Intel announced in Germany its future high-end GPU for HPC and AI-oriented data centers, we are talking about Rialto Bridgea chip that can unite up to 160 Xe cores32 more than the Ponte Vecchio, i.e. the number of cores increases by 25 percent.
The big performance boost will come on the bandwidth side, and is that although it can retain the memory interface of 8192 bitthe company will employ a memory HBM3 over the current 128GB HBM2e to add a drastic performance boost.
Intel also claims that its new GPU will have improved dies thanks to a new, more advanced manufacturing process, which will provide greater density, performance and energy efficiency. Intel hasn’t confirmed which lithography it will use, but there are only three possible options: Intel 4 and 4nm or 3nm from TSMCsince Ponte Vecchio combines several manufacturing processes: Intel 7 with 5nm and 7nm from TSMC.
Although energy efficiency is mentioned, a format will be used OAM v2 Able to bear up to 800W of powertherefore the efficiency will not be linked to the loss of performance for the simple fact of consuming less, and that is that Ponte Vecchio moves in the 600W. Intel points up to 30% increase for applications compared to its current graphics solution. For now, we’ll have plenty of time to wait for more details, and that’s because the first Rialto Bridge test chips shouldn’t see the light of day. until mid 2023.
|GPU HPC 2022-203|
|VideoCardz.com||NVIDIA H100SXM||AMD Instinct MI250X OAM||Intel Ponte Vecchio OAM||Intel’s Rialto Bridge OAM|
|GPUs||GH100||Aldebaran||Ponte Vecchio||Rialto Bridge|
|Transistors||80,000M||58,200M||100,000M||N / A|
|Die size||814mm²||2x ~790mm²||2x 640mm²||N / A|
|Manufacturing node||TSMC N4||TSMC-N6||Intel 7, TSMC N5+N7||Intel 4, TSMC N3 + N4 (?)|
|GPU clusters||132 (MS)||220 (UA)||128 Xe-cores||160 Xe-cores|
|L2 cache||50 MB||32 MB||408 MB||N / A|
|Interface||5120 bit||8192 bit||8192 bit||8192bit(?)|
|Memory||80 GB HBM3||128 GB HBM2e||128 GB HBM2e||HBM3|
|Interface/Format||SXM5/PCIe Gen5||OAM/PCIe Gen5||OAM/PCIe Gen5||OAM-V2|
via: video cardz