SupermicroAS-2115HS-TNR-01-G2

A 96-core AMD EPYC 9655P, 1.5 TB of DDR5 and an NVIDIA BlueField-3 DPU with two 200 GbE or InfiniBand NDR200 ports, in a single-socket 2U Hyper with 24 bays.

Supermicro AS-2115HS-TNR-01-G2Rack 2U · Gold Series G2

The exact configuration

Official datasheet

Front view diagram: hover over each component.
Processor1× AMD EPYC 9655P96 cores · 2.6 GHz
Memory1.5 TB24× 64 GB DDR5-6400
Storage2× 7.6 TB U.2 NVMe
NetworkNVIDIA BlueField-3 B3220 DPU, 2-port 200 GbE / NDR200
Form factorRack 2U
Warranty
3 years parts and labour
On-site engineer
Optional: next business day, for 3 years
Software
Supermicro Data Center Management Suite, licence per node

In-memory databases serving inference

Vector databases, feature caches and the index a model queries before it answers are all in-memory databases serving inference, and two things matter for them: the whole dataset has to fit in RAM, and 1.5 TB goes a long way, and it has to reach the network without the CPU spending its time moving packets. That is the job of the BlueField-3, a DPU (a network card with its own Arm processors) that takes over encryption, remote storage access and network isolation, leaving the 96 cores for the database. Compared with a dual-socket server fitted with a conventional 200G network card, here that networking and encryption work never touches the CPU.

With a single processor, the 24 slots run at two modules per channel and the speed drops to 4,000 MT/s, so if your workload depends more on bandwidth than on capacity, the -02 gives you half the memory at 5,200 MT/s. The DPU, for its part, only pays off if your platform makes use of it or someone is going to program it; for a conventional 25 GbE network it is more than you need.

What we do

Most of the work here is on the network side. We install RHEL or Ubuntu with the BlueField-3 drivers and firmware, set it to Ethernet or InfiniBand mode depending on what sits at the other end of the cable and, if the models run on Kubernetes or OpenShift, connect the node as that cluster's data layer. We explain how it fits with the other pieces in AI inference on Ceph, OpenStack and Kubernetes.

Compared with the rest of the family

ModelWhat it hasWhen to choose it
AS-1126HS-TN-01-G22x AMD EPYC 9355 · 1.5 TBTwo high-frequency EPYC in just 1U.
AS-2115HS-TNR-02-G21x AMD EPYC 9655P · 768 GBThe same 96 cores with faster memory and 25 GbE.
AS-2126HS-TN-01-G22x AMD EPYC 9965 · 1.5 TB384 cores to consolidate, with little memory per core.
SYS-112H-TN-02-G21x Intel Xeon 6731P · 1 TBOne Xeon, 32 GB per core and 4×25 GbE.
SYS-222H-TN-01-G22x Intel Xeon 6740P · 1 TBTwo Xeon in 2U with spare slots and GPU space.
AS-2115GT-HNTR-01-G21x AMD EPYC 9534 · 512 GBFour 64-core servers in 2U, built for clustering.
AS-1115CS-TNR-01-G21x AMD EPYC 9555P · 384 GBMetadata node with four 200G ports.
AS-1115CS-TNR-02-G21x AMD EPYC 9115 · 64 GBEntry-level EPYC 9005 with room to grow inside.
SYS-122C-TN-01-G22x Intel Xeon 6736P · 1 TBTwo Xeon 6 in 1U, for many identical nodes.
AS-3015MR-H8TNR-01-G21x AMD EPYC 4565P · 64 GBEight small, fast nodes in 3U, one per customer.
AS-1015A-MT-01-G21x AMD EPYC 4564P · 32 GBRemote office in 1U and 43 cm deep.

See all rack servers

What the delivery includes

From the factory

  • Configuration assembled and tested by Supermicro
  • Warranty: 3 years parts and labour

From us

  • Rack mounting
  • Remote management (BMC) and up-to-date firmware
  • Operating system and the stack that runs on top
  • Testing with your workload before it goes into production

If you need it

  • 24×7 on-call support with SLA
  • On-site engineer next business day for the full 3 years

Questions about the AS-2115HS-TNR-01-G2

What is the difference from the AS-2115HS-TNR-02-G2?

Same chassis and same EPYC 9655P. The -01 doubles the memory to 1.5 TB, although it runs slower with two modules per channel, and swaps the 25 GbE networking for the 200 GbE BlueField-3. The -02 stays at 768 GB at 5,200 MT/s and is the sensible choice if you are not going to make use of the DPU.

Does the BlueField-3 run as Ethernet or InfiniBand?

Both ports work as 200 GbE or as InfiniBand NDR200 (the InfiniBand generation at 200 Gb/s per port), and the mode is set in configuration. We prepare it for the switch it will connect to, because that is the first thing to decide before ordering cables and transceivers.

Does it take GPUs?

The chassis accepts up to two double-width GPUs, such as the NVIDIA RTX 6000 Ada or the RTX A6000, depending on how the PCIe slots are shared with the DPU. The Gold configuration has none, so if you want the same node to run the model as well as serve the data, tell us when you order so we can check slots, power supplies and cooling.

Is the AS-2115HS-TNR-01-G2 right for what you're building?

Tell us about the workload and the data centre it's going into, and we'll come back with the price, a confirmed lead time and what it would take to get it running.