs into meaningful insights, such as predictions, labels, or recommendations. These outputs must be integrated into business workflows or customer-facing applications in a way that’s easy to understand and actionable.We help streamline this step by offering APIs and integration tools that allow businesses to seamlessly incorporate inference results into their operations, so outputs are accessible and actionable in real time.A real-life exampleLet’s look at how this works in practice. Consider a retail business implementing AI for inventory management. The system continuously:Receives data from point-of-sale systems and warehouse scannersProcesses this information through trained AI modelsGenerates predictions about future inventory needsAdjusts order quantities and timing automaticallyAll of this happens in milliseconds, making real-time decisions possible. However, the speed and efficiency depend on choosing the right infrastructure for your needs.The technology stack behind inferenceTo make this process work smoothly, specialized computing infrastructure and software need to work together.Computing infrastructureModern AI inference relies on specialized hardware designed to process mathematical operations quickly. While training AI models often requires expensive, high-powered graphics processors (GPUs), inference can run on more cost-effective hardware options:CPUs: Suitable for smaller-scale applicationsEdge devices: For processing data locally on smartphones or IoT devices or other hardware closer to the data source, resulting in low latency and better privacy.Cloud-based inference servers: Designed for handling large-scale operations, enabling centralized processing and flexible scaling.When evaluating computing infrastructure for AI, businesses should prioritize solutions that address latency, scalability, and ease of use. Edge inference capabilities are essential for deploying models closer to end users, which optimizes performance globally even during peak demand. Flexible access to diverse hardware options like GPUs, CPUs, and advanced accelerators ensures adaptability, while user-friendly tools and automated scaling enable seamless management and consistent performance.Software optimizationThe efficiency of inference depends heavily on software optimization. When done right, software optimization ensures that AI applications are fast, responsive, and scalable, making them practical for real-world use.Look for the following to identify a solution that reduces inference processing time and supports optimized results:Model compression and optimization: The computational load is reduced and inference occurs faster—without sacrificing accuracy.Workload distribution and automation: This means that resources are allocated efficiently and cost-effectively.Integration: Look for APIs and tools that connect seamlessly with existing business systems.The future of AI inferenceWe anticipate three major trends for the future of AI inference.First, we’re seeing a dramatic shift toward specialized AI accelerators and custom silicon. New chips are being developed and existing ones optimized specifically for inference workloads. These purpose-built processors are delivering significant improvements in both performance and energy efficiency compared to traditional GPUs. This specialization is making AI inference more cost-effective and environmentally sustainable, particularly for companies running large-scale operations.The second major trend is the emergence of lightweight, efficient models designed specifically for inference. While large language models like GPT-4 showcase the potential of AI, many businesses are finding that smaller, task-specific models can deliver comparable or better results for their particular needs. These “small language models” (SLMs) and domain-adapted models are trained on focused datasets and optimized for specific tasks, making them more practical for real-world deployment. This approach is particularly valuable for edge computing scenarios where computing resources are limited.Finally, the infrastructure for AI inference is becoming more sophisticated and accessible. Advanced orchestration tools are automating the complex process of model deployment, scaling, and monitoring. These platforms can automatically optimize model performance based on factors like latency requirements, cost constraints, and traffic patterns. This automation is making it possible for companies to deploy AI solutions without maintaining large specialized teams of ML engineers.Dive into more of our predictions for AI inference in 2025 and beyond in our dedicated article.Accelerate inference adoption for your businessAI inference is rapidly becoming a differentiator for businesses. By applying trained AI models to new data, companies can make instant predictions, automate decision-making, and optimize operations across industries. However, achieving these benefits depends on having the right infrastructure and expertise behind the scenes. This is where the choice of inference provider plays a critical role. The provider’s infrastructure determines latency, scalability, and overall efficiency, which directly affect business outcomes. A well-equipped provider allows businesses to maximize the value of their AI investments.At Gcore, we are uniquely positioned to meet these needs with our edge inference solution. Leveraging a secure, global network of over 180 points of presence equipped with NVIDIA GPUs, we deliver ultra-fast, low-latency inference capabilities. Intuitively deploy and scale open-source or custom models on our powerful platform that accelerates AI adoption for a competitive edge in an increasingly AI-driven world.Get a complimentary consultation about your AI inference needs
AI model selection simplified: your guide to Gcore-supported model selection
2024 has been an exceptional year for advancements in artificial intelligence (AI). The variety of models has grown significantly, with impressive strides in performance across domains. Whether it’s text or image classification, text and image generation, speech models, or multimodal capabilities, businesses now face the challenge of navigating an ever-expanding catalog of open-source models. Understanding the differences in tasks and metrics targeted by these models is crucial to making informed decisions.At Gcore, we’ve been expanding our model catalog to simplify AI model testing and deployment. As businesses scale their AI applications across various units, identifying the best model for specific tasks becomes critical. For example, some applications, like cancer screening, prioritize accuracy over latency. On the other hand, time-sensitive use cases like fraud detection demand rapid processing, while cost may drive decisions for lightweight applications like chatbot development.This guide provides a comprehensive overview of the AI models supported on the Gcore platform, their characteristics, and their most effective use cases to help you choose the right model for your needs. Our inference solution also supports custom AI models.Large language models (LLMs)LLMs are foundational for applications requiring human-like understanding and generation of text, making them crucial for customer service, research, and educational tools. These models are versatile and cover a range of applications:Text generation (e.g., creative writing, content creation)SummarizationQuestion answeringInstruction following (specific to instruct-tuned models)Sentiment analysisTranslationCode generation and debugging (if fine-tuned for programming tasks)Models supported by GcoreGcore supports the following models for inference, available in the Gcore Customer Portal. Activate them at the click of a button.Model nameProviderParametersKey characteristicsLLaMA-Pro-8BMeta AI8 BillionBalanced trade-off between cost and power, suitable for real-time applications.Llama-3.2-1B-InstructMeta AI1 BillionIdeal for lightweight tasks with minimal computational needs.Llama-3.2-3B-InstructMeta AI3 BillionOffers lower latency for moderate task complexity.Llama-3.1-8B-InstructMeta AI8 BillionOptimized for instruction following.Mistral-7B-Instruct-v0.3Mistral AI7 BillionExcellent for nuanced instruction-based responses.Mistral-Nemo-Instruct-2407Mistral AI & Nvidia7 BillionHigh efficiency with robust instruction-following capabilities.Qwen2.5-7B-InstructQwen7 BillionExcels in multilingual tasks and general-purpose applications.QwQ-32B-PreviewQwen32 BillionSuited for complex, multi-turn conversations and strategic decision-making.Marco-o1AIDC-AI1-5 Billion (est.)Designed for structured and open-ended problem-solving tasks.Business applicationsLLMs play a pivotal role in various business scenarios; choosing the right model will be primarily influenced by task complexity. For lightweight tasks like chatbot development and FAQ automation, models like Llama-3.2-1B-Instruct are highly effective. Medium complexity tasks, including document summarization and multilingual sentiment analysis, can leverage models like Llama-3.2-3B-Instruct and Qwen2.5-7B-Instruct. For high-performance needs like real-time customer service or healthcare diagnostics, models like LLaMA-Pro-8B and Mistral-Nemo-Instruct-2407 provide robust solutions. Complex, large-scale applications, like market forecasting and legal document synthesis, are ideally suited for advanced models like QwQ-32B-Preview. Additionally, specialized solutions for niche industries can benefit from Marco-o1’s unique capabilities.Image generationImage generation models empower industries like entertainment, advertising, and e-commerce to create engaging content that captures the audience’s attention. These models excel in producing creative and high-quality visuals. Key tasks include:Generating photorealistic imagesArtistic rendering (e.g., illustrations, concept art)Image enhancement (e.g., super-resolution, inpainting)Marketing and branding visualsModels supported by GcoreWe currently support six models via the Gcore Customer Portal, or you can bring your own image generation model to our inference platform.Model nameProviderParametersKey characteristicsByteDance/SDXL-LightningByteDance100-400 MillionLightning-fast text-to-image generation with 1024px outputs.stable-cascadeStability AI20M-3.6 BillionWorks on smaller latent spaces for faster and cheaper inference.stable-diffusion-xlStability AI~3.5B Base + 1.2B RefinementPhotorealistic outputs with detailed composition.stable-diffusion-3.5-large-turboStability AI8 BillionBalances high-quality outputs with faster inference.FLUX.1-schnellBlack Forest Labs12 BillionDesigned for fast, local development.FLUX.1-devBlack Forest Labs12 BillionOpen-weight model for non-commercial applications.Business applicationsIn high-quality image generation, models like stable-diffusion-xl and stable-cascade are commonly employed for creating marketing visuals, concept art for gaming, and detailed e-commerce product visualizations. Real-time applications, such as AR/VR customizations and interactive customer tools, benefit from the speed of ByteDance/SDXL-Lightning and FLUX.1-schnell. FLUX.1-dev and stable-diffusion-3.5-large-turbo are excellent options for experimentation and development, allowing startups and enterprises to prototype generative AI workflows cost-effectively. Specialized use cases, such as ultra-high-quality visuals for luxury goods or architectural renders, also find tailored solutions with stable-cascade.Speech recognitionSpeech recognition models are essential for industries like media, healthcare, and education, where transcription accuracy and speed directly impact their efficacy. They facilitate:Accurate speech-to-text transcriptionLow-latency live audio conversionMultilingual speech processing and translationAutomated note-taking and content creationModels supported by GcoreAt Gcore, our inference service supports two Whisper models, as well as custom speech recognition models.Model nameProviderParametersKey characteristicswhisper-large-v3-turboOpenAI809 MillionOptimized for speed with minimal accuracy trade-offs.whisper-large-v3OpenAI1.55 BillionHigh-quality multilingual speech-to-text and translation with reduced error rates.Business applicationsSpeech recognition technology supports a wide range of business functions, all requiring precision and accuracy, delivered at speed. For real-time transcription, whisper-large-v3-turbo is ideal for live captioning and speech analytics applications. High-accuracy tasks, including legal transcription, academic research, and multilingual content localization, leverage the advanced capabilities of whisper-large-v3. These models enable faster, more accurate workflows in sectors where precise audio-to-text conversion is crucial.Multimodal modelsBy bridging text, image, and other data modalities, multimodel models unlock innovative solutions for industries requiring complex data analysis. These models integrate diverse data types for applications in:Image captioningVisual question answeringMultilingual document processingRobotic visionModels supported by GcoreWe currently support the following multimodal models:Model nameProviderParametersKey characteristicsPixtral-12B-2409Mistral AI12 BillionExcels in instruction-following tasks with text and image integration.Qwen2-VL-7B-InstructQwen7 BillionAdvanced visual understanding and multilingual support.Business applicationsFor tasks like image captioning and visual question answering, Pixtral-12B-2409 provides robust capabilities in generating descriptive text and answering questions based on visual content. Qwen2-VL-7B-Instruct supports document analysis and robotic vision, enabling systems to extract insights from documents or understand their physical surroundings. These applications are transformative for industries ranging from digital media to robotics.A multitude of models, supported by GcoreStart developing on the Gcore platform today, leveraging top-tier GPUs for seamless AI model training and deployment. Simplify large-scale, cross-regional AI operations with our inference-at-the-edge solutions, backed by over a decade of CDN expertise.Get started with Inference at the Edge today
How to Run Hugging Face Spaces on Gcore Inference at the Edge
Running machine learning models, especially large-scale models like GPT 3 or BERT, requires a lot of computing power and comes with a lot of latency. This makes real-time applications resource-intensive and challenging to deliver. Running ML models at the edge is a lightweight approach offering significant advantages for latency, privacy, and resource optimization. Gcore Inference at the Edge makes it simple to deploy and manage custom models efficiently, giving you the ability to deploy and scale your favorite Hugging Face models globally in just a few clicks. In this guide, we’ll walk you through how easy it is to harness the power of Gcore’s edge AI infrastructure to deploy a Hugging Face Space model. Whether you’re developing NLP solutions or cutting-edge computer vision applications, deploying at the edge has never been simpler—or more powerful. Step 1: Log In to the Gcore Customer PortalGo to gcore.com and log in to the Gcore Customer Portal. If you don’t yet have an account, go ahead and create one—it’s free. Step 2: Go to Inference at the EdgeIn the Gcore Customer Portal, click Inference at the Edge from the left navigation menu. Then click Deploy custom model. Step 3: Choose a Hugging Face ModelOpen huggingface.com and browse the available models. Select the model you want to deploy. Navigate to the corresponding Hugging Face Space for the model. Click on Files in the Space and locate the Docker option. Copy the Docker image link and startup command from Hugging Face Space. Step 4: Deploy the Model on GcoreReturn to the Gcore Customer Portal deployment page and enter the following details: Model image URL: registry.hf.space/ethux-mistral-pixtral-demo:latest Startup command: python app.py Container port: 7860 Configure the pod as follows: GPU-optimized: 1x L40S vCPUs: 16 RAM: 232GiB For optimal performance, choose any available region for routing placement. Name your deployment and click Deploy.Step 5: Interact with Your ModelOnce the model is up and running, you’ll be provided with an endpoint. You can now interact with the model via this endpoint to test and use your deployed model at the edge.Powerful, Simple AI Deployment with GcoreGcore Inference at the Edge is the future of AI deployment, combining the ease of Hugging Face integration with the robust infrastructure needed for real-time, scalable, and global solutions. By leveraging edge computing, you can optimize model performance and simultaneously futureproof your business in a world that increasingly demands fast, secure, and localized AI applications. Deploying models to the edge allows you to capitalize on real-time insights, improve customer experiences, and outpace your competitors. Whether you’re leading a team of developers or spearheading a new AI initiative, Gcore Inference at the Edge offers the tools you need to innovate at the speed of tomorrow. Explore Gcore Inference at the Edge
How to Choose Between Bare Metal GPUs and Virtual GPUs for AI Workloads
Choosing the right GPU type for your AI project can make a huge difference in cost and business outcomes. The first consideration is often whether you need a bare metal or virtual GPU. With a bare metal GPU, you get a physical server with an entire GPU chip (or chips) installed that is completely dedicated to the workloads you run on the server, whereas a virtual GPU means you share GPU resources with other virtual machines.Read on to discover the key differences between bare metal GPUs and virtual GPUs, including performance and scalability, to help you make an informed decision.The Difference Between Bare Metal and Virtual GPUsThe main difference between bare metal GPUs and virtual GPUs is how they use physical GPU resources. With a bare metal GPU, you get a physical server with an entire GPU chip (or chips) installed that is completely dedicated to the workloads you run on the server. There is no hypervisor layer between the operating system (OS) and the hardware, so applications use the GPU resources directly.With a virtual GPU, you get a virtual machine (VM) and uses one of two types of GPU virtualization, depending on your or a cloud provider’s capabilities:An entire, dedicated GPU used by a VM, also known as a passthrough GPUA shared GPU used by multiple VMs, also known as a vGPUAlthough a passthrough GPU VM gets the entire GPU, applications access it through the layers of a guest OS and hypervisor. Also, unlike a bare metal GPU instance, other critical VM resources that applications use, such as RAM, storage, and networking, are also virtualized.The difference between running applications with bare metal and virtual GPUsThese architectural features affect the following key aspects:Performance and latency: Applications running on a VM with a virtual GPU, especially vGPU, will have lower processing power and higher latency for the same GPU characteristics than those running on bare metal with a physical GPU.Cost: As a result of the above, bare metal GPUs are more expensive than virtual GPUs.Scalability: Virtual GPUs are easier to scale than bare metal GPUs because scaling the latter requires a new physical server. In contrast, a new GPU instance can be provisioned in the cloud in minutes or even seconds.Control over GPU hardware: This can be critical for certain configurations and optimizations. For example, when training massive deep learning models with a billion parameters, total control means the ability to optimize performance optimization—and that can have a big impact on training efficiency for massive datasets.Resource utilization: GPU virtualization can lead to underutilization if the tasks being performed don’t need the full power of the GPU, resulting in wasted resources.Below is a table summarizing the benefits and drawbacks of each approach: Bare metal GPUVirtual GPUPassthrough GPUvGPUBenefitsDedicated GPU resourcesHigh performance for demanding AI workloadsLower costSimple scalabilitySuitable for occasional or variable workloadsLowest costSimple scalabilitySuitable for occasional or variable workloadsDrawbacksHigh cost compared to virtual GPUsLess flexible and scalable than virtual GPUsLow performanceNot suitable for demanding AI workloadsLowest performanceNot suitable for demanding AI workloadsShould You Use Bare Metal or Virtual GPUs?Bare metal GPUs and virtual GPUs are typically used for different types of workloads. Your choice will depend on what AI tasks you’re looking to perform.Bare metal GPUs are better suited for compute-intensive AI workloads that require maximum performance and speed, such as training large language models. They are also a good choice for workloads that must run 24/7 without interruption, such as some production AI inference services. Finally, bare metal GPUs are preferred for real-time AI tasks, such as robotic surgery or high-frequency trading analytics.Virtual GPUs are a more suitable choice for the early stages of AI/ML and iteration on AI models, where flexibility and cost-effectiveness are more important than top performance. Workloads with variable or unpredictable resource requirements can also run on this type of GPU, such as training and fine-tuning small models or AI inference tasks that are not sensitive to latency and performance. Virtual GPUs are also great for occasional, short-term, and collaborative AI/ML projects that don’t require dedicated hardware—for example, an academic collaboration that includes multiple institutions.To choose the right type of GPU, consider these three factors:Performance requirements. Is the raw GPU speed critical for your AI workloads? If so, bare metal GPUs are a superior choice.Scalability and flexibility. Do you need GPUs that can easily scale up and down to handle dynamic workloads? If yes, opt for virtual GPUs.Budget. Depending on the cloud provider, bare metal GPU servers can be more expensive than virtual GPU instances. Virtual GPUs typically offer more flexible pricing, which may be appropriate for occasional or variable workloads.Your final choice between bare metal GPUs and virtual GPUs depends on the specific requirements of the AI/ML project, including performance needs, scalability requirements, workload types, and budget constraints. Evaluating these factors can help determine the most appropriate GPU option.Choose Gcore for Best-in-Class AI GPUsGcore offers bare metal servers with NVIDIA H100, A100, and L40S GPUs. Using the 3.2 Tbps InfiniBand interface, you can combine H100 or A100 servers into scalable GPU clusters for training and tuning massive ML models or for high-performance computing (HPC).If you are looking for a scalable and low-latency solution for global AI inference, explore Gcore Inference at the Edge. It especially benefits latency-sensitive, real-time applications, such as generative AI and object recognition.Discover Gcore bare metal GPUs
Subscribe to our newsletter
Get the latest industry trends, exclusive insights, and Gcore updates delivered straight to your inbox.