Either you had inside information or you made a very good guess prior to the presentation, because Nvidia has announced precisely that.
The only possible difference is that Nvidia did not specify whether the ARM core is Neoverse V1 or Neoverse N2, as both are successors to the current Neoverse cores.
Nvidia's been hiring server CPU folks for a while and Power10 has no NVlink. Taken together, that pointed pretty strongly to a server CPU being announced at some point, though I didn't seriously think it would be this early (figured it would wait until after there's some resolution to the ARM acquisition.)
"I am so happy today to announce nVidia's new business model: Vaporware GPUs! From now on we can paper launch any GPU with any specifications we care to create without the added burden of having to actually make and ship the hardware that will do all the wonderful things we claim it will do! And people will believe it--because we can blame the miners, the FABs, and the UFOs for the almost total shortages! It Just Works! I'm so happy today! Thank you very much. Our next vaporware GPUs coming in 3-2-1..."
Sorry, but I'm tuned out until general availability of gaming GPUs improves to a point where either lottery levels of luck and/or enriching scalpers is no longer mandatory.
Just a FYI: the headline reads "The NVIDIA GTC 2021 Keynote Live Blog (Starts at 8:30am PT/16:30 UTC)", but it started at 15:30 UTC. Since UTC doesn't have daylight savings, the conversion was wrong.
"NVIDIA will not stop supporting x86 Instead they'll support both Arm and x86"
For now. Nvidia is in it's final face to create a 100% Nvidia ecosystem, something they started with their first Tegra. And they want to finish ARM's acquisition before moving against the X86 platform.
Hey nVidia, the only care you have on the shelf is the GT 1030, less performance than the GTX 750 from 2014 for more money, $100 USD. Start with releasing decent entry level cards and get them on the shelf. GTX is everything that is wrong with nVidia. MONEY MONEY MONEY. You, gamers, pay for nVidia to enter the datacenter and machine learning. There's a lot of opportunity for some company one day to actually focus on making a video card again.
Low-margin products are going to be the last ones to migrate onto newer and more expensive process nodes. Until there's more production capacity on 7 or 8 nm, don't expect to see new, low-end products on it.
Samsung's node is not newer or expensive. nVidia should do a public service here and create volume over profit for a few months. This is not like last time, there have literally been no gaming GPUs on the shelf for half a year already. Nothing to buy.
Wait a sec. Why do they claim PCIe 4.0 is only 16 GB/s? If it's a x16 slot, it should be 32 GB/s per direction (so, 64 GB/s if counting the same way they count NVLink bandwidth).
Do they use just x8 or x4 PCIe 4.0 links, for the A100? Or did someone make a convenient mistake?
In the livestream, Nvidia was talking about their existing DGX products (and their top end model, in particular), which use a massive PCIe 4.0 switch to share the dual AMD CPUs 128 (total) PCIe lanes among NVMe, 200Gbps NICs, and GPUs.
The NICs alone take 128 lanes. GPUs take another 128 lanes. SSDs another 16 lanes. There are options for another two NICs, each taking 16 lanes.
Even with a max PCIe lane config from AMD (2 x EPYC CPUs for 192 lanes, at the cost of CPU-CPU interconnect bandwidth), that's not nearly enough without using a PCIe switch (or rather, a lot of PCIe 4.0 switches). At the system's max throughput, each GPU can only shuffle off 16GB/s of data.
In the end, Nvidia usually compares their next gen products to their current products. IMO, they see the PCIe 4.0 switches as a bottleneck, especially w.r.t. throughput of the whole system.
Some people would probably rather Nvidia compare their top end DGX configs against competing systems, but... what systems? Cerberus' wafer scale chip with no PCIe link to really care about? Google TPU with it's own network? Fujitsu's Fukagu's Tofu? POWER9 + NVLink in Sierra/Summit? None of them are really all that similar, even if they are still turing complete.
It's a nice presentation, but from where I sit, all the stuff about graphics is lip service. Nvidia 'graphics' cards are now just crypto miners. There is next to zero consumer market availability.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
29 Comments
Back to Article
SarahKerrigan - Monday, April 12, 2021 - link
In which... Nvidia announces a Neoverse-V1 server CPU with integrated NVlink.(Probably not, but it would be cool.)
FreckledTrout - Monday, April 12, 2021 - link
What about new spatulas?SarahKerrigan - Monday, April 12, 2021 - link
I WAS KIDDING.AdrianBc - Monday, April 12, 2021 - link
Either you had inside information or you made a very good guess prior to the presentation, because Nvidia has announced precisely that.The only possible difference is that Nvidia did not specify whether the ARM core is Neoverse V1 or Neoverse N2, as both are successors to the current Neoverse cores.
SarahKerrigan - Monday, April 12, 2021 - link
Nvidia's been hiring server CPU folks for a while and Power10 has no NVlink. Taken together, that pointed pretty strongly to a server CPU being announced at some point, though I didn't seriously think it would be this early (figured it would wait until after there's some resolution to the ARM acquisition.)No inside information.
Operandi - Monday, April 12, 2021 - link
My bet is egg whisks this time around. At least 25 of them, and a multitude of colors of course.WaltC - Monday, April 12, 2021 - link
From the keynote:"I am so happy today to announce nVidia's new business model: Vaporware GPUs! From now on we can paper launch any GPU with any specifications we care to create without the added burden of having to actually make and ship the hardware that will do all the wonderful things we claim it will do! And people will believe it--because we can blame the miners, the FABs, and the UFOs for the almost total shortages! It Just Works! I'm so happy today! Thank you very much. Our next vaporware GPUs coming in 3-2-1..."
Unashamed_unoriginal_username_x86 - Monday, April 12, 2021 - link
Yeah the profit margin from 0 expenses on 0 revenue is insane. Expecting more companies to pretend to make products soonmrvco - Monday, April 12, 2021 - link
Sorry, but I'm tuned out until general availability of gaming GPUs improves to a point where either lottery levels of luck and/or enriching scalpers is no longer mandatory.Unashamed_unoriginal_username_x86 - Monday, April 12, 2021 - link
Good, this keynote isn't for gaming GPUs.mbucdn - Monday, April 12, 2021 - link
Why bother having an announcement we have no products on the shelf. This company is wasting time with a keynote speech.mode_13h - Monday, April 12, 2021 - link
I'm pretty sure the products and services discussed in the talk *are* available, of course with the exception of the explicitly unreleased ones.This high-margin stuff is where a lot of the production capacity is going.
Peskarik - Monday, April 12, 2021 - link
Unicorn with integrated Unicornlink.jamesindevon - Monday, April 12, 2021 - link
Just a FYI: the headline reads "The NVIDIA GTC 2021 Keynote Live Blog (Starts at 8:30am PT/16:30 UTC)", but it started at 15:30 UTC. Since UTC doesn't have daylight savings, the conversion was wrong.Ryan Smith - Monday, April 12, 2021 - link
Oh for the love of Pete...Thanks!
yannigr2 - Monday, April 12, 2021 - link
"NVIDIA will not stop supporting x86Instead they'll support both Arm and x86"
For now. Nvidia is in it's final face to create a 100% Nvidia ecosystem, something they started with their first Tegra. And they want to finish ARM's acquisition before moving against the X86 platform.
mode_13h - Monday, April 12, 2021 - link
I'm sure they'll continue to support x86 as long as it holds a significant marketshare.UltraWide - Monday, April 12, 2021 - link
CPU + GPU + DPU, they are going hard after the last piece of the data center: Intel's high-margin CPUs.Alistair - Monday, April 12, 2021 - link
They are not going hard, they basically don't have a product for the next 4 years even...Alistair - Monday, April 12, 2021 - link
Hey nVidia, the only care you have on the shelf is the GT 1030, less performance than the GTX 750 from 2014 for more money, $100 USD. Start with releasing decent entry level cards and get them on the shelf. GTX is everything that is wrong with nVidia. MONEY MONEY MONEY. You, gamers, pay for nVidia to enter the datacenter and machine learning. There's a lot of opportunity for some company one day to actually focus on making a video card again.gescom - Monday, April 12, 2021 - link
Why cry? It's a business model and it works. Want a Nvidia GPU? 1500+ EUR and it's yours.mode_13h - Monday, April 12, 2021 - link
Low-margin products are going to be the last ones to migrate onto newer and more expensive process nodes. Until there's more production capacity on 7 or 8 nm, don't expect to see new, low-end products on it.Alistair - Monday, April 12, 2021 - link
Samsung's node is not newer or expensive. nVidia should do a public service here and create volume over profit for a few months. This is not like last time, there have literally been no gaming GPUs on the shelf for half a year already. Nothing to buy.jesuscat - Monday, April 12, 2021 - link
I think that's as much a Crypto issue as anything else.Even if NVIDIA had better yields, Crypto will create this shortage.
mode_13h - Monday, April 12, 2021 - link
Wait a sec. Why do they claim PCIe 4.0 is only 16 GB/s? If it's a x16 slot, it should be 32 GB/s per direction (so, 64 GB/s if counting the same way they count NVLink bandwidth).Do they use just x8 or x4 PCIe 4.0 links, for the A100? Or did someone make a convenient mistake?
jeremyshaw - Tuesday, April 13, 2021 - link
In the livestream, Nvidia was talking about their existing DGX products (and their top end model, in particular), which use a massive PCIe 4.0 switch to share the dual AMD CPUs 128 (total) PCIe lanes among NVMe, 200Gbps NICs, and GPUs.Eight 200Gbps NICs (each taking 16 lanes of PCie 4.0), 4 NVMe SSDs (4 lanes each), and 8 GPUs (16 lanes each).
The NICs alone take 128 lanes. GPUs take another 128 lanes. SSDs another 16 lanes. There are options for another two NICs, each taking 16 lanes.
Even with a max PCIe lane config from AMD (2 x EPYC CPUs for 192 lanes, at the cost of CPU-CPU interconnect bandwidth), that's not nearly enough without using a PCIe switch (or rather, a lot of PCIe 4.0 switches). At the system's max throughput, each GPU can only shuffle off 16GB/s of data.
In the end, Nvidia usually compares their next gen products to their current products. IMO, they see the PCIe 4.0 switches as a bottleneck, especially w.r.t. throughput of the whole system.
Some people would probably rather Nvidia compare their top end DGX configs against competing systems, but... what systems? Cerberus' wafer scale chip with no PCIe link to really care about? Google TPU with it's own network? Fujitsu's Fukagu's Tofu? POWER9 + NVLink in Sierra/Summit? None of them are really all that similar, even if they are still turing complete.
mode_13h - Tuesday, April 13, 2021 - link
Ah, thanks for that detailed answer!chrysrobyn - Tuesday, April 13, 2021 - link
It's a nice presentation, but from where I sit, all the stuff about graphics is lip service. Nvidia 'graphics' cards are now just crypto miners. There is next to zero consumer market availability.edzieba - Tuesday, April 13, 2021 - link
I'm going to be disappointed if the next in-person Nvidia keynote does not feature a 'Jensen's kitchen' set.