NVIDIA has released its 50-series cards, and AMD has released its 9000-series cards, and everyone is very excited to buy into this new generation of graphics hardware, but what do their specs actually mean?
Decoding the Specs of NVIDIA’s 50-Series Cards
If you open up the page for any of NVIDIA’s latest cards, you’ll encounter a few common terms, and may be wondering what exactly the words and numbers mean. So eat this GPU elephant one bite at a time.
GPU Architecture
First, we have Architecture, which in the case of the 50-series is Blackwell, named after mathematician David Blackwell. You don’t need to know all the specifics of chip architecture, but this is basically the fundamental design of the chip, and each generation gets its own code name. This is where a company like NVIDIA works to improve the GPU, make it more efficient, more powerful, and add more features. It’s important to know that you can’t directly compare the specifications of GPUs that use different architectures, for reasons I will explain below.
CUDA Cores
Next we have CUDA (Compute Unified Device Architecture) cores. These are the main processors of the GPU that do most of the work of rendering your graphics. These are also the most versatile processor cores on the GPU, though not as versatile as the cores in your computer’s main CPU. The more of these cores a GPU has, the better it will perform. However, you can’t compare the number of CUDA cores across generations, since newer CUDA cores may run at higher clock speeds, be more efficient at some jobs, or get more work done at the same clock speeds as previous generations. When you see a GPU’s performance rated in “teraflops” it’s the performance of these cores that are being referenced, though for various reasons, using teraflops as a way to compare cards doesn’t make much sense.
Tensor Cores
Tensor Cores are specialized processing cores that are built for tensor math calculations that are crucial for computing neural nets as used in AI (Artificial Intelligence) applications. For example, NVIDIA’s DLSS upscaling technology uses the tensor cores to accelerate its AI upscaling solution so that it’s fast enough to use in real time at hundreds of frames per second. It’s now common for AI-focused processors like these to have their performance measured in TOPS or Trillions of Operations Per Second, with the top-spec RTX 5090 offering a whopping 3352 AI TOPS, compared to the 40 AI TOPS you get in an AI-ready laptop.
Related
DLSS vs. XeSS vs. FSR: Which Upscaling Technology Should You Choose?
Ready for a crash course in all things upscaling?
RT Cores
RT or “Ray Tracing” cores are why NVIDIA’s last few generation of cards have been called “RTX” cards, and these specialized cores make it possible to use a realistic light simulation method known as “Ray Tracing” in real time. That’s as opposed to pre-rendering ray-traced graphics, as has been the norm for CG effects and movies in Hollywood for decades.
Related
Ray Tracing Isn’t a Gimmick, It’s a Way Bigger Deal Than You Think
The future’s bright for ray tracing.
The RTX-50 series offers 4th-generation RT cores, and each generation has been more capable and efficient, so the number of cores between generations isn’t that important. Likewise, the official specifications for 50-series cards may list a teraflop number for the RT cores, but that doesn’t tell you much, other than how powerful each card in the stack is relative to each other. For actual real-world performance, you’ll have to look up benchmarks for games with ray-tracing, such as Cyberpunk 2077.
Related
Why the Witcher 4 Trailer Being “Pre-rendered” on an RTX 5090 Means Nothing
Peak marketing tomfoolery.
Clock Speed
The clock speed of a GPU is a measure of how many cycles of operations it can run in a single second. This is measured in Hertz, and these days in Gigahertz. For example, the RTX 5090 has a base clock of 2.01GHz, and a boost clock of 2.41GHz. The base clock is the number the card is guaranteed to run at, and the boost clock is how high it will push itself assuming there’s enough power and cooling available.
Like other specs on these GPUs, the clock speed doesn’t tell you much, since total performance depends on so many factors. For example, a card with half the CUDA cores and twice the clock speed might perform roughly the same as its opposite. However, some third-party cards of the same model may offer higher clock speeds, and you can compare this as long as the cards use the exact same chip.
Memory Capacity, Speed, Bandwidth and Type
Discrete graphics cards have their own onboard memory, and will specify how much of it you get, how fast it is, how much bandwidth it has, and what generation of memory it is. Using the RTX 5090 as an example again, this card offers a whopping 32GB of memory with a 512-bit memory bus. The memory is GDDR7, and runs at 28Gbps. All of this culminates in a final memory bandwidth figure of 1.792TB/s.
That final number is what really matters, but cards that are aimed at lower resolutions (e.g. 1080p and 1440p) tend to have narrower memory buses. So a 1080p card might have a 128- or 192-bit bus, but a 1440p card might offer 256- to 384-bits and top-tier cards meant for 4K gaming with all the bells and whistles offer 512-bit memory buses. That doesn’t mean these cards can’t run higher resolutions; it’s just that they are usually most efficient within a certain resolution range.
Related
Make Sure Your Next GPU Has More Than 8GB of Memory
If you have one, that’s OK, but don’t fall for it again!
These are the core specs for modern NVIDIA cards you’re most likely to see, but things can be more granular than this when looking at the actual layout of each card’s GPU, but as someone looking to buy a card, that level of detail isn’t necessary.
As of this writing, there are four 50-series desktop cards for sale (if you can find them):
There are also laptops with GPUs that have the same names as some of these desktop GPUs, but it’s important to be aware that they are not the same, and will perform significantly worse than their desktop namesakes.
Understanding the AMD 9000-Series
That was a lot of ground to cover, and we haven’t even looked at AMD’s latest GPUs meant to compete with the 50-Series from NVIDIA. However, we don’t have to cover all of that ground again. AMD’s GPUs have specs that are largely comparable to NVIDIA’s cards; it’s just that the names may be different.
The architecture of the 9000-series cards is RDNA 4.0. Rather than CUDA cores, AMD GPUs have “stream processors”. The numbers can’t be compared between brands because their performance per processor is completely different. Instead of Tensor Cores, we get “AI accelerators”, and instead of RT cores we get “Ray Accelerators”. These do the same jobs, but with different levels of success and quality.
You can compare memory specifications directly between any two cards, since those fundamental details are the same on any GPU, but other than that it’s not that meaningful.
Unique Features of Each Team
GPUs are not just about the hardware, but also about the software features each brand offers. This can be a real deciding factor when two cards are otherwise similar in performance and/or price.
Upscalers are a big deal, and until the advent of the RX 9070 and 9070 XT cards, AMD has been far behind the curve. NVIDIA’s DLSS, now on DLSS 4, offers amazing upscaling quality, which lets a GPU render a game at a lower resolution and then increase the detail using AI for little to know visual loss and higher frame rates. Thousands of games support DLSS, and the latest DLSS 4 Transformer Model brings a huge jump in quality and will work on every RTX card, not just the 50-series.
The 9000-series of AMD RX cards brings FSR 4, which is AMD’s first AI-based upscaler comparable to DLSS. Only the 9070 and 9070 XT (and future) cards support this technology, and a very small number of titles work with FSR 4 right now. However reviews for FSR 4 technology are largely positive, and I’m sure more games will be added to that list in the near future.
NVIDIA’s ray-tracing performance remains much better than AMD’s, but AMD has made huge strides and this is now a feature you can actually turn on with the 9000-series cards without quite such a big hit to performance.
Both cards offer some form of frame-generation, where the GPU creates additional frames that don’t come from the game to improve visual smoothness, but onlt the 50-series cards offer multi-frame generation, where the smoothness can be pushed up to hundreds of frames per second.
How These Cards Stack Up
The last piece of the puzzle here are how these cards stack up relative to one another.
At the top of the hill is the RTX 5090, which is the fastest GPU in existence as of this writing. You’ll have to cough up an enormous amount of money to own one, however, and that’s if you can find stock.
The RTX 5080 is around 60% less performance depending on the game, but it’s still a beast of a card in its own right, and should be more than enough for most 4K gamers in modern titles, or those who want to absolutely push the frame rate at 1440p.
The RTX 5070 Ti and RX 9070 XT are very comparable cards in terms of raw performance, though the 5070 Ti is significantly better at ray tracing. The AMD card is quite a lot cheaper and more plentiful however, so when it comes to price versus performance, it’s the clear winner. These are primo 1440p cards, but they’ll run 4K games at reasonable quality settings, or great settings if you use their respective upscaling solutions.
The RTX 5070 and RX 9070 are also very comparable, but neither card makes a lot of sense at their current prices, and the 5070 in particular only comes with 12GB of VRAM, which means it will be a relatively short-lived card.
There are currently no entry-level cards from either company (such as an RTX 5060 or RX 9060), but I expect they’ll come along sooner rather than later, and probably outsell all the other cards on this list combined.
