There is no argument that AMD, NVIDIA, Intel, and more companies are looking toward the future of artificial intelligence, machine learning, and supercomputers. We have seen AMD reveal their Instinct MI250X OAM module based on the dual-tile Aldebaran GPU and NVIDIA, with their H100 SXM5 high-performance computing module showcasing the GH100 GPU. Now, Intel is entering the scene with the Ponte Vecchio GPU will help the company reach more compute algorithms utilizing graphics card technology to access the data center and machine learning systems over the next decade.
Intel promises 100 billion transistors in the new Ponte Vecchio compute GPU to bring the company into more mainstream AI & high-performance computing applications
Last week, Intel confirmed on Twitter that the company has begun to test and sample the Ponte Vecchio compute GPU. Their new compute GPU will allow the company to give consumers the ability to access more development and functions through AI and HPC applications. In Western Europe, Mikael Moreau, Intel's PR manager, revealed the module that offers a unique cooling component. For users, this tells us that the company has already begun delivery of parts to partners located in the United States.
Voici Ponte Vecchio 🤩 pic.twitter.com/zbXf184nHk
— Miakel MoreIAu / Mr Intel by le JDH / Olympics COE (@fragtalife) May 6, 2022
Intel's next-gen Ponte Vecchio compute GPU offers up to 47 tiles, with the complete set containing a staggering 100 billion transistors. A compute GPU of this magnitude will require a hefty amount of power consumption, and the new OAM add module is planned to require 600W. Since it will be running under such constraints, the system will need adequate cooling. Intel has chosen to look at liquid cooling techniques for the new compute GPU, but in a minimum footprint compared to larger system liquid coolers. What is unknown from the Twitter post is whether we are looking at the standard Ponte Vecchio compute GPU or an advanced XT variation of the next-gen GPU.
The Intel Ponte Vecchio appears to be quite large, almost to the size of a standard tablet. However, this is not shocking due to the amount of computing power required and the number of transistors. The Ponte Vecchio is slated to offer "PetaFLOPS-class AI performance," a terminology that Raja Koduri of Intel has quoted.
Moreau also posted this on Twitter:
Tiens je crois que j’ai un sapphire rapids HBM et un Ponte Vecchio entre les mains 🤩
— Miakel MoreIAu / Mr Intel by le JDH / Olympics COE (@fragtalife) April 27, 2022
Translated, the Tweet from Moreau reads:
Here I think I have a sapphire rapids HBM and a Ponte Vecchio in my hands
If this is true, both components will process the 2 ExaFLOPS Aurora supercomputers scheduled to activate later this year.
Next-Gen Data Center GPU Accelerators
| GPU Name | AMD Instinct MI250X | NVIDIA Hopper GH100 | Intel Ponte Vecchio |
|---|---|---|---|
| Packaging Design | MCM (Infinity Fabric) | Monolithic | MCM (EMIB + Foveros) |
| GPU Architecture | Aldebaran (CDNA 2) | Hopper GH100 | Xe-HPC |
| GPU Process Node | 6nm | 4N | 7nm (Intel 4) |
| GPU Cores | 14,080 | 16,896 | 16,384 ALUs (128 Xe Cores) |
| GPU Clock Speed | 1700 MHz | ~1780 MHz | TBA |
| L2 / L3 Cache | 2 x 8 MB | 50 MB | 2 x 204 MB |
| FP16 Compute | 383 TOPs | 2000 TFLOPs | TBA |
| FP32 Compute | 95.7 TFLOPs | 1000 TFLOPs | ~45 TFLOPs (A0 Silicon) |
| FP64 Compute | 47.9 TFLOPs | 60 TFLOPs | TBA |
| Memory Capacity | 128 GB HBM2E | 80 GB HBM3 | 128 GB HBM2e |
| Memory Clock | 3.2 Gbps | 3.2 Gbps | TBA |
| Memory Bus | 8192-bit | 5120-bit | 8192-bit |
| Memory Bandwidth | 3.2 TB/s | 3.0 TB/s | ~3 TB/s |
| Form Factor | OAM | OAM | OAM |
| Cooling | Passive Cooling Liquid Cooling | Passive Cooling Liquid Cooling | Passive Cooling Liquid Cooling |
| TDP | 560W | 700W | TBD |
| Launch | Q4 2021 | 2H 2022 | 2022? |
