AMD Says It Is Possible To Develop An NVIDIA RTX 4090 Competitor With RDNA 3 GPUs But They Decided Not To Due To Increased Cost & Power

AMD has said that it is possible for them to develop an RDNA 3 GPU that competes in the ultra-enthusiast market with NVIDIA's GeForce RTX 4090 graphics card but they decided not to go down that route due to increase power and cost concerns.

During an interview with ITMedia, AMD's EVP, Rick Bergman, and AMD SVP, David Wang, sat to discuss their goals with the RDNA 3 and CDNA 3 architectures. The most interesting question is asked right at the start of the beginning about why AMD didn't release an RDNA 3 GPU under its Radeon RX 7000 lineup that competes in the ultra high-end enthusiast segment such as NVIDIA's RTX 4090.

Rick replied that it is entirely possible for AMD to develop a specification based on the RDNA 3 GPU that competes with the NVIDIA GeForce RTX 4090 but that wasn't the strategy AMD was going for in its Radeon RX 7000 lineup. It sounds like AMD can definitely squeeze some more juice out of RDNA 3 if they wanted to but it is also stated that such a specification will result in a higher power and also higher costs which is something they weren't going for in the first place.

It makes sense but at the same time, NVIDIA's RTX 4090 was said to be a power-hog post-release but that didn't end up entirely true. Sure the card is rated anywhere between 450W (reference) and 500W+ for custom AIB models but the card itself is very efficient compared to AMD's high-end, the Radeon RX 7900 XTX as seen in various reviews, even our own.

Technically, it is possible to develop a GPU with specs that compete with theirs (NVIDIA) . However, the GPU developed in this way was introduced to the market as a graphics card with a TDP (thermal design power) of 600W and a reference price of $1,600 (about 219,000 yen)'', and was accepted by general PC gaming fans . After thinking about it, we chose not to adopt such a strategy.

The RDNA 3-based GPU "Radeon RX 7900XTX" released this time is targeted at $ 999 (about 136,000 yen), which is considered to be the "upper price" assumed by high-end users among general PC gaming fans. Developed . The "Radeon RX 7900XT" below it is said to be $ 699 (about 95,000 yen).

The price strategy is the same as the previous RDNA 2 (Radeon RX 6000 series), with the top-end "Radeon RX 6900XT" and "Radeon RX 6800XT" targeting $999 and $699, respectively. However, the target price changes for each GPU generation .

We take this strategy to fit into the mainstream infrastructure (hardware environment) utilized by today's PC gaming enthusiasts . At the same time as demanding high performance, it should be possible to operate with an existing common sense'' power supply unit, ``be able to cool the inside of the case with common sense'', and can be installed without requiring an extremely large case.'' ――The Radeon RX high-end product group was designed with these in mind.

AMD EVP, Rick Bergman (Machine Translated via ITMedia)

Continuing the questions regarding AMD's ultra-high-end enthusiast product positioning, it was stated that while Ryzen exists for gamers, there's also Ryzen Threadripper which exists for highest end of users and AMD has no cost/power concerns in that specific segment to which Bergman said:

Certainly, Ryzen Threadripper is crazy high performance (laughs). However, we don't think it was developed as a CPU for general users or gamers . The CPU socket is also unique, as evidenced by the fact that it doesn't fit into the mainstream Socket AM4 socket. However, I believe that the Ryzen series, which applies 3D V-Cache technology , which stacks cache memory dies , has become a receptacle for such ultra-high-end gamers.

AMD EVP, Rick Bergman (Machine Translated via ITMedia)

If I may say a few words, we at AMD are developing and releasing ultra-high-performance GPUs. For example, two years ago, we announced the " Instinct MI200 series " as the world's first multi-die GPU.

The top model of the series, " Instinct MI250X " marked the world's fastest theoretical performance of FP32 (single-precision floating-point arithmetic) at about 48 TFLOPS at the time. Since this is an Instinct series, it is not a GPU for gaming. However, if you look at the Instinct series, you should know that AMD can develop (ultra-high-end GPUs) if they want to.

We just don't think such GPUs are suitable for consumer use.

We haven't planned a "$1600 GPU" like the competition (NVIDIA) for PC gaming fans in recent years. Instead, we are focusing on planning a GPU that fits in the $1,000 class with a good balance between performance and cost.

There is a "difference" of about $600 from them (laughs), but I think that using that $600 to procure other parts, such as a CPU, would lead to a better gaming experience. I'm here.

AMD SVP, David Wang (Machine Translated via ITMedia)

Talking about future tech, David Wang said that while chiplet architectures need to be realized for future GPUs, they decided to postpone the realization of a true Multi-GPU die for the RDNA 3 gen of GPUs. AMD's RDNA 3 GPUs are still the first gaming architecture to feature a chiplet architecture but the actual chiplets are for the memory controller dies, also known as MCDs. The GPU die is still a singular monolithic package but that seems to change in the future.

Chiplet architecture can be used to realize high-performance processors while reducing manufacturing costs.

As you know, current high-end GPU cores contain more than 10,000 arithmetic cores (floating point arithmetic units). This is over 1000 times the number of CPU cores. If you try to interconnect (connect) the GPU dies in this state, the number of connection points will be enormous, and reliable electrical signal transmission cannot be guaranteed . So, at the moment, it is difficult not only in man-hours but also in terms of cost to connect the GPU die with the same glue as the CPU die. There's nothing you can't do, but... Rather than doing so, it is more efficient and less costly to create a large-scale GPU (core) at the moment .

Therefore, in the current generation (RDNA 3 architecture), we decided to postpone the realization of "multi-GPU die.

AMD SVP, David Wang (Machine Translated via ITMedia)

Moving forward, it looks like AMD has still room left to squeeze out of its RDNA 3 and it might be possible that a future enthusiast product launch might be imminent. As for RDNA 4, David Wang says that ray tracing is going to be focused a lot more and they want to continue improving the performance at the same pace as RDNA 3 over RDNA 2 which was around a 60% uplift. Updates that improve execution efficiency and delivering a flexible processing system are up in the charts for RDNA 4.