AI-powered businesses are now stocking up on gaming GPUs

AMD (unlike Qualcomm) is a joy to buy chips from. 7.38 PFLOPS of compute was delivered today, with exaflops more to come. Nice to see them more active on the ROCm GitHub, we are working on the layer above @tinygrad

Eagle-eyed readers will be able to make out the XFX Merc 310 GPU from an AMD Radeon RX 7900 XTX box, with the whole batch supposedly delivering 7.38 petaflops of processing power. If we base this on the FP16 performance of the RX 7900 XTX, which yields 123 teraflops, we can deduce that this single purchase contains 60 GPUs, which means Hotz had to part with nearly $60,000.

Hotz has no intention of stopping there, as he's revealed his plans to acquire additional exaflops of processing power. To put this into context, an exaflop consists of an incredible 1,000 petaflops. So, even if we consider a conservative estimate, let's say Hotz acquires a minimum of 1.045 exaflops of processing power, that translates to no less than 8,500 AMD GPUs funneling to a single company.

Curious gamers might wonder why Hotz would opt to invest a considerable sum in the RX 7900 XTX, when he could spend a fraction of that on Nvidia's GeForce RTX 4060, which delivers a comparable level of performance.

The simple answer is that AI tasks heavily rely on memory capacity. The AMD RX 7900 XTX boasts an impressive 24GB GDDR6 VRAM, while Nvidia's RTX 4060 offers just 8GB GDDR6 VRAM (which is even less than the RTX 3060's 12GB VRAM).

What implications does this have for the gaming landscape? Well, there's both good news and bad news. On the bright side, lower memory cards will still be more accessible, as AI-powered businesses will predominantly choose higher memory options. The drawback is that the surge in demand for higher memory cards will likely lead to increased prices.

Although some reviewers downplay the importance of plentiful VRAM, gamers who've played recent games like the Resident Evil remakes know the harsh reality of dealing with frame rate drops and low-resolution textures on higher settings.

Amidst the backdrop of US sanctions targeting China, the demand for GPUs has surged to new heights. This surge has significantly impacted consumer GPUs used for AI purposes, and there are growing concerns about the possibility of even stricter sanctions in the future.

To navigate this, Nvidia has taken the approach of de-optimizing some of its AI oriented GPUs, like the Ampere A100, now rebranded as the A800, with a reduced transfer rate from 600GB/s to 400GB/s.

As a result, AMD has decided to join the fray and develop GPUs that comply with the sanctions for the Chinese market. Dr. Lisa Su, AMD's CEO, shared this strategic move during the company's latest earnings call, saying: China is a very important market for us, certainly across our portfolio. As we think about certainly the accelerator market, our plan is to, of course, be fully compliant with U.S. export controls. But we do believe there's an opportunity to develop products for our customer set in China, that is looking for AI solutions, and we'll continue to work in that direction. Some NVIDIA AI GPUs are selling as high as $70,000 in China.

Our verdict is that if you happen to have some extra funds and you're stubbornly willing to stick to PC gaming, it might be wise to invest in a GPU with at least 16GB VRAM before the AI bros buy them out of stock.