The app for independent voices

At first I wasn't sure what I was looking at.

A liquid-cooled GPU tray, but this one has 8 GPUs and I don't see any Grace CPUs?

Fortunately for me, the team from PEGATRON SVR were there to help!

---

All of the B300 + x86 air cooled servers I've seen are huge 8 or 10U tanks.

After all, the cooling tower on a 1400W GPU has to be a skyscraper relative to anything else in the box.

When I said in my article (lnkd.in/ezibVS78) that air-cooled GPU server chassis size is now dictated by cooling tower height, I was a little nervous that I'd be wrong and called out - But I guess I'm right!

---

Part of a 2U HGX server (Codenamed "AS208-2A1"), the tray shown here has all 8 B300 GPUs with NVLink (switches likely underneath?).

Apparently the custom baseboard on the left that holds the 8 x Connect-X8 ASICs and OSFP ports (not attached here) was co-designed with NVIDIA.

No PCIe form-factor NICs!

Also on the board at the very left are board-board PCIe connectors, I believe it would be 128 lanes of gen. 5 in total for the PCIe 6.0 switch in the CX8's to connect to the CPUs.

---

All of this comes in a rack-scale solution that houses 16 of these 2U servers along with 3 x SN2201 MGMT switches and 4 x power shelves.

That's a total of 128 B300 GPUs in one rack! From my rough estimate, that's a 220kW+ rack!

It's important to note that this doesn't directly compete with the NVL72/144s though, since those are a single NVLink domain (all GPUs P2P interconnected) and this is 16 separate NVL8/16 domains.

---

Anyway, enough talk, here are the specs:

Server (AS208-2A1)

- 2 x AMD EPYC 9005 (Turin) CPUs

- 16 (8) x NVIDIA Blackwell "Ultra" B300 GPUs

- 8 x Connect-X8 ASICs (custom baseboard, not NICs)

- 8 x 800G OSFP ports (to be used as 2 x 400G each for a dual-plane topology)

- 4 TB (realistically) DDR5 @ 6400 MT/s (32 slots)

Note 1: the core count of the CPUs is unclear since there is no public information on this on Pegatron's website, but I would assume given the wide range of core counts available with Turin that it would be variable based on the customer.

Note 2: Each of the CPUs only has 12 physical memory channels, so a total of 24 on this dual-socket CPU board. However the server here is advertised as containing 32 DIMM slots. I think this means that 16 of the channels are used as 1DPC and 8 as 2DPC to connect to the 32 physical slots. I wonder what this means for memory bandwidth and allocation, if anything.

---

Thanks to NexGen Cloud for taking me to GTCParis2025, a great event with great company!

---

Image sources:

Me

Links:

Pegatron’s news article on the AS208-2A1: linkedin.com/company/pe…

Not much more detail on this yet sadly.

Sep 9
at
10:35 AM
Relevant people

Log in or sign up

Join the most interesting and insightful discussions.