ARM announces cache-coherent GPU, "Mimir"

[NTMBK]

http://semiaccurate.com/2015/10/27/arm-cci-550-interconnect-mimir-gpu/

Should be useful for HSA.

 

[AtenRa]

HSA is dead.









/sarcasm

 

[Phynaz]

ARM announced an interconnect, not a GPU.

https://www.arm.com/about/newsroom/...on-for-next-generation-heterogeneous-socs.php

Charlie gets it wrong again.

 

[Exophase]

ARM announced an interconnect, not a GPU.

https://www.arm.com/about/newsroom/...on-for-next-generation-heterogeneous-socs.php

Charlie gets it wrong again.

Charlie's article covers all this. It just also includes a block diagram that uses a fully-coherent "Mimir" Mali GPU, hence that part of his announcement. He didn't get anything wrong there.

If there's something he got wrong it's when he says that the Mali-T8xx line is not-coherent. This is half-true; they are one-way coherent. Meaning that the GPU can snoop CPU-side data but the CPU can't snoop GPU-side data. This arguably provides the much bigger portion of the benefit, if not all that's necessary for high performance hUMA/HSA. AFAIK this is also the coherency configuration on PS4 and XB1 (although PS4 has GPU-side cache bypass options)

 

[NTMBK]

ARM announced an interconnect, not a GPU.

https://www.arm.com/about/newsroom/...on-for-next-generation-heterogeneous-socs.php

Charlie gets it wrong again.

So Charlie and Anandtech both imagined this slide?

The one with "Mali Mimir Fully Coherent GPU" on it?

 

[monstercameron]

Glad that Sammy, mediatek, imgtech, and even arm like the idea of hsa.

 

[Exophase]

Glad that Sammy, mediatek, imgtech, and even arm like the idea of hsa.

Both IMG and Qualcomm will need to support it in their GPUs for it to be worth anything on ARM's. They won't be able to drive even a small amount of ecosystem adoption (which is limited pretty much 100% to Android) without that much. And even if they did, it'd take years for developers to start feeling taking advantage of it. IMG looks on board but I don't know about Qualcomm.

It's kind of amazing that we've moved to inter-core, inter-cluster, and now CPU/GPU full data coherency but still lack coherency between L1 instruction and data caches on ARM CPUs. Even when it's more important now than it was years ago.

 

[Phynaz]

So Charlie and Anandtech both imagined this slide?

The one with "Mali Mimir Fully Coherent GPU" on it?

Derp 😳

But in my defense it's true they didn't mention the GPU in the press release.

 

[NTMBK]

Derp 😳

But in my defense it's true they didn't mention the GPU in the press release.

Yup, ARM could have been a lot clearer 😛 I hope they get Anandtech to do another deep dive, the Midgard one was very interesting!