nvidia.jpg

B200 new image.png

The Foundation for your AI Factory. NVIDIA B200 is a next-generation AI server built for training and deploying the most demanding AI workloads. Powered by 8 Blackwell GPUs and fifth-gen NVLink interconnect, it delivers up to 3x training and 15x inference performance over its predecessor. Ideal for LLMs, recommender systems, and real-time inference applications, B200 is a high-performance system designed for teams scaling production AI infrastructure with confidence.

Warranty

Nvidia B200 Specifications

Specifications

h200.png

The NVIDIA H200 is a cutting-edge GPU designed to excel in AI, machine learning, and high-performance computing (HPC) environments. Built on the advanced Hopper architecture, it features improved computational power and memory capabilities, making it well-suited for tasks like generative AI, large language models (LLMs), and complex scientific simulations. The H200 is the first to use HBM3e memory, which provides faster data throughput and larger capacity, enabling the processing of more substantial and complex workloads. It is engineered to meet the growing demands of data-intensive applications across industries.


Overview

H200 SXM

H200 NVL

rtx-4090.png

NVIDIA unveiled the GeForce RTX 4090, a top-tier graphics card for enthusiasts, on September 20th, 2022. Fabricated with a 5 nm process and based on the AD102 graphics processor, specifically the AD102-300-A1 variant, the card boasts support for DirectX 12 Ultimate. This feature ensures compatibility with all contemporary games and provides future-proofing for upcoming titles with support for hardware-ray tracing, variable-rate shading, and more. The AD102 is a substantial chip, possessing a die size of 609 mm² and 76.3 billion transistors. The GeForce RTX 4090, which has fewer shading units activated compared to the fully enabled TITAN Ada that utilizes the same GPU, includes 16384 shading units, 512 texture mapping units, and 176 ROPs. Furthermore, the card is equipped with 512 tensor cores to enhance machine learning application performance and 128 cores dedicated to ray tracing acceleration. NVIDIA has paired the card with 24 GB of GDDR6X memory, connected via a 384-bit interface. The GPU operates at a base frequency of 2235 MHz, with a boost capability up to 2520 MHz, while the memory speed is effective at 21 Gbps (1313 MHz). As a triple-slot card, the GeForce RTX 4090 requires a single 16-pin power connector and has a maximum power consumption of 450 W. It features display outputs of 1x HDMI 2.1 and 3x DisplayPort 1.4a, connecting to the system through a PCI-Express 4.0 x16 interface. The dimensions of the card are 304 mm by 137 mm by 61 mm, with a triple-slot cooling design.

GeForce 210

GeForce 9400 GT

Radeon HD 4550

Radeon HD 5450

Radeon HD 6450

GeForce GT 520

GeForce GT 220

GeForce GT 430

Radeon HD 5570

Radeon HD 4670

GeForce GT 440

GeForce GT 240

GeForce 9600 GT

Radeon HD 5670

GeForce GT 640

GeForce 9800 GT

Radeon HD 6670

Radeon HD 4830

Radeon HD 4770

Radeon HD 4850

GeForce GTS 250

GeForce GTS 450

Radeon HD 5750

Radeon HD 7750

Radeon HD 5770

GeForce GTX 260

Radeon HD 4870

Radeon Vega 8

GeForce GTX 550 Ti

GeForce GTX 650

Radeon HD 5830

Radeon HD 6790

Radeon HD 4890

GeForce GTX 460

Radeon RX Vega 11

GeForce GT 1030

GeForce GTX 275

Radeon HD 7770 GHz Edition

GeForce GTX 280

GeForce GTX 465

Radeon HD 6850

GeForce GTX 285

Radeon HD 5850

GeForce GTX 650 Ti

Radeon HD 7790

Radeon RX 550

GeForce GTX 470

Radeon HD 6870

Radeon HD 5870

GeForce GTX 560 Ti

Radeon HD 4870 X2

GeForce GTX 750 Ti

Radeon HD 6950

GeForce GTX 295

GeForce GTX 650 Ti Boost

Radeon HD 7850

GeForce GTX 480

Radeon HD 6970

Radeon R7 265

Radeon RX 460

Radeon R7 370

GeForce GTX 660

GeForce GTX 570

Radeon RX 560

GeForce GTX 950

Radeon HD 7870 GHz Edition

Radeon R9 270X

GeForce GTX 660 Ti

GeForce GTX 580

GeForce GTX 1050

Radeon HD 7950

GeForce GTX 1630

Radeon HD 5970

GeForce GTX 760

GeForce GTX 670

GeForce GTX 960

Radeon R9 380

Radeon R9 285

Radeon HD 7970

GeForce GTX 680

GeForce GTX 1050 Ti

GeForce GTX 770

Radeon HD 6990

Radeon R9 280X

GeForce GTX 590

Radeon HD 7970 GHz Edition

GeForce GTX 780

Arc A380

Radeon RX 6400

GeForce GTX 1650

Radeon RX 470

Radeon R9 290

Radeon RX 570

Radeon R9 390

Radeon R9 290X

GeForce GTX 970

GeForce GTX TITAN

GeForce GTX 780 Ti

Radeon R9 390X

Radeon HD 7990

Relative Performance

Graphic Processor

Graphics Card

Clock Speeds

Memory

Render Config

Theoretical Performance

Board Design

Graphic Features

nvidia-rtx-5090.png

The GeForce RTX 5090 will be a graphics card by NVIDIA, that is expected to launch in 2025. Built on the unknown process, and based on the GB202 graphics processor, in its GB202-400-A1 variant, the card supports DirectX 12 Ultimate. This ensures that all modern games will run on GeForce RTX 5090. Additionally, the DirectX 12 Ultimate capability guarantees support for hardware-raytracing, variable-rate shading and more, in upcoming video games. It features 21760 shading units, 680 texture mapping units, and 192 ROPs. Also included are 680 tensor cores which help improve the speed of machine learning applications. The card also has 170 raytracing acceleration cores. NVIDIA has paired 32 GB GDDR7 memory with the GeForce RTX 5090, which are connected using a 512-bit memory interface. The GPU is operating at a frequency of 2235 MHz, which can be boosted up to 2520 MHz, memory is running at 1875 MHz (23.8 Gbps effective).
Being a triple-slot card, the NVIDIA GeForce RTX 5090 draws power from 1x 16-pin power connector, with power draw rated at 500 W maximum. Display outputs include: 1x HDMI 2.1, 3x DisplayPort 1.4a. GeForce RTX 5090 is connected to the rest of the system using a PCI-Express 5.0 x16 interface. The card's dimensions are 304 mm x 137 mm x 61 mm, and it features a triple-slot cooling solution.

Graphics Processor

grace-hopper.png

The NVIDIA Grace Hopper SuperChip combines NVIDIA Grace CPU with NVIDIA Hopper GPU, linked by high-bandwidth NVLink-C2C, focusing on AI and HPC applications. It features up to 72 Arm Neoverse V2 CPU cores for high-performance and energy efficiency, alongside the NVIDIA Hopper GPU designed for significant advancements in AI and HPC tasks. The superchip boasts up to 512 GB of LPDDR5X memory with a bandwidth of 546 GB/s, expandable through NVLink-C2C, facilitating a total of 900 GB/s bandwidth accessible to the GPU, enhancing data access and processing efficiency.

rtx-a6000.png

Launched by NVIDIA on October 5th, 2020, the RTX A6000 is a high-end professional graphics card designed for enthusiasts. This card is developed using an 8 nm manufacturing process and is based on the GA102 graphics processor, featuring support for DirectX 12 Ultimate. The GA102 is a considerable chip, with a die area of 628 mm² and 28.3 billion transistors. It comes equipped with 10752 shading units, 336 texture mapping units, and 112 ROPs. Additionally, the card includes 336 tensor cores to accelerate machine learning applications and 84 cores dedicated to ray tracing acceleration. The RTX A6000 is outfitted with 48 GB of GDDR6 memory, which utilizes a 384-bit interface for connectivity. The GPU operates at a base frequency of 1410 MHz, with the capability to reach up to 1800 MHz, and the memory has an effective speed of 16 Gbps (2000 MHz). As a dual-slot graphics card, the RTX A6000 requires an 8-pin EPS connector for power and has a maximum power consumption of 300 W. It provides 4x DisplayPort 1.4a outputs for display connectivity and interfaces with the system via a PCI-Express 4.0 x16 connection. The card's dimensions are 267 mm in length and 112 mm in width, featuring a dual-slot cooling design. 

amd-logo.png

mi250.png

The AMD Radeon Instinct MI300, released on January 4, 2023, is a specialized graphics card developed by AMD, utilizing a 5 nm manufacturing process and the Aqua Vanjaram GPU. This card is not equipped with DirectX support, limiting its ability to run the latest gaming titles. The Aqua Vanjaram GPU boasts a substantial size with a 1017 mm² die area and contains 153 billion transistors. Unlike the Radeon Instinct MI300X, which shares the same GPU but has all shaders active, the MI300 has some shaders disabled to meet its shader count goals, featuring 14080 shading units and 880 tensor cores, beneficial for machine learning tasks. AMD pairs the card with 128 GB of HBM3 memory via an 8192-bit interface, with a base GPU frequency of 1000 MHz and a boost up to 1700 MHz. The card's power consumption peaks at 600 W, and it lacks display outputs, interfacing with the system through a PCI-Express 5.0 x16 slot.

Graphics Features

l40.png

NVIDIA launched the L40, a high-end professional graphics card, on October 13th, 2022. This card is crafted using the 5 nm technology and features the AD102 graphics processor in the AD102-895-A1 variant, supporting DirectX 12 Ultimate. The AD102 chip is substantial, with a 609 mm² die area and 76.3 billion transistors. Unlike the TITAN Ada, which is a fully unlocked version with the same GPU but all 18432 shaders enabled, the L40 has a reduced number of active shading units to meet its specific target, featuring 18176 shading units, 568 texture mapping units, and 192 ROPs. Additionally, the card includes 568 tensor cores for enhanced machine learning application performance and 142 cores dedicated to ray tracing acceleration. NVIDIA has equipped the L40 with 48 GB of GDDR6 memory, utilizing a 384-bit interface for connectivity. The GPU operates at a base frequency of 735 MHz with the capability to boost up to 2490 MHz, while the memory has an effective speed of 18 Gbps (2250 MHz). As a dual-slot graphics card, the L40 uses a single 16-pin power connector and has a power consumption rated at 300 W maximum. It comes with 4x DisplayPort 1.4a outputs and connects to the system via a PCI-Express 4.0 x16 interface. The card's physical dimensions are 267 mm in length and 111 mm in width, incorporating a dual-slot cooling design.

On October 13th, 2022, NVIDIA introduced the L40S, a professional graphics card crafted with the 5 nm manufacturing process. It leverages the AD102 graphics processor, specified as the AD102-???-A1 variant, and offers support for DirectX 12 Ultimate. The AD102 is a significant processor, featuring a 609 mm² die area and 76.3 billion transistors. This model, similar to the TITAN Ada which shares the same GPU but enables all 18432 shaders, has a reduced number of shading units activated in the L40S to achieve its intended shader count. It boasts 18176 shading units, 568 texture mapping units, and 192 ROPs. The card is further enhanced with 568 tensor cores to facilitate machine learning tasks and 142 cores for ray tracing acceleration. NVIDIA pairs the L40S with 48 GB of GDDR6 memory, connected through a 384-bit interface. The GPU operates at a base frequency of 1110 MHz, with a boost potential up to 2520 MHz, while the memory is set at an effective speed of 18 Gbps (2250 MHz). Designed as a dual-slot card, the L40S utilizes a single 16-pin power connector and has a maximum power consumption of 300 W. It features 1x HDMI 2.1 and 3x DisplayPort 1.4a for display outputs, ensuring connectivity to various display setups. The L40S interfaces with the system via a PCI-Express 4.0 x16 connection and measures 267 mm in length and 111 mm in width, incorporating a dual-slot cooling architecture.

rtx-3090.png

The NVIDIA GeForce RTX 3090, a high-end graphics card, was released on September 1, 2020. This card is manufactured using the 8 nm technology and is built around the GA102 graphics processor, specifically the GA102-300-A1 version. It is compatible with DirectX 12 Ultimate, ensuring it can run the latest video games and support advanced features like hardware-ray tracing, variable-rate shading, and more. The GA102 chip is substantial, with a 628 mm² die size and 28.3 billion transistors. The GeForce RTX 3090 has 10,496 shading units, compared to the 3090 Ti, which has all 10,752 shaders activated. It also includes 328 texture mapping units, 112 raster operations pipelines, 328 tensor cores for machine learning acceleration, and 82 cores dedicated to ray tracing. The card is equipped with 24 GB of GDDR6X memory, a 384-bit interface, and operates at a base clock of 1395 MHz, with a boost clock up to 1695 MHz. The memory's effective speed is 19.5 Gbps. Designed as a triple-slot card, it requires a 12-pin power connector and has a maximum power consumption of 350 W. Its output ports are 1x HDMI 2.1 and 3x DisplayPort 1.4a, and it uses a PCI-Express 4.0 x16 connection to the system. The dimensions of the card are 336 mm by 140 mm by 61 mm, and it comes with a triple-slot cooling design.

AMD's Radeon Instinct MI250, introduced on November 8, 2021, is a specialized graphics card designed for professional use. It is crafted using the 6 nm manufacturing process and is powered by the Aldebaran graphics processor, specifically the Aldebaran variant. This card is not equipped to support DirectX, meaning it may not be compatible with all recent games. The Aldebaran GPU is notably large, featuring a 724 mm² die area and housing 58,200 million transistors. The Radeon Instinct MI250X, another model with the same GPU, has all 14080 shaders activated, whereas the MI250 has fewer—13312 shading units—to meet its design specifications. Additionally, the MI250 includes 832 texture mapping units and does not have raster operation pipelines (ROPs). AMD has incorporated 128 GB of HBM2e memory into the MI250, connected via an 8192-bit interface. The GPU's base operating frequency is set at 1000 MHz, with the potential to boost up to 1700 MHz, while the memory operates at 1600 MHz. Designed as a dual-slot card, the MI250 requires two 8-pin power connectors and has a maximum power consumption of 500 W. It does not offer display outputs, aligning with its non-consumer focus. The card connects to systems through a PCI-Express 4.0 x16 interface and is dimensioned at 267 mm by 111 mm, with a dual-slot cooling system.

Procurement

Procurement
as a Service

GPU Procurement:The Wild Westof AI

Why Hydra Host

Build Your Quote

Meet with Gabriel Perez

Procurementas a Service

GPU Procurement:The Wild Westof AI

Why Hydra Host

Build Your Quote

Meet with Gabriel Perez

Procurement
as a Service