This is a review of the current and future state of GPUs. It has my take on the future which speculative.Current Server:- AMD server flag GPU is the MI355X. It's a 8 vertically stacked chiplet design.- Nvidia server flag GPU is the B200. It's a two glued big die design.- AMD is beating Nvidia in a single rack in cost and performance.- Nvidia is beating AMD in multiple racks connected to each other (up to 72 vs up to 8).Future Server:- AMD next architecture is the MI400. It's a redesign that is expected to double the performance of the MI355X.- Nvidia next updates are R100 and F100. R100 is just an update but F100 will have 3D stacking, but only LPUs.Server take:- AMD MI400 and MI500 are most likely going to beat Nvidia in general computing.- Nvidia F100 might beat AMD in LLM. However, LLM algorithms are changing, and Nvidia face thermal problems for sure. This strategy could be a big mistake.Desktop take:- Intel vs AMD vs Nvidia on Steam is currently 72/8/20, more or less.- AMD is cheaper with GDDR6. They invested in bigger cache, to mask latency.- Nvidia is beating AMD in performance, path tracing, DLSS, and fake frames.- Nvidia may have given up on the best desktop GPU. If they're only going 3D stack LPUs it's not going to be useful for desktops.- AMD solved the GPU chiplet latency problem, but might not have 3D stack on desktop on the near future. It's expensive and they need drivers.- RTX 60 was delayed to early 2028. RDNA5/UDNA1 is on schedule for mid-2027 but they might also delay to 2028 if they switch to GDDR7.
What's the latest one I can afford if I want 16gb VRAM or 24gb for the lo lo
>>109101691You can probably get an AMD 9070XT at reasonable price right now. Nvidia forget it, they're all overpriced.
Is ROCm competitive yet?
>>109101660>This is a review of the [...] future state of GPUsMight you be a time traveler?
>>109101765I'm actually a faggot but don't tell the other anons.
>>109101755Apparently yes. It doesn't perform as good but works both in Windows and Linux. You'll have better luck on Linux though. The regular amdgpu + mesa drivers work. It used to be a mess but now works with most regular software. This is probably going to improve further with UDNA but for now your best bet is having RDNA4.
>>109101777>I'm actually a [...]Which type of time travel? Monolithic, vehicular, portable, or Internet-based?
>>109101691Intel B50
>>109101824You can get away with an Intel B50 but ROCm is more mature than oneAPI/SYCL and the 9070XT easily beats the B50 in performance. The advantages of B50 is less power consumption and virtualization with SR-IOV.However $350 vs $650. If you're on a tight budget, yeah sure, the B50. You're have to deal with some quirks in gaming.
>>109101660amd isn't offering much of a value proposition in terms of price. Are they really trying to matter, or just fool investors?
>>109101691>for the lo lowhat
>>109101691you need a blackwell 6000
>>109101955Is one really enough for uh... "for the lo lo"?
>>109101948Now it is, inadvertently, and only because GDDR6 is cheap while GDDR7 isn't.Ignoring RAM prices, node shrinking costs more than it's worth and that's what got us into expensive GPU problem. Nvidia knows this and is trying to workaround with features. AMD and Intel also know this and both have chiplets in GPUs, although differently.It will take a while to have a big GPU upgrade at reasonable prices. I'd say 2028, minimum.
