This week, we grabbed a few photos at AMD’s Advancing AI 2025 event. We thought they would be worth showing since AMD said it is already shipping the Instinct MI350 series to customers.
AMD Instinct MI350 Package
Here is a straight-on shot of the AMD Instinct MI350 package. You can see the eight HBM3E 12-high stacks around the central compute tiles and two I/O dies.
Here is the package from a different angle.
AMD had the pad side affixed to a stand, but here is what the pads look like on the other side.
The GPU package is then soldered onto its OAM board to make an OAM package that can be installed in servers.
That OAM package is installed into an OCP UBB.
AMD Instinct MI350 UBB
Here is the MI350 OAM package installed into a UBB alongside seven other GPUs for a total of eight.
Here is another angle of that.
Here is a look at the entire UBB with eight GPUs installed.
In a lot of ways, this is similar to the previous generation AMD Instinct MI325X board, and that is the point.
On one end, we have the UBB connectors and a heatsink for the PCIe retimers.
There is also a SMC for management.
Beyond the board itself, is also the cooling.
AMD Instinct MI350X Air Cooling
Here is an OAM module with a big air cooling heatsink. This air cooling is the AMD Instinct MI350X.
Here are eight of these on the UBB. This is similar to what we saw above, just with the eight big heatsins.
Here is another view of the heatsinks from the SMC and handle side.
For some reference, here is the AMD MI300X’s UBB:
AMD also has the MI355X liquid-cooled version of this which allows for higher TDP and higher performance per card.
Final Words
Is always neat to see new generations of AI chips. We still have some AMD Instinct MI325X content coming up, but this new generation we expect to ramp this year. With 2304GB of HMB3E on each UBB, this has a lot of memory.
Do you know what they are using for gpu to gpu communication for this system?
Also, what are they using for N/S communication?
Seeing the fully populated baseboard and all eight accelerators with heatsinks, I can’t imagine this thing weighs any less than 600 to 800 pounds when completely assembled, case and all. What a monster. I’d bet just ONE of the complete OAM cards has to weigh a good 15 or 20 pounds just due to all the copper and steel bolted to it.
@Stephen, search for:
GPU-AMD-MI300X-OAM-0045H “6 kg”
It says six kg for 8 OAMs, like what is shown in the last three images.
Do you know if there’s a vendor which allows for not fully populated configurations? I would like to buy a single card and keep adding more with time (for financial reasons)
Thanks, @Rob.
6 kilos is in the ballpark of my estimate of the weight of one OAM card, plus heatsink and support frames. That translates to just over 13 pounds per module. So yeah, just under the lower bound of my estimate. For the full set of eight OAMs, that’s 48 kilograms, or just shy of 106 pounds just for the GPU complement! OUCH! Not the two to three hundred pounds I initially expected but still very heavy! That means the whole server when fully assembled as shipped from the factory is probably more like 300 or 400 pounds. Still a lot of mass.