Wednesday, August 24th 2022
NVIDIA Grace CPU Specs Remind Us Why Intel Never Shared x86 with the Green Team
NVIDIA designed the Grace CPU, a processor in the classical sense, to replace the Intel Xeon or AMD EPYC processors it was having to cram into its pre-built HPC compute servers for serial-processing roles, and mainly because those half-a-dozen GPU HPC processors need to be interconnected by a CPU. The company studied the CPU-level limitations and bottlenecks not just with I/O, but also the machine-architecture, and realized its compute servers need a CPU purpose-built for the role, with an architecture that's heavily optimized for NVIDIA's APIs. This, the NVIDIA Grace CPU was born.
This is NVIDIA's first outing with a CPU with a processing footprint rivaling server processors from Intel and AMD. Built on the TSMC N4 (4 nm EUV) silicon fabrication process, it is a monolithic chip that's deployed standalone with an H100 HPC processor on a single board that NVIDIA calls a "Superchip." A board with a Grace and an H100, makes up a "Grace Hopper" Superchip. A board with two Grace CPUs makes a Grace CPU Superchip. Each Grace CPU contains a 900 GB/s switching fabric, a coherent interface, which has seven times the bandwidth of PCI-Express 5.0 x16. This is key to connecting the companion H100 processor, or neighboring Superchips on the node, with coherent memory access.
Serial processing muscle on the NVIDIA Grace CPU is care of a 72-core Arm v9 64-bit CPU. A Superchip would contain 144 cores. The main memory interface is LPDDR5x, with each "socket" having a maximum memory bandwidth of 1 TB/s (or rivaling that of over 24 channels of DDR5). This includes ECC. A key serial-IO interface is PCI-Express Gen 5, with 68 lanes on offer. These are mainly to wire out NVMe storage devices. The chip has a TDP rating of 500 W peak.
The Grace CPU demonstrates the engineering muscle of NVIDIA at designing large multi-core processors for enterprise and HPC applications. With Arm achieving near-parity with x86-64 in performance, efficiency, and IPC, we're beginning to understand why NVIDIA couldn't become an x86 licensee. It would have achieved a winning enterprise processor rivaling Intel's much before. Future generations of NVIDIA's DGX compute nodes, as well as pre-built workstations and servers, spanning a multitude of applications, could see NVIDIA wean away from x86-based CPUs, replacing them with Grace and its successors.
Source:
Wccftech
This is NVIDIA's first outing with a CPU with a processing footprint rivaling server processors from Intel and AMD. Built on the TSMC N4 (4 nm EUV) silicon fabrication process, it is a monolithic chip that's deployed standalone with an H100 HPC processor on a single board that NVIDIA calls a "Superchip." A board with a Grace and an H100, makes up a "Grace Hopper" Superchip. A board with two Grace CPUs makes a Grace CPU Superchip. Each Grace CPU contains a 900 GB/s switching fabric, a coherent interface, which has seven times the bandwidth of PCI-Express 5.0 x16. This is key to connecting the companion H100 processor, or neighboring Superchips on the node, with coherent memory access.
The Grace CPU demonstrates the engineering muscle of NVIDIA at designing large multi-core processors for enterprise and HPC applications. With Arm achieving near-parity with x86-64 in performance, efficiency, and IPC, we're beginning to understand why NVIDIA couldn't become an x86 licensee. It would have achieved a winning enterprise processor rivaling Intel's much before. Future generations of NVIDIA's DGX compute nodes, as well as pre-built workstations and servers, spanning a multitude of applications, could see NVIDIA wean away from x86-based CPUs, replacing them with Grace and its successors.
54 Comments on NVIDIA Grace CPU Specs Remind Us Why Intel Never Shared x86 with the Green Team
edit:
My point => .
The heads of all the people telling me that this is a completely different market segment, which I am well aware of: O O O
edit: ok it looks like sapphire rapids will have over 80 PCIe lanes so Nvidia’s solution is behind on this spec.
Zen 4 EPYC will have 128 PCIe 5.0 lanes if you want apples-to-apples.
And yes, the above is a joke.
Anyway, an Nvidia desktop CPU would also be welcome. Of course it won’t have or need as many PCIe lanes as 68 but more choices are always nice.
NVIDIA is fully invested in x86 replacement chips based on ARM - on all segments.
That said, it looks like the competition will really be between NVIDIA and AMD. AMD added Xlinx to their portfolio, also own an ARM license, and are jointly working with Samsung to integrate RDNA with elements of ARM (via Exynos), which would help them combat NVIDIA across all platforms too. This is assuming NVIDIA also ports elements of this CPU down into their next-gen gaming tablets (and the next-gen Switch, assuming Nintendo sticks with NVIDIA), and even some gaming laptops running either Steam OS or Windows ARM.
Meanwhile, Intel, despite all their recent acquisitions, haven't really gotten anything to show for it, aside from Foveros, and it'll be awhile longer before their own compute and gaming GPUs can prove reliable enough in the high-value markets. Kind of wild to see such a dramatic shift the last 5 years.
Nvidia did a mistake to not REALLY concentrate on ARM sooner and produce products like Grace long ago. Not this kind of huge processors for servers from the beginning maybe, but SOCs for laptops and desktops, or if not desktops, at least mini PCs, running Windows on ARM, or Linux, or Android, or all of them. Qualcomm is a sleeping, boring, failure in that area.
They'll probably start accelerating in the ARM platform now. They lost time waiting to see if they can first have the absolute control of ARM. No one wanted them, so it's good to see that their pride and arrogance - which is part of their business mentality, sometimes helps them, mosts times, it doesn't - is not becoming an obstacle to their plans to start developing CPUs also.
As much as Intel needs GPUs for it's future, the same Nvidia needs CPUs for it's future. We all saw what happened to Nvidia's financials this quarter, because they only stand on one foot. GPUs. Hit that foot and the whole company trembles.
I haven't miss something.All in on ARM? Really? Tell me some products that where made for mass production and availability to the general public. Except the obvious mention to Switch, what else is there? Shield tablet?
You both missed my point.No, he is right. Nvidia had always better vision than AMD, it was more ambitious, and was in better position to utilize the hardware it was making, thanks to it's software and better promote it's products thanks to the much more aggressive marketing.
The only time in history when AMD did a bold move, that eventually saved it, was when it bought ATI.
(not the whole list obviously, just a few examples)