Post
199
Something concerning is happening with Nvidia's GPU pricing and benchmark transparency.
Look at how other hardware generations work: a new iPhone costs roughly what the previous one did. Same with Samsung Galaxy phones, same with AMD Ryzen CPUs. Price increases are modest and the specs are straightforward to compare. Nvidia has broken from this pattern. Each new generation costs meaningfully more than the last, while the benchmarks get harder to interpret, not easier.
Look at what's happening with the numbers. BF16 and FP16 performance figures, the ones most relevant for real workloads, are increasingly buried or absent. Sparsity is baked into headline numbers without always being clearly flagged. The B200 is being benchmarked as an NVLink pair rather than a single card in some comparisons. Each of these is defensible in isolation, but together they make it genuinely hard to do apples-to-apples comparisons across generations.
The inference vs training tension is worth flagging too. Lower precision formats are good for inference but introduce real tradeoffs for training. As Nvidia pushes further into inference optimization, NVFP4 being the latest example, it raises a fair question about where training workloads go. NVFP4 training exists, but documentation and real-world results are thin enough that most teams aren't seriously evaluating it yet.
The "ROCm is immature" narrative is also getting a bit stale. I've personally run workloads on the MI300X and honestly it felt smoother than I expected, even compared to CUDA in some cases. And the price to performance ratio isn't even close. AMD isn't a consolation prize anymore.
At some point, labs and enterprises doing the actual buying will have enough alternatives like AMD MI300X, custom silicon, and inference-optimized ASICs to push back. Nvidia is standing on thin ice, and a few more steps like this might be enough to crack it.
Look at how other hardware generations work: a new iPhone costs roughly what the previous one did. Same with Samsung Galaxy phones, same with AMD Ryzen CPUs. Price increases are modest and the specs are straightforward to compare. Nvidia has broken from this pattern. Each new generation costs meaningfully more than the last, while the benchmarks get harder to interpret, not easier.
Look at what's happening with the numbers. BF16 and FP16 performance figures, the ones most relevant for real workloads, are increasingly buried or absent. Sparsity is baked into headline numbers without always being clearly flagged. The B200 is being benchmarked as an NVLink pair rather than a single card in some comparisons. Each of these is defensible in isolation, but together they make it genuinely hard to do apples-to-apples comparisons across generations.
The inference vs training tension is worth flagging too. Lower precision formats are good for inference but introduce real tradeoffs for training. As Nvidia pushes further into inference optimization, NVFP4 being the latest example, it raises a fair question about where training workloads go. NVFP4 training exists, but documentation and real-world results are thin enough that most teams aren't seriously evaluating it yet.
The "ROCm is immature" narrative is also getting a bit stale. I've personally run workloads on the MI300X and honestly it felt smoother than I expected, even compared to CUDA in some cases. And the price to performance ratio isn't even close. AMD isn't a consolation prize anymore.
At some point, labs and enterprises doing the actual buying will have enough alternatives like AMD MI300X, custom silicon, and inference-optimized ASICs to push back. Nvidia is standing on thin ice, and a few more steps like this might be enough to crack it.