In this article, we will compare the new product from Nvidia, the GeForce RTX 4090, with various professional cards from this manufacturer and try to answer the question, "Is it more beneficial to use a new graphics card in workflows, or is it still better to use ?" server graphics cards have a number of significant differences, determined by the purpose of use: Professional and gaming GPU cards Server graphics cards are used in ML development, rendering and modeling of complex objects, scientific research, cinematography, and more. Gaming graphics cards are designed for individual use. Scope of application. The professional card cooling system blows hot air out of the server or workstation. Their cooling turbine is designed for continuous operation. Gaming cards blow air over the cards. They must be installed in special cases with a good ventilation system. Gaming card fans are not designed for long-term operation and fail during long-term, continuous operation. Cooling. Professional GPUs allow you to do more computing with less power. This feature largely accounts for the high cost of server video cards. Performance and energy efficiency. Quality control in the manufacture of professional cards is stricter than in the creation of gaming cards. Production features. Professional cards are not equipped with connectors (HDMI, DVI) for video output - there is only DisplayPort. Connectors. Not all server GPUs can be used for gaming. Additional functionality. GeForce RTX 4090 Technology Overview The GeForce RTX 4090 GPU was released at the end of 2022 and represents a continuation of the line of desktop accelerators from , which aroused great interest among players around the world. NVIDIA The key features of the card are: As with the entire GeForce RTX 40 line, the new AD10x GPUs (AD102 in the 4090) are based on the Ada Lovelace architecture and use the 4N process (TSMC). Improved performance of ray tracing and machine compute operations on tensor cores. The 4N technological process allows them to increase energy efficiency by several percentage points. The size of the card (304 by 137 mm, 3 slots) makes it difficult to mount it in both desktop PCs and servers. The gaming cooling system often makes it impossible to mount the 4090 in GPU servers. Compared to the 3090, the AD102 has 70% more CUDA cores. NVIDIA DLSS 3 technology uses motion vector analysis and OFA algorithms. The low-latency NVIDIA Reflex platform enhances the gaming experience for professional gamers. 8th generation NVEnc encoder with AV1 encoding support. NVIDIA Broadcast app. NVIDIA Studio. Specifications for NVIDIA RTX A4000, NVIDIA RTX A5000, NVIDIA RTX 3090 and NVIDIA RTX 4090 Graphics Cards RTX A4000 RTX A5000 RTX 3090 RTX 4090 Archi­tec­ture Ampere Ampere Ampere Ada Lovelace Techno­logy pro­cess 8 nm 8 nm 8 nm 4N GPU GA104 GA102 GA102 AD102 Number of transis­tors (billion) 17,4 28,3 28,3 76,3 Clock frequency (GHz) 0,74 1.17 1,39 2,23 Boost Clock (GHz) 1.56 1.70 1.70 2,52 Memory frequency (MHz) 1,750 2,000 1,219 1,325 Memory Band­width (Gb/s) 448 768 936.2 1008 GPU memory(Gb) 16 Gb 24 24 24 Memory type GDDR6 GDDR6 GDDR6X GDDR6X Cache memory (Mb) 4 6 6 72 ECC memory yes yes no no CUDA cores 6 144 8192 10496 16384 Tensor cores 192 256 328 512 RT cores 48 64 82 128 Number of texture modules 192 256 328 512 Maximum power (W) 140 230 350 450 Compu­ting perfor­mance FP16 (half) (tera­flops) 19.2 27.8 35.6 82.6 Compu­ting perfor­mance FP32 (float) (tera­flops) 19.2 27.8 35.6 Up to 82,6 Compu­ting perfor­mance FP64 (double) 599 giga­flops 867.8 giga­flops 556 giga­flops 1.3 tera­flops Theore­tical maximum fill rate (giga­pixels/s) 149.8 162.7 189.8 444 Theore­tical texture fetch rate (giga­texels/s) 149.8 433.9 566 1290 Inter­face PCI-E 4.0 x16 PCI-E 4.0 x16 PCI-E 4.0 x16 PCI-E 4.0 x16 NVIDIA DLSS no no yes 3 Nvlink no 2-board low-profile confi­gura­tion (2-slot and 3-slot bridges) no no CUDA Support 8.6 8.6 8.6 8.9 VULKAN Support 1.3 1.3 1.2 1.3 DirectX 12 Ulti­mate 12 Ulti­mate 12 Ulti­mate 12 Ulti­mate Shader Model 6.6 6.6 6.7 6.7 OpenGL 4.6 4.6 4.6 4.6 OpenCL 3.0 3.0 3.0 3.0 Virtual GPU (vGPU) software support — NVIDIA Virtual PC (vPC) and Vir­tual Appli­cations (vApps), NVIDIA RTX vWS, NVIDIA Vir­tual Com­pute Ser­ver — — Cost (€) 1 500 3 000 1 600 from 1 900 The new architecture, memory bandwidth, number of tensor cores, DLSS 3 technology, and other characteristics of the GeForce RTX 4090 define a wide range of applications for the GPU. It’s not only for gaming but also for working with artificial intelligence and complex calculations. Testing the RTX 4090 and A5000 Description of the test environment AMD Ryzen 9 5900 X 12-Core Processor (3.80 GHz) 32 GB DDR4-3200 ECC DDR4 SDRAM 1600 MHz Microsoft Windows 10 Professional 64-bit V-Ray GPU CUDA test V-Ray GPU RTX Test Blender Benchmark In this test and in LuxMark, we will only compare the RTX A5000 and RTX 4090 cards, as they are the most relevant in the context of this article. LuxMark We measured relative GPU performance in rendering. The performance of the GeForce RTX 4090 in the tests looks impressive and almost doubles not only the results of the but also the professional GPUs. The V-Ray GPU RTX benchmark demonstrates ray-traced GPU performance. The RTX 4090 also outperforms the RTX 3090 by two times. RTX 3090 "Dogs vs. Cats" To compare GPU performance for neural networks, we used the Dogs vs. Cats dataset. The test analyzes the contents of a photo and distinguishes between a cat or a dog in the photo. All necessary initial data is . We ran this test on different GPUs and in different cloud services, and we got the following results: here Full cycle of training (min.) The full training cycle of the test neural network took from 31 to 60 minutes. The result for the GeForce RTX 4090 was 31 minutes and surpassed the performance of all the other GPUs. The most noticeable difference in the results was between the RTX 3090 and RTX 4090 cards: the new generation of GPUs from NVIDIA coped with the calculations almost twice as fast as the previous Nvidia generation. Tests have shown that the closest competitor to the 4090 card is the A5000. It remains to compare these cards in terms of their price-quality ratio. In all the tests, the new card from Nvidia showed a result that exceeded the RTX A5000 by about two times. At the same time, the cost of the RTX 4090 is much lower: $1,599 USD (minimum price) compared to $2,599 USD. It would seem that the choice is obvious, but there are nuances. The A5000 GPU consumes significantly less power and can be a cost-effective solution for tasks with a constant high load on the GPU over the long term. The RTX A5000 also supports NVLink technology, which is useful when training neural networks. A5000 GPUs have no restrictions on using NVENC/NVDEC for parallel video transcoding tasks. With the purchase of a specialized license, the A5000-class professional GPUs can be virtualized and available in the server as several smaller virtual GPUs. Another problem is on using drivers for its gaming cards in data centers and remotely outside the office. Nvidia's ban Although NVidia's promotional photo shows a lot of 4090s with large fans and a 3U format, in reality, this configuration is almost impossible to buy. In stock, there are only large-size gaming cards for 4 units with an increased height and fans blowing out up and down. Such cards cannot be used in servers and in most workstations. Conclusion The transition to the new Ada Lovelace architecture has significantly increased the performance of the GeForce RTX 4090. Improved tensor cores and RT cores significantly improve the quality and expand the capabilities of real-time ray tracing. The 24 GB memory capacity allows you to process large amounts of data. The GeForce RTX 4090 is primarily designed for gaming and is great for various types of computing solutions: AI, data analysis, and machine learning. The new architecture greatly outperforms the previous generation of NVIDIA GPUs. An important limitation in the professional use of this video card is its high power consumption and the inability to combine several cards using Nvlink. An alternative to purchasing a video card is to rent a GPU server. Our calculations show that monthly rentals for the GeForce RTX 4090 and RTX A5000 cards are comparable in price. Accordingly, if you need to perform professional tasks, renting a GeForce RTX 4090 card can be beneficial due to its high performance. or a server with a with professional-grade NVIDIA RTX 5500 / 5000 / A4000 cards. are also available . The GPU card is dedicated to the VM and cannot be used by other clients. GPU performance in virtual machines matches GPU performance in dedicated servers. Rent GPU servers with instant deployment custom configuration VPS with dedicated GPU cards

End-to-End Data Center Equipment Supply and Support in Europe

Reanimating Old Server Cases with HP Common Slot Power Supplies

Up to 22% discounts on instant servers!

Too Long; Didn't Read

Make resilience your competitive advantage

Which GPU is Better for Business, the RTX 4090 or a Server RTX A5000?

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Untitled Story

A Quick WireGuard VPN Setup Guide

3 Easy Ways to Improve The Performance Of Your Python Code

5 Performance Tips For .Net Developers

A Benchmark of Low Code Platforms for Test Automation

A brief overview of Automatic Machine Learning solutions (AutoML)

A Quick WireGuard VPN Setup Guide

3 Easy Ways to Improve The Performance Of Your Python Code

5 Performance Tips For .Net Developers

A Benchmark of Low Code Platforms for Test Automation

A brief overview of Automatic Machine Learning solutions (AutoML)

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps