As a rule, the most expensive component of a gaming computer is the video card. The quality of digital content imagery is steadily increasing, and a performance GPU capable of outputting detailed and realistic images on the screen is becoming increasingly important.
This creates a race among manufacturers competing to create the most performance video card. Today we will look at the best of the best.
The very best video cards for gaming
Modern games are huge worlds with a high degree of detail. To render them dynamically, performance GPUs are needed. For the user, the question of whether their computer will handle a new game often comes down to the video card.
As of today, NVIDIA confidently holds the lead in the field of gaming graphics cards.
NVIDIA GeForce RTX 4090
A kind of star among all NVIDIA models, the most performance video card in all respects:
- The most advanced generation of GeForce 40 graphics processors on the Ada Lovelace architecture.
- The largest number of transistors - 76.3 billion.
- The largest number of scalar processors (CUDA cores - 16384), tensor cores (512).
- The highest performance - 88.2 TFLOPS.
The card uses the latest GDDR6X video memory standard with a volume of 24 GB, a PCI Express 4.0x16 bus, supports the latest version of Direct3D 12_2, OpenGL 4.6, and other latest APIs.
The abbreviation RTX stands for Ray Tracing Texel eXtreme: support for ray tracing technology, which provides a new level of light, shadows, reflections, and ultimately, realism of the game.
You have to pay for all this with high energy consumption - 450 W, so a computer with this video card needs a performance power supply. It is also not cheap in terms of money: the manufacturer's recommended price is $1600, often more expensive in retail.
NVIDIA GeForce RTX 4080
This video card is one step below the flagship, but with the same architecture, it already significantly loses to it. In particular, it has 45.9 billion transistors, 9728 CUDA cores, and 304 tensor cores. Performance, accordingly, drops to 48.7 TFLOPS.
Comparing in relative numbers, the older brother surpasses RTX 4080 in different tests by 1.6—1.8 times. At the same time, the difference in price is not so significant. GeForce RTX 4080 costs $1200, so the more sophisticated RTX 4090 is only a third more expensive. If you calculate in dollars per unit of performance, the older model is a more profitable purchase.
Nevertheless, GeForce RTX 4080 is still a very cool video card that will confidently handle any modern game. And if the difference of $400 is critical for you, or if RTX 4090 is not for sale and is not expected, GeForce RTX 4080 will be an excellent choice for a gamer PC.
AMD Radeon RX 7900 XTX
Speaking of AMD - NVIDIA's main competitor in the graphics card market, their flagship today is Radeon RX 7900 XTX. It must be honestly said that so far it is inferior to the top models of the opponent:
- The video card is made on the TSMC N5 technology process, while for GeForce RTX 4090 it is already TSMC N4.
- The number of transistors in AMD's leader is 57.7 billion.
- The used GDDR6 memory also lags behind the highest standard, although in volume it is similar 24 GB
- The video card has received 96 new ray tracing accelerators, which have improved performance by 80% compared to the previous generation, but in this indicator, the card is still inferior to NVIDIA.
Losing in performance, Radeon RX 7900 XTX can quite make up for it in price. The recommended $1000 is noticeably lower than the cost of GeForce RTX 4090 and slightly cheaper than RTX 4080. However, since in terms of performance it is inferior to both, it makes sense to buy this card only for convinced AMD fans and those for whom NVIDIA solutions are not available for some reason.
Top video cards for professional design
Creating video games, digital content, designing machines and buildings places its own demands on GPUs. And since this activity brings in money, developers can afford models much more expensive than those for whom the video card only opens access to cool games.
AMD Radeon Pro W7900
Introduced in April this year, this graphics card is AMD's most performance solution for workstations. In terms of the number of transistors, it's very similar to the Radeon RX 7900 XTX: the same 57.7 billion transistors, 96 compute units, and 6144 stream processors. The Radeon Pro 7900 also supports hardware ray tracing.
The difference in memory is significant: the Radeon Pro W7900 has 48 GB, twice as much as the gaming variant, and this memory includes error correction (ECC support). Another advantage of this professional AMD graphics card is its lower power consumption: 295 watts versus 355 watts.
However, for the double memory capacity and other advantages, the AMD Radeon Pro W7900 comes at a substantial cost: the price of the card starts at 4,000 dollars. There's a slightly less performance model, the Radeon Pro W7800, with 32 GB of memory, priced significantly lower at 2,500 dollars.
NVIDIA Quadro RTX 6000 Ada Generation
In NVIDIA's lineup, professional graphics design cards are traditionally released under the Quadro brand. The most performance solution from NVIDIA today is the RTX 6000 Ada Generation.
This graphics card is built on the Ada Lovelace architecture processors, with 18176 CUDA cores, 142 third-generation ray tracing cores, and 568 fourth-generation tensor cores. Its total computational power is immense: in single-precision computations, the card delivers 91.1 TFLOPS, and in tensor calculations — 1457 TFLOPS.
The graphics card has 48 GB of memory and supports up to 4 monitors with a resolution of 4096x2160 or 2 monitors with 7680x4320 resolution.
Despite its gigantic computational power, the RTX 6000 Ada's power consumption remains moderate at 300 watts. However, its price is substantial — 6,800 dollars, affordable only for professionals tackling serious tasks.
NVIDIA Quadro RTX A6000
The previous version of NVIDIA's workstation graphics cards, the RTX A6000, is built on the Ampere architecture. It has the same 48 GB of memory but significantly fewer CUDA cores — 10752. It also has fewer tensor cores — 336, and importantly, they are only third generation. This results in a significant performance gap compared to the RTX 6000 Ada Generation.
With 38.7 TFLOPS in single-precision computations and 309.7 TFLOPS in tensor calculations, the difference with the leader is not in percentages, but in multiples. Meanwhile, the price of the older graphics card remains substantial at 3,900 dollars, and is obviously no longer justified if you can buy the NVIDIA RTX 6000 Ada Generation.
Graphics Cards for Science, Data Centers, AI
The most performance graphics cards are not for gaming or design, but those intended for truly big tasks: scientific research, data processing in data centers, and training of graphic neural networks. Solutions here are often specific, optimized for particular calculations, and their prices are several times higher than those of regular GPUs.
NVIDIA Quadro GV100
This graphics card was released in 2018 and used the most advanced architecture of that time, Volta. At the time, it was a very impressive graphics card: 5120 CUDA cores, 640 tensor cores, and an impressive 32 GB of ECC-support memory. For comparison, the top gaming cards of the GeForce 20 series at that time only had 6–8 GB of memory.
The Quadro GV100 used NVLink technology, allowing multiple graphics cards to be connected for data center operations. At launch, the official price of the GV100 was 9,000 dollars, with market prices reaching up to 12,000 dollars.
AMD Instinct MI100
Advanced Micro Devices presented its first specialized graphics card for data centers in 2020. Unlike gaming models on the RDNA architecture, the MI100 card used the CDNA architecture, more suitable for processing large volumes of scientific data and AI training.
This card has 7680 cores, manufactured using 7 nm technology, and 32 GB of memory. In terms of performance at that time, it was already inferior to the GV100, and NVIDIA was already preparing to release the A100 series. Nevertheless, the price of 7,200 dollars left AMD with chances to find its customers.
NVIDIA Tesla A100
Impressive a few years ago, the Instinct MI100 and Quadro GV100 are now far from leaders. Today, NVIDIA's flagship in the science and AI segment is the A100 series.
This graphics card has a whopping 80 GB of HBM2e standard memory, with memory bandwidth of 1900–2040 GB/s. In double-precision floating-point computations, its performance is 9.7 TFLOPS, and in the case of FP16 tensor cores, it reaches 312 TFLOPS.
Designed for data centers engaged in artificial intelligence research, the NVIDIA A100 should be used as part of certified NVIDIA servers, housing up to 8 such GPUs. It's not available on the regular market (and wouldn't be needed there), and the price of one graphics card exceeds 17,000 dollars.
NVIDIA Hopper H100
If the Ampere architecture was used by developers both in regular (GeForce 30 series) and server (Tesla A100) graphics cards, they then decided to separate architectures for different applications. For regular GPUs, it's Ada Lovelace in the GeForce 40 series, and for servers, it's Hopper in the H100 cards.
This graphics accelerator has 80 billion transistors, 80 GB of memory, a bandwidth of 2 Tb/s, and uses the PCIe 5.0 or SMX5 standard. Its performance compared to the NVIDIA A100 is three times higher. The price also jumped significantly: over 40,000 dollars.
Due to the specificity of AI training tasks, in gaming tests, the Hopper H100 shows results at the level of fairly mediocre graphics cards. Thus, it offers nothing to the average user — it is a tool for specific applications.
What to Look for When Choosing a GPU
As we can see, big money doesn't always mean top results in games. So what should you focus on when choosing a GPU?
If you're building a new computer, it's better to choose the most performance among gaming graphics cards — currently, this is the GeForce RTX 4090. It will be more than enough for the most demanding modern games and those that will be released in the foreseeable future. New expenses will not be required any time soon.
If you are looking for a graphics card to upgrade your existing computer, then you need to consider compatibility with your hardware. You won't achieve maximum performance, but the quality of the gaming picture will be quite decent.
The staff at HYPERPC are ready to professionally consult you on all issues related to choosing components and compiling an optimal assembly for your tasks.
Arabic 
English