Ingrasys GB6181I vs Supermicro X14 Gaudi 3 AI Training Platform

Both are built on the NVIDIA Gaudi 3 HLB-325 Baseboard baseboard, so their accelerator performance is identical rather than merely similar. The choice between them is chassis, cooling, expansion and supplier, not FLOPS.

What actually differs

SpecificationIngrasys GB6181ISupermicro X14 Gaudi 3 AI Training Platform
VendorIngrasysSupermicro
Part numberGB6181I
Form factorrackrack
Released2024
Accelerator8x Gaudi 38x Gaudi 3
Aggregate memory1.0 TB
Aggregate bandwidth
System power
InterconnectEight Gaudi 3 accelerators on a universal baseboard, with twenty PCIe 5.0 x16 slots available for scale-out networkingEight Gaudi 3 accelerators all-to-all on an OAM 2.0 universal baseboard, with six on-board OSFP 800GbE ports for scale-out over standard Ethernet

Rows shown in grey are identical between the two.

Ingrasys GB6181I

Ingrasys' Gaudi node, and the second Intel accelerator system in this catalogue. Eight Gaudi 3 accelerators sit on a universal baseboard in a six OpenU liquid-cooled chassis, so like its AMD sibling this is an Open Compute part rather than a 19 inch one. What stands out against the rest of the field is the I/O: thirty-two NVMe drive bays and twenty PCIe 5.0 x16 slots, far more than the eight to twelve typical of an eight-accelerator machine, which suits Gaudi's design of scaling out over ordinary Ethernet rather than a proprietary fabric. Power is ten 3,000 W Platinum supplies in 5+5 redundancy, which is capacity rather than draw, so no system power is shown. Ingrasys publishes no throughput figure; per-accelerator compute is on the Gaudi 3 chip entry and the baseboard's figures on the linked Intel HLB-325 entry.

Vendor product page

Supermicro X14 Gaudi 3 AI Training Platform

Supermicro's X14 platform for Intel Gaudi 3, and by Supermicro's account the first AI server built on the accelerator. Eight Gaudi 3 cards sit on an OAM 2.0 universal baseboard driven by two Intel Xeon 6 processors, with 24 DIMM slots taking up to 6 TB of host DDR5 in one-DIMM-per-channel, up to eight hot-swap PCIe 5.0 NVMe drives, and four further PCIe 5.0 slots. Scale-out is six on-board OSFP 800GbE ports over ordinary Ethernet rather than a proprietary fabric, which is the Gaudi platform's central pitch alongside an open software stack with no licensing cost. Power is eight 3,000 W Titanium supplies in a 4+4 redundant arrangement, which is supply capacity rather than a draw figure, so no system power is recorded. Air-cooled builds use the HL-325L accelerator and liquid-cooled builds the HL-335, which is a part this catalogue does not yet carry. Supermicro publishes no throughput figure for the system, so none is shown; per-accelerator compute is on the Gaudi 3 chip entry. Supermicro describes the baseboard only as an OAM 2.0 or OAM universal baseboard and never names Intel's HLB-325, though that is what the description matches. Search results give the part number as SYS-822GA-NGR3 in an 8U chassis, but neither Supermicro page carrying that detail could be loaded, so it is reported here rather than recorded as a specification.

Vendor product page