NVIDIA GeForce vs. AMD Radeon Vulkan Neural Network Performance With NCNN

[angrypie]

Looks like on the AMD side that the best compute bang for the power budget is the Radeon RX 5700. I wonder going forward with AMD's bifurcation of graphics cards (RDNA) and compute cards (CDNA) that we will be forced to choose between the two for different use cases and work loads as opposed to buying a more general all purpose card as we can now with a Vega based card ??

Yeah, I'd say so, this way they could argue "well no wonder our RX 6969 XT xXx has no SR-IOV, it's a gaming card, please sell a kidney and buy a Radeon PRO (didn't say it has to be your kidney, wink wink)."

[Imout0]

It's nice to see a program running on AMD GPUs for a change. And nicely on top of that, it's like an endangered species or something.

[nranger]

Great to see Vega doing exceptionally well in this workload. Too bad that most of the tech press doesn't care about these scientific / professional workloads that its bad reputation will stick to the end of its product life.

As a Vega 56 owner, I appreciate Michael including one in the benchmarks.

[TOMBOMBADIL]

Thank you for this test, Michael.
It's nice to see compute tests including AMD cards and even nicer to see them compete well.

[Danny3]

I would imagine that since HBM is so hot and expensive AMD will reserve HBM for their CDNA compute cards and keep scads of GDDR6 on their RDNA cards with the possible exception of some one off RDNA 2 Geforce killer. Plus GDDR6x is right around the corner.

HBM2 cannot be more expensive than 3 years ago when Radeon VII was launched.
Compute cards are very hard to resell as second hand, so not very attractive to buy in the first place if then you are stuck with it.
GDDR6 is still not quite there performance wise and I thing it's much more power hungry.

[nranger]

HBM2 cannot be more expensive than 3 years ago when Radeon VII was launched.
Compute cards are very hard to resell as second hand, so not very attractive to buy in the first place if then you are stuck with it.
GDDR6 is still not quite there performance wise and I thing it's much more power hungry.

Radeon VII launched February 2019, less than 2 years ago. And they're selling on ebay right now for $100-$200 OVER msrp. It's still one of the best options for crypto-miners.

HBM2 is likely less expensive now that it was for last generation, but it's going to be more money than GDDR6, and more expensive to package. It makes sense in a premium gaming product if it can capture high enough margins, and the power budget saved can go the GPU core. I wouldn't be surprised if AMD has an HBM2e equipped "3080 killer" in the $800-$1000 range.

[qarium]

this benchmark is really amazing. it looks like as soon as Nvidia bullshit technology like CUDA or OpenCL is gone and we do real technology Compute with vulkan AMD hardware shine and is faster than nvidia.

this means we just need to ignore CUDA/OpenCL and go full vulkan for Compute tasks and we will have a very good future perspective.

by the way: FUCK YOU NVIDIA

[Jumbotron]

HBM2 cannot be more expensive than 3 years ago when Radeon VII was launched.
Compute cards are very hard to resell as second hand, so not very attractive to buy in the first place if then you are stuck with it.
GDDR6 is still not quite there performance wise and I thing it's much more power hungry.

Sorry to burst your bubble. But HBM is STILL more expensive and power hungry than GDDR5 and even GDDR6. HBM is not only more expensive to produce it also adds a cost to the board manufacturers to implement thus increasing the overall BOM (billable order of materials).

It's also more expensive to implement on die or on package to something such as a GPU, not to mention an APU.

[Imout0]

But HBM is STILL more expensive and power hungry than GDDR5 and even GDDR6.

HBM(2) is far more energy efficient than anything GDDR.

BOM (billable order of materials).

Bill Of Materials.

[thelongdivider]

this benchmark is really amazing. it looks like as soon as Nvidia bullshit technology like CUDA or OpenCL is gone and we do real technology Compute with vulkan AMD hardware shine and is faster than nvidia.

this means we just need to ignore CUDA/OpenCL and go full vulkan for Compute tasks and we will have a very good future perspective.

by the way: FUCK YOU NVIDIA

Yeah, ignore the most performant option because it tells the story you like best. That seems like a reasonable scientific approach.