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
I was very clear in my first post. I was very clear in my reply. Instead of acknowledging what I meant, you chose to play dump and throw me a list of companies that maybe produced some products 10 years ago, when Nvidia did tried to get in smartphones.
Now that you can't continue that charade, you run away saying that I am the one shifting the goalposts.
Go and play your games elsewhere.
Source.
For all that terabyte of bandwidth and juvenile “SuPErCHiP” naming, it’s merely “competitive” (read: slower but not a whole lot slower) with current offerings. Imagine designing your processor with an exotic memory interface and giving it a silly grandiose name, only for it to be slower than year old processors using bog-standard DDR4-3200. IPC uplift leaks suggest 96-core Genoa will be faster than the whole dual-processor “sUpERchIp” package :laugh:
I agree with the article, maybe if Nvidia had an x86 license they could have come up with something more competitive. This is a great (although very costly) demonstration that the reports of x86 being dead are greatly exaggerated :sleep:
There's also a myriad of smartphones and the gaming series with Ouya, Shield Portable and Nintendo Switch series.
All in all, NVIDIA has a long history with ARM and has scored multiple products with them from the biggest OEMs. Were they unsuccessful? Arguably mostly yes, but not usually as a consequence of technical merit.To be honest the STH article you linked explains it pretty well. Grace is not supposed to compete with x86 in raw compute, but is a platform for NVIDIAs GPU efforts. The CPU-GPU coherency along with NVLink networking is a big deal with scaling their AI/ML solutions.
Historically NVIDIA went from IBM POWER to Intel x86 to AMD x86 as the basis of their GPU clusters, so it makes sense for them to utilize previous ARM expertise and complete the walled garden with an in-house CPU platform ;)
With the Mellanox acquisition they are pretty well vertically integrated now.
I also explained in that first post what I meant by saying sooner. It's there. Just READ it. Don't avoid parts of my posts just to create not existent excuses, to give a different meaning of what I wrote. You misread what I write, and then accuse me for what you misunderstood, or chose to give a different explanation, an explanation of your own. You think I don't know Nvidia's history or that I haven't seen their first efforts in the ARM platform? But they just gave up for something REALLY serious in the retail market. I think their excuse was Qualcomms anticompetitive tactics back then. Give me a brake, my first Nvidia product was a GeForce 2 MX.
The products you show are 10 years old. Haven't seen what Nvidia was doing in the Chromebook market to be honest, but the last 5-10 years they where mostly making specialized boards and nothing else with the exception of Shield and Swift. Do you understand that "REALLY" in my first post, or are you determent to NOT understand what I am saying even from that first post? When we have so many companies succeeding in the ARM market, companies (much) smaller than Nvidia and knowing Nvidia's potential, it's easy to assume that they just lost interest and only now they are coming back. Hope not just to support their server vision, but for more.
As for Nvidia's long history, did I said they just got a license? What? I am shifting the golposts again? No. You just DON'T READ what I write.
first dual core soc in a phone, even
People read your posts, but quite often something is missing from your version of history
en.m.wikipedia.org/wiki/LG_Optimus_2X
They also pushed on the GPU angle but (graphics intensive-) gaming in phones never really took off;
developer.nvidia.com/embedded/buy/tegra-k1-processor
And this was their niche, the unique selling point Nvidia had and now expands on with Grace as well: gpu acceleration.
Overall I dont think Nvidia can be blamed for lack of trying to get into this market... the ultimate push cost them dearly;
nvidianews.nvidia.com/news/nvidia-to-acquire-arm-for-40-billion-creating-worlds-premier-computing-company-for-the-age-of-ai
Gaming on phones doesn't really needs graphics. At least most games in smartphones don't. They are very simple in graphics. i believe 3D intensive gaming in smartphones is a minority, but i could be wrong here.
I think Nvidia wasn't really investing on the ARM platform, because it could not control the platform and what direction that platform would go. That's why they tried to buy ARM. They had a license, they even had a license to build custom cores, but they probably also had reasons to not invest heavily on that platform. If they had bought ARM, no Qualcomm could sabotage their negotiations with big smartphone companies and also they could turn the ship more in the server direction, than the smartphone market. They couldn't buy ARM, but they can't do something about that either. They need CPUs for their server aspirations at least.
Nvidia never got to an efficient enough design to compete. They sold their SoC to LG and their dual core phone was fast but not efficient at that. (I owned one..) Tegra K1 similarly guzzles power. And Tegra STILL shines in one device: Shield TV. Why? It doesnt run on a battery ;) And its still arguably the best android Tv device you can hook onto a TV.
The ARM market is open. At the server end, many core CPUs like ThunderX trumped their design too for server.
You can spin it however you feel but the reality is, Nvidia certainly did try, but their advantage in tech just couldnt get translated into something meaningful while competitors already had some iterations and refinements under the hood. The whole reason Grace is the subject now, again is because it helps their GPU division. Not because they built the most efficient CPU. They have shareholders too; these adventures cant last forever.
You even answered the issue yourself by stating smartphones dont really need graphics in a big way. Its not entirely realistic to then expect them to push further on stuff they are behind on anyway...
As for the olde idea of Nvidia wanting market control and not just share... you might be right about that. But the company really did try the fair play approach; let the design/tech speak for itself. Its all Nvidia really does and did at its core. Their GPU leadership is not the result of market control. Its a result of consecutive design wins. An approach they tried just the same with Tegra.
On another note, cracking into the x86 CPU business is probably impossible today. AMD is very lucky to have x86_64 in their courtyard, that forces Intel to cross-license the rest of the instruction set. A third player would have to get a license from both Intel and AMD and will probably never be able to compete on price because of that.
Still come 2025-27 I think we are in for a treat. AMD is not ignoring Apple either.
Nvidia was always trying to gain market control. That's what forces them to innovate and makes them look like they are one or two steps in front of the competition. CUDA, PhysX(they bought the tech, but who else had a chance to promote it stronger?), G-Sync, DLSS, RayTracing(it's not their, but the same as PhysX). Fair game and Nvidia doesn't usually go together and when I was reading 10 years or so ago Nvidia's protests about Qualcomm's anticompetitive tactics in the SOC market, I was laughing. Nvidia wasn't lying. Qualcomm had to face fines if I remember correctly from the courts and agree to play nice in the future.Nvidia is as arrogant as Apple. But they don't have the huge customer base Apple does. Apple sells devices, Nvidia sells parts, not whole devices. It's not easy to make loyal customers who buy your parts, to convince them to buy whole devices, just because of your logo on them. But Nvidia could start offering ARM SOCs for smartphones, laptops, (mini) desktop PCs and build reference, all Nvidia hardware, devices for others to build.
Apple will never sell it's stuff to others, because if everyone could build iPhones and Macs, those devices could lose their premium image. Many people buy iPhones and Macs with the same mentality they buy jewelery.
Because honestly saying that yet another ARM clone had been rolled out today, this time, by NV, is not flashy enough.Ah, so that is why both Microsoft and Apple told it to go have solo kamasutra: "better vision".
Apple paywalled people by anything from Apple anyway.
Why would the rest of the world bother?
In the case of consoles, Nvidia either didn't had the hardware, a SOC that would combine both a CPU and a GPU and be competitively priced compared to AMD's APUs, or the low profit margins made them ignore that market (and I thank them for that choice, if that was the case, because I could be forced to be buying quad core Intel's for the rest of my life).
The need for massively parallel processing has exactly 0 to do with NV.That is one way to refer to "they bought PhysX".Proprietary crap.That is dead, and good riddance.Used that thing to milk more money off the customer. The tech itself will likely follow the path of GSYNC.And at the moment, the most noticeable hint at "RT is on" is the massive drop in framerates. At the same time, this thing is not uses NV's RT:
If memory serves me well Nvidia pioneered programmable shaders.
Which once again brings us back to their attempted acquisition of Arm; I still struggle to see the reasoning behind it. The argument that it was to integrate NVIDIA graphics into Arm CPUs doesn't wash because NVIDIA's focus has always been high-performance high-power graphics, not low-end low-power ones as found in typical Arm applications, so they would essentially have to build an entirely new product. The thing is though, that doesn't require them to buy Arm; if NVIDIA already has a low-power GPU capable of competing with what's typically found in smartphones, there's absolutely nothing stopping them from just licensing or selling it as a standalone product.
The cynical take is that it's simply so NVIDIA could increase Arm licensing fees and reap the profits, but I really don't see that panning out well for them; it would almost certainly have pushed a lot of Arm licensees towards the royalty-free RISC-V, which makes it a self-defeating proposition.
But, besides trying to develop something new in the open taking a lot more time, look at the compute situation. Closed approach: CUDA, doing its job just fine and is now nearly ubiquitous. Open approach: OpenCL, a spec everyone hates to use, thus almost no one uses it. It doesn't always pan out like that, but what I'm trying to say is going the open route is not an automatic win, I will not fault anyone for going for the closed route.
A Ryzen on Intel 20A? I want to see it, to be honest with you.
I think Nvidia having a... dramatic past was one of the reasons for them trying to buy ARM. Intel refusing to give them a license for it's back then new CPUs drive Nvidia's chipset department to an end. Qualcomm's anticompetitive tactics probably played a major role in Nvidia not pushing in the smartphone market. They probably thought that, in the case of ARM, there could be another company trying to buy them today or in the future and they should move first. Think of Apple buying ARM. They have the money, the power and the customer base to just terminate all licenses and say to the other companies "Go and try your lack with RISC-V, we just don't care!". A simplified thought that last one, but, who knows? I think Nvidia didn't wanted to invest heavily in something they couldn't control. Also buying ARM, as I said, it could give them, not just an advantage over competition, keeping the best design for themselves, but also to choose where the ARM platform will focus. Now they get what ARM is making and then change that to suit their needs. With them having control over ARM, they would create cores for their needs and then let their licensees to spend time and money to make those cores suits their needs.
Anyway, just speculating here.