AMD CDNA Discussion Thread

trinibwoy said:
It doesn't say whether the host interface is dedicated or uses one of the 8 external links. The latter certainly seems to be the case from the picture. 2 links to the CPU in PCIe mode and 6 links to other GPUs.
Click to expand...

upload_2021-11-10_23-29-29.png

This is the one used by Frontier not the one in page 8. (only Description is in page 8). The CPU connection uses the external IF links.
One link goes to Slingshot 200Gbps
 
Last edited by a moderator:
ethernity said:
This is the one used by Frontier not the one in page 8. (only Description is in page 8). The CPU connection uses the external IF links.
One link goes to Slingshot 200Gbps
Click to expand...

Right, it seems the CPU and NIC connections ride the IF links. So GPU P2P can't use the full 800GB/s in practice.

Another interesting thing is that IF tops out at 25Gbps per pin over an interposer while NVLink is at 50GBps per pin over a PCB. Shouldn't IF be much, much faster on the interposer?
 
trinibwoy said:
Another interesting thing is that IF tops out at 25Gbps per pin over an interposer while NVLink is at 50GBps per pin over a PCB. Shouldn't IF be much, much faster on the interposer?
Click to expand...
IF is couple of types. IFOP and IFIS. IFIS is designed for long range. IFOP has higher BW but designed for low energy inter die. IFOP have more BW than IFIS.
Both are designed for low latency instead of BW because of CPU roots.
 
trinibwoy said:
Right, it seems the CPU and NIC connections ride the IF links. So GPU P2P can't use the full 800GB/s in practice.

Another interesting thing is that IF tops out at 25Gbps per pin over an interposer while NVLink is at 50GBps per pin over a PCB. Shouldn't IF be much, much faster on the interposer?
Click to expand...
NICs are behind MI200's extra PCIe links, not IF.
 
DavidGraham said:
Bandwidth is still actually 200GB/s directional. Power consumption is 560w on water cooling.
Click to expand...
Well, I didn't mention any names, but if you ask… As for your original posting about bandwidth:

For CDNA you choosed to use lower / directional value (without mentioning it) and negative connotation:
DavidGraham said:
the link between the two dies can only handle 200GB/s between each die
Click to expand...

For A100 you choosed to use higher / bi-directional value (again, without mentioning it) and positive connotation:
DavidGraham said:
As explained above, big part of this increase is the upgraded NVLink with 600GB/s.
Click to expand...

Hard to keep technological discussion, when like half of the posts have to be dedicated to deciphering of biased and misleading content, mixing apples and oranges etc.
 
no-X said:
For A100 you choosed to use higher / bi-directional value (again, without mentioning it) and positive connotation:
Click to expand...
This isn't positive or negative, it was in the context of explaining power budget. And it wasn't definitely in the context of comparing inter die link speeds, which has to be fast enough to handle communication between two GPUs.


no-X said:
Hard to keep technological discussion, when like half of the posts have to be dedicated to deciphering of biased and misleading content, mixing apples and oranges etc.
Click to expand...
Maybe if you stopped digging through years old posts and started discussing technical points based on their merits alone.
 
Bondrewd said:
Where's the NV exascale?

Even Intel won (effectively) 2 systems.
Click to expand...
there:
Atos and NVIDIA to Advance Climate and Healthcare Research With Exascale Computing

Atos will develop an exascale-class BullSequana X supercomputer with NVIDIA’s Arm-based Grace CPU, NVIDIA’s next-generation GPU, Atos BXI Exascale Interconnect and NVIDIA Quantum-2 InfiniBand networking platform.
Click to expand...
https://www.hpcwire.com/off-the-wir...-healthcare-research-with-exascale-computing/
 
You must log in or register to reply here.
Back
Top