MIAOW: An Open-Source GPU Design Based On AMD's Southern Islands

[phoronix]

Phoronix: MIAOW: An Open-Source GPU Design Based On AMD's Southern Islands

A new open-source GPU design has been published designed to run on FPGAs... What makes this "open hardware" project more interesting than past designs is that their compute unit was designed around AMD's public "Southern Islands" instruction set architecture...

http://www.phoronix.com/vr.php?view=MTgyNTE

[profoundWHALE]

This is pretty awesome. I hope it'll actually succeed.

[rxonda]

What changed?

This is wonderful news.
But what did change to make this project a hit?
Because he is using AMD instruction set?
They will use HyperZ and HSA API too?
Well, I do hope he will suceed.

[gigaplex]

This is wonderful news.
But what did change to make this project a hit?
Because he is using AMD instruction set?
They will use HyperZ and HSA API too?
Well, I do hope he will suceed.

How do you propose HSA would work with a discrete card?

[zanny]

How do you propose HSA would work with a discrete card?

Same way NUMA works. You treat the discrete GPU as a physically attached coprocessor but use scheduling priority to allocate work where its closest to its physically stored memory.

It would require CPU + GPU awareness and special instructions between them to do it, though.

[gigaplex]

Same way NUMA works. You treat the discrete GPU as a physically attached coprocessor but use scheduling priority to allocate work where its closest to its physically stored memory.

It would require CPU + GPU awareness and special instructions between them to do it, though.

The benefit of HSA was that there was zero copy between CPU and GPU, you just use virtual address mapping. The discrete card can't access the system memory, it needs to be copied into GPU memory, which means you're no longer using HSA.

[glisse]

The benefit of HSA was that there was zero copy between CPU and GPU, you just use virtual address mapping. The discrete card can't access the system memory, it needs to be copied into GPU memory, which means you're no longer using HSA.

Discrete GPU can access all system memory with no trouble. No need for special instruction or anything special, it is all part of existing PCI/PCIE specifications and have been for years, decades ...

[V10lator]

Discrete GPU can access all system memory with no trouble. No need for special instruction or anything special, it is all part of existing PCI/PCIE specifications and have been for years, decades ...

You're talking about DMA? Yes, that's true but the PCIe bus is way slower than the memory bus, so to not slow the GPU down a lot you have to copy the data into VRAM.

[glisse]

You're talking about DMA? Yes, that's true but the PCIe bus is way slower than the memory bus, so to not slow the GPU down a lot you have to copy the data into VRAM.

Nothing forbid you to use system memory directly without copy and in this case DMA is not the right word to use, of course it's all a matter of how you define DMA but in my book its definition is not appropriate for direct system memory access. But yes VRAM most of the times have more bandwidth the PCIe links. Nonetheless you can do HSA with discret GPU and AMD discret GPU are doing that, they just do not use VRAM.

[rxonda]

The benefit of HSA was that there was zero copy between CPU and GPU, you just use virtual address mapping. The discrete card can't access the system memory, it needs to be copied into GPU memory, which means you're no longer using HSA.

I'm no HSA expert, but I believe the real benefit of is a single API to DMA and fence data between resources (whether is GPU or CPU).