We Only Train on an RTX 5090 and 3090, So Is NVIDIA Still the Safe Buy Over AMD ROCm?

A first-person hardware account reports that an ASUS ROG Astral RTX 5090 and an RTX 3090 remain the only training cards in use, with the author arguing NVIDIA stays the default buy for training because the gap with AMD is software, not silicon. The account cites Compute Market's finding that AMD undercuts NVIDIA by 15-40% on price for inference, budget builds and datacenter serving, and bestgpuforai's example of an RX 9070 XT at about $550 carrying the same 16GB as a $1,400 RTX 5080, while noting the RTX 5090 lists at $1,999 but reportedly sells for roughly $4,300 to $5,000+. A Wan 2.2 speech-to-video render on the RTX 5090 peaked at 31.9 GB of VRAM on a 32GB card, and after a separate render exhausted VRAM and froze the desktop the author added a gate that defers renders unless 25 GB is free.

For training, buy NVIDIA unless you have a specific reason not to. Those reasons exist, and I’ll lay them out below. First, a disclosure: we haven’t trained on AMD hardware. Our cards are an ASUS ROG Astral RTX 5090 https://www.gladlabs.io/go/asus-rog-astral-nvidia-geforce-rtx and an RTX 3090. Everything below about AMD comes from published sources, and everything about NVIDIA comes from our own machine. That matters for how you read this. I’m not going to hand you a benchmark table I didn’t run. The gap is software, not silicon The most useful line I found came from bestgpuforai: the gap is software, not silicon. RDNA 4 ships with 128 AI accelerators and published FP8 and INT4 throughput. The hardware is real. But CUDA is what every framework, tool and tutorial assumes. Training magnifies that. Inference is one loop you can swap a backend under. Training is a stack: your framework, your custom kernels, your optimizer extensions, your distributed library, your logging. One CUDA-only extension anywhere in that chain and the cheap card becomes an expensive project. Compute Market puts the same thing in friendlier terms. CUDA’s maturity means day-one framework support, 4M+ developers and fewer driver headaches. https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026 It calls NVIDIA the lower-friction choice and AMD the value play if you can handle ROCm. “If you can handle ROCm” is the whole decision. It’s a staffing question, not a spec-sheet question. Memory is the training budget Forget TFLOPS for a minute. Training dies on VRAM. Here’s a concrete example from our box. We ran a Wan 2.2 speech-to-video render on the RTX 5090. It peaked at 31.9 GB. On a 32GB card, that’s a rounding error away from falling over. While the render held 28 GB, Ollama on the same GPU saw 3.9 GB free, and our integration tests crashed. That wasn’t a training run. It was a render. But the failure mode is identical to what you hit fine-tuning a model that almost fits. After a separate incident where a video render exhausted VRAM and froze the desktop, we added a gate. Renders now defer unless 25 GB is free. In its first production check it saw 8.1 GB free and correctly deferred. Boring, effective, and nothing to do with which vendor made the card. This is where AMD’s pitch gets interesting. Compute Market says the MI300X’s 192GB of memory has proven itself in large-scale datacenter serving. https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026 If your bottleneck is fitting the model, a card with that much memory changes what you can attempt. Then bandwidth. The RTX 5090 moves roughly 1,790 GB/s, per the spec figure on Compute Market and our own notes. https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026 Our RTX 3090 moves 936 GB/s. Same driver, same tooling, half the pipe. That gap shows up whenever you’re memory-bound. Bandwidth still decides a lot, and AMD’s marketing rarely leads with it. Price: AMD wins on paper, the street disagrees The case for AMD is money. Compute Market says AMD undercuts NVIDIA by 15-40% on price for inference, budget builds and datacenter serving. https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026 bestgpuforai makes it concrete: an RX 9070 XT at about $550 carries the same 16GB as a $1,400 RTX 5080. https://bestgpuforai.com/articles/nvidia-vs-amd-for-ai/ That’s a real gap. It holds only if nothing in your stack needs a CUDA-only extension. The flagship story is messier. The RTX 5090 lists at $1,999, and Compute Market shows a $1,999-$2,199 range. But One Day Advisor reports it selling for roughly $4,300 to $5,000+. https://www.onedayadvisor.com/2025/10/Nvidia-vs-AMD-GPUs-AI.html At that price, AMD’s 32GB Radeon AI PRO R9700 starts to look like a serious option rather than a compromise. Check current listings before you budget. Prices move week to week and the trackers disagree. Don’t mix vendors If you’re thinking about a Frankenstein build, one NVIDIA card plus one AMD card to pool VRAM, stop. We added an RTX 3090 beside the 5090. That gave us 24GB plus 32GB, 56GB in total. It works because both cards speak the same stack. Our notes are blunt about the alternative: Ollama runs each model load on a single compute backend, so NVIDIA and AMD cards can’t be pooled for one model. That’s an Ollama limit, not a law of physics. But the underlying logic travels. Two vendors means two driver stacks and two sets of kernels, and your tooling will pick one. Even same-vendor pooling bit us, just in a dumber place. Our GPU exporter parses output from NVIDIA’s System Management Interface tool. It assumed exactly one GPU’s worth of values. Add a second card and the unpacking broke. The metrics went dark right when we wanted to see them. Plan for that. Multi-GPU isn’t just hardware. It’s every script that ever assumed there was one card. Power is the other physical reality. Our machine idles around 40W and pulls about 480W under an Ollama load. Two big cards, a Ryzen 9 9950X3D, a PSU you can hear. Budget the wall socket, not just the card. Where our experience stops Our actual training work is small-scale. We’ve written up how we got GRPO running across separate jobs by syncing LoRA adapters instead of relying on NCCL, in Skip NCCL https://www.gladlabs.io/posts/skip-nccl-how-lora-adapter-syncing-makes-grpo-trai-f555bedc . That was about dodging a communication bottleneck, not about vendor choice. Large-scale pretraining on a rack of accelerators is a different sport. For that tier, one datacenter guide we’ve seen calls the H100 80GB the standard workhorse and the B200 and GB200 better suited to newer frontier training. An industry roundup lists AMD’s MI355X as the credible NVIDIA alternative. We haven’t run any of them. Treat those as pointers. If you want the consumer side ranked by use case, we put that together in Best NVIDIA GPUs for Local LLMs in 2026 https://www.gladlabs.io/posts/best-nvidia-gpus-for-local-llms-in-2026-ranked-by-dde9c4f9 . How to decide Here’s the rule I’d actually use. Pick NVIDIA if: - You’re one person or a small team and your time is the expensive part. - Your stack leans on community kernels, extensions or tutorials. - You want the least surprising path from “pip install” to a loss curve. Pick AMD if: - Your workload is memory-bound and you can price a big-memory card against NVIDIA’s equivalent. - You’ve already confirmed every dependency in your stack runs on ROCm. Confirmed, not assumed. - Someone on the team is willing to own the driver and kernel debugging. Either way: - Stay on one vendor per machine. - Fix your telemetry before you add the second GPU. - Set a free-VRAM floor and defer work that doesn’t clear it. The verdict For training, NVIDIA is still the safer bet. Every source I read lands there, and our own experience with one vendor’s stack backs it up: when something broke, it broke in our scripts, not in the framework. AMD is no longer a bad choice. The silicon is real, the memory is generous, and the price gap is not small. But you pay for it in compatibility work, and training is where that bill comes due. If you have the time to pay it, run the numbers. If you don’t, buy the card that everything already assumes you own.