I don't think AMD needs to support 5+ year old GPUs personally. And all the recent generations are already practically supported.
AMD only claims support for a select few GPUs, but in my testing I find all the GPUs work fine if the architecture is supported. I've tested rx6600, rx6700xt for example and even though they aren't officially supported, they work fine on ROCm.
You can paper launch as much as you want but Blackwell isn't shipping until Q4.
Also Blackwell's lead will be short lived, because mi350x is coming out next year, and it will have a node and architecture advantage. So AMD will be ahead again.
None of those things matter, if all you're looking to do is run your existing workloads on mi300x in the cloud. You get more bang per buck by going AMD.
ROCm has improved a lot. And you can rent mi300x in the cloud now. So if you have a program that runs on Nvidia GPUs, it takes no time to test it on a cloud mi300x. If it works you can use it and save some money in the process.
I don't understand why they should use TensorRT. vLLM is much more popular and it was actually written for Nvidia. It also supports AMD hardware, so it's the appropriate tool to compare.
People used the same argument when saying AMD would never beat Intel in CPUs. Intel has a lot of software engineers. Also these days AMD has a good number of software folks, thanks to the Xilinx acquisition and the organic investments in this area.
For us hobbyists used 3090 or new 7900xtx seem to be the way. But even then you still need to build a machine with 3 or 4 of these GPUs to get enough VRAM to play with big models.
zendnn is for CPUs. Pytorch ROCm version doesn't require you to change a single line of code. So all the same semantics like `device = "cuda"` work the same when using AMD GPUs.
The author also says that AMD doesn't have Tensor cores. No shit, Tensor core is an Nvidia trademark. AMD has matrix multiplication units in their CDNA GPUs.
AMD's driver is much better on Linux than Nvidia's. And ROCm HIP stuff isn't obscure. It's what's needed to bridge Nvidia's proprietary vendor lock ins.
There is also an article about how they did submit a score from one of these Chinese "exaflop" systems, for a different benchmark and it turns out it can only achieve the claimed performance at half precision:
> I guess AM4 also means no real improvements on the PCIe lane count
The new Ryzen 3000 CPUs support PCIe Gen4.. so while the number of lanes will remain the same, their bandwidth could be doubled. Just announced Navi GPUs also support Gen4.
It's only a space heater when it's being used. These CPUs consume very little power when idle. So if you have a task that can leverage 32 cores.. that task should in theory complete that much faster than on your 65W CPU. So the difference in total heat generated shouldn't be too bad.
AMD only claims support for a select few GPUs, but in my testing I find all the GPUs work fine if the architecture is supported. I've tested rx6600, rx6700xt for example and even though they aren't officially supported, they work fine on ROCm.