We have some new information on Intel’s first discrete graphics processing unit. Dubbed “Ponte Vecchio”, the GPU was teased during Intel’s recent HPC Developer Conference in Denver, Colorado.
Ponte Vecchio is based on Intel’s 7nm process node based dedicated Graphics Architecture, “Xe” (stylized as Xe). Intel has called this new graphics card its first “exascale graphics card” and has announced that it will debut in 2021 as part of the world’s first exascale class supercomputer, the Aurora Supercomputer at Argonne National Laboratory in Lemont, Illinois.
The keynote to the demonstration on Monday was delivered by Raja Koduri. Koduri use to work at Intel’s competitor, AMD, before switching to Intel in 2017.
Intel Does Graphics Cards?
Intel is known foremost as the world’s leading CPU makers. And while competition with AMD is tough at the moment, they still dominate a huge chunk of the CPU market. But in 2018, Intel announced that they were putting together a team to get in on the discrete graphics game. Currently the market leader in this sector is Nvidia. They are followed by Intel’s long time CPU competitors, AMD. The GPU market is booming at the moment, and Intel wants in.
After stealing Raja Koduri from AMD, the project has been on full throttle. The Ponte Vecchio graphics card that they recently teased is an exascale GPU, suitable for high end computing. Unlike it’s 7nm process node, the first consumer graphics card from Intel will be based on Intel’s fabled 10nm process node and may be available as early as 2020. Intel’s current and next gen CPUs are still based on the 14nm process node. They only plan on switching to the 10nm process for the 11th gen Core-i CPUs, out in 2021.
The Xe Graphics Architecture
But that’s enough on CPUs. This article is about Intel entering the discrete graphics race. Traditionally, Intel has only been focusing on Integrated Graphics Processors (IGPs). These are like tiny GPUs built into the CPU. But with the announcement of the Xe Graphics Architecture, they are moving on to a whole new level.
The Xe Architecture has been divided into three distinct designs. Each of these designs serve a different end. The three designs are:
- HP: Suitable for consumer GPUs, data centers and AI use cases
- LP: Integrated graphics processors that come with their CPUs
- HPC: Suitable for heavy computational tasks
Here’s how Raja Koduri summarized the decision to separate the designs:
“Architecture is a software compatibility contract. We originally were planning for two microarchitectures within Xe, our architecture (LP and HP), but we saw an opportunity for a third within HPC.”
The Architecture
While Intel has refrained from sharing key information on the specs of its upcoming GPUs, they still shared some critical infos, that has us got us excited.
For instance, unlike other architectures currently employed in the discrete graphics industry, Intel are choosing not to go with one large monolithic die for its graphics cards. Instead, they have adopted a multi-chip module (MCM) design. MCM consists of several different computational units that are held together by a fabric, in the case of Xe, the XE Memory Fabric (XEMF).
XEMF
“At the heart of Xe architecture we have a new fabric called XEMF. It is the heart of the performance of these machines. We called it the Rambo Cache. It is a unified cache that is accessible to CPU and GPU memory.”
XEMF is a new scalable fabric by Intel designed to hold the computational and memory units together within the MCM design of the Xe cards. Intel has also announced that XEMF can be scaled to thousands of Xe nodes. Koduri says it the heart of the Xe architecture and then termed it “Rambo Cache”. Coupled with Foveros technology, the XEMF provides a large amount of memory bandwidth that can be accessed by both the CPU and the GPU simultaneously.
The MCM design for the Xe Graphics card, means that the graphics and computational workload will be distributed among the multiple chips in the card. And coincidentally, it was also recently revealed that the Xe Graphics drivers would support multiple GPUs. So, let us connect the dots here. The Xe Graphics drivers will most likely play a huge role in managing the multi-chip architecture of Intel’s upcoming discrete graphics cards.
Intel has also announced that its upcoming Xe GPUs will support both SIMT (GPU) and SIMD (CPU) vector widths. This means that the Xe cards will be able to maximize performance by processing SIMT (Single-Instruction Multiple Threads) and SIMD (Single-Instruction Multiple Dispatch) simultaneously. With SIMD, the Xe GPUs will be able to provide AVX-type processing power to data center accelerators.
Intel’s 7nm Strategy
The “Ponte Vecchio” graphics card, when it debuts in 2021, will be the first product from Intel to feature a 7nm process node. Intel’s current stack of CPUs are based on the 14nm process node, and it plans on transitioning to a 10nm process with the 11th gen Core-i processors in 2021. Its competitor AMD has already found massive success this year with the 7nm process for its Ryzen line of CPUs. But how good will the 7nm process serve the Xe Graphics Architecture. As it turns out: Very Well!
The inner design of a GPU consists of many repeated structures which makes it easy for the unit to cope with defects. Also, adding more redundant structures in areas like the critical pathways will help prevent risks of manufacturing the die during early stages of the node. And to add to all this, Intel has always been using small chiplets which can drastically improve the yield.
Intel will be using one its coolest new asset, the Foveros 3D chip packaging technology, to actualize the Xe Graphics Architecture. The technology uses a 3D stacked die design that lets them use larger process nodes for non-computational elements like I/O. And in order to tie the HBM packages to the compute die, Intel makes use of an Embedded Multi-Die Bridge (EMIB).
One final cool thing about the architecture is that the Xe cards will be able to scale up to many thousand EUs (execution units). The EUs will be spread across multiple chiplets. And each EU will improve the double-precision floating point performance of the card by 40 folds.
Ponte Vecchio & The Aurora Supercomputer
During the event, Intel announced that it the Ponte Vecchio graphics card would make its debut in 2021. It will be powering the Aurora supercomputer at Argonne National Laboratory in Lemont, Illinois.
So far there isn’t much we know about the Aurora supercomputer. We know that it is a joint project between the DOE and Intel. It will also become the world’s very first exascale supercomputer. Intel has released some information on the architecture of Aurora. It will have two Sapphire Rapids processors and six Ponte Vecchio graphics units. Yes, that is a lot of firepower. The Aurora will be one of the world’s most powerful supercomputers. The architecture also features an all-in-all connections within the nodes. This will insure that the data will be delivered through all eight fabric endpoints in each node.
Conclusion
So it seems pretty clear now that Intel mean business with their entrance into the discrete graphics market. The first of these Xe Architecture graphics cards from Intel to have been announced is the Ponte Vecchio which will debut as part of the Aurora Supercomputer at Argonne National Laboratory in 2021. We expect to hear more news about the consumer level Xe graphics card in 2020. We will keep you posted.
Traditionally the GPU game has been a dual between Nvidia and AMD. But the days of Green vs Red are coming to and end. Enter the days of Green vs Red vs Blue.
Feel free to leave a comment below.
