It's not a video card,
but how many frames per second playing Doom?
NEC will soon sell as a standalone item the PCIe accelerator cards it uses to power modest supercomputers. The device is built around the tech giant's custom Vector Engine Processor, and NEC suggests it as a fine way to handle tasks such as weather forecasting, climate modelling, and fluid analysis. NEC offers the chipset in …
RockPro64 has a 4? lane PCIE. Similar money. Just upconvert to a 16 lane with an adapter. A roll of duct tape. The NEC card can't burn more than 500 or so watts, so you'll need an interesting power arrangement. Does someone make an active pcie backplane with an ATX power plug? Does this beast have a seperate power inpuf?
AMD's new MI100 Instinct accelerator card offers 11.5 Teraflops of 64-bit double-precision floating-point oomph, and is probably somewhat less expensive than the NEC card with only 2.45 Teraflops, but still in the same ballpark.
But the NEC card does have an important advantage. Accelerator cards are designed like GPUs, and so they can only perform a fairly restricted type of calculations.
The NEC card, on the other hand, is architecturally similar to the Cray I computer and similar vector supercomputers. As a result, a larger fraction of a typical calculation can take advantage of those 2.45 Teraflops of vector power. As well, it isn't an accelerator card - it's a self-contained computer, and so the scalar parts of the calculation are also performed within the card, so delays in moving data in or out of the card do not take place during a calculation.
That's the dumbest thing ever,
Especially since you can get Intel Xeon Phi cards for like $200-300 online nowadays, and that's like nearly a decade old technology!
It also makes no sense knowing one can just buy mini Threadripper or Epyc server blades, or systems.
Not to mention, GPUs like Nvidia RTX 3090 is running +20x faster, for an 8x lower price.
Just like GPU are good at some things and not at others, vectors are better at some things, including those that GPU are not good at. Horses for courses and it's certainly nothing like a Phi.
HPC and certain workloads are not only about TFLOPS and they certainly don't burn 500W, but that would only require a cursory glance (via your favourite search engine) to see that:
https://www.nec.com/en/global/solutions/hpc/sx/vector_engine.html?