I'm equally surprised by this. I guess with the reputation and global reach Nuvia has it doesn't hurt to name your organization after them, and it likely helps they have no presence in the USA, but what happens when NUVIA (ARM engineering) enters the Canadian market where Nuvia (nuclear engineering) has offices?
So, Amazon, Google, and Microsoft. Are there any other potential customers for this? At least 2 of those are already involved with designing their own chips.
it have one big gain, they offset risk to someone else. so even if they have their own chip and this unrelated team manages to make something incredible, those 3 will merge nuvia chip magic with their own designs gaining performance. " so alll they do is branched out chip design that might bring unexpected gains that would normally die inside corpo design procedures. its not a high cost if they really get to something useful.
I don't think Amazon is going to be a large customer. They have invested billions in Annapurna which makes hyperscale ARM designs, while NUVIA doesn't appear in the short term to be targetting that segment...but who knows, there is little information here on actual material products because they don't have one yet.
There is definitely room to fill a void though, that is, HPC and nearline server rackspace, where there are few legitimate options.
That's true, but getting adoption for an entirely new IP is going to take some huge reach. In the short term I suspect they will make ARM-compatible designs, unless they have some amazing devkit that can port software or emulate instructions via hardware translation (ie Transmeta Crusoe.)
They won't be able to break into a flooded market with an incompatible product. ARM hasn't even been able to pull that off in most markets outside of mobile and they've been trying for 30 years.
You wouldn't want to replace ARM, the risk of something goes wrong is too high. You would want to just pay a license fee for new cores that everyone else use anyways, so even if there's a problem everyone will have a problem or a solution for everyone.
The question that matters is: why did GW3 do this? The naive answer is "Apple wouldn't let him create server cores". To me that sounds insane; obviously Apple will soon be using the same tech they'll use for iMac Pro and Mac Pro in data centers.
So the answer must be something else, in other words: - get very rich or - create an empire.
If THOSE two are the goals, then what helps further them? One obvious answer is acquisition. Obviously helps with the very rich, and done properly allows empire creation.
So who might engage in such an acquisition? It's easy to suggest possibilities, without being able to say anything certain until both the nVidia/ARM deal and the future course of ARM servers are clear (the latter perhaps by this time next year, once we see what Amazon does with V1, and what the "just you wait for AMD-next and Intel-next" camp have delivered. But some possibilities that seem plausible include - nVidia (especially if the ARM deal does not go through) - Microsoft (we can do anything Apple can do, only better) - AMD (we can see the twilight of x86 as well as anyone, and we won't fumble it this time)
Less plausible (but not insane, IMHO) possibilities - Cisco (full service data center provider...) - IBM (POWER had a good run, but what if we couple our memory/IO to ARM's raw compute performance?) - Intel (look, we screwed with x86, OK, we admit it. But we're the company you trust for your enterprise IT, and boy do we have a good range of new server chips for you. And your autos will also love them.)
"Apple wouldn't let him create server cores" is exactly the reason. I've heard from multiple primary sources. Now that may change in the future. Apple doesn't sell the chips it makes anywhere else - if it made its own server processors, it would only use them themselves, and they'd be rebuilding their entire back-end infrastructure from x86 to Arm. You have to wonder what benefit for Apple that would be.
I think it would be interesting if one of these companies released an arm workstation running Windows on ARm. Server back ends have had different architectures off on but if one of these new ARM startups could break the biggest rock in the industry, wintel, It would get people talking. Right now, Microsoft is trying and failing to get off x86 due to the anemic performance of Qualcomm when compared with other modern intel and arm cpu's.
I see Apple at some point launching the equivalent of AWS for Apple developers -- a way to provide remote compute service for non-huge companies that provides Apple-like APIs, security, and hooks into the Apple ecosystem. Something like that is much more desirable if it's running the same ISA as the rest of Apple.
More generally, if ARM servers are overall a better value proposition (cost, performance, security) for everyone, the same is true for Apple. So they can buy ARM servers from Ampere, or they can use their own chips. Using their own chips seems cheaper and gives more control.
Ultimately few people are going to say "the reason I'm leaving my company to start a new one is because I want to become Silicon Valley Rich" or "start my own empire". I mean, that's just human nature. You clothe your exit (maybe even your tell yourself) it's somehow for the benefit of humanity.
You don't make crazy money in HW. Especially in licensable CPU cores, look at ARM considering their size they're not making crazy money on CPU licensing. Spending a few years making wattsapp or byteDance or SurveillancePhoneApp2.0 will get you a far higher return in SV than designing HW that requires another company to integrate it into their products.
The primary functionality that matters for servers but is not captured by GB5 (or SPEC...) is large instruction footprints. BUT - Apple clearly cares about this. Witness the 128K I-cache. - Better I prefetching is a mostly solved problem over the past two years or so. A nimble company (Apple and presumably Nuvia, right away, ARM a year later, Intel some time around 2030) will implement those better prefetch algorithms.
There's simply no reason, on technical grounds, to expect that Nuvia can't deliver. Now maybe they can't deliver on execution grounds -- none of us know yet how well they can execute, though obviously the top management have all executed will in other contexts.
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
21 Comments
Back to Article
BedfordTim - Thursday, September 24, 2020 - link
I guess naming your company after a nuclear waste handler worked. It also helps keep their website well hidden.Samus - Friday, September 25, 2020 - link
I'm equally surprised by this. I guess with the reputation and global reach Nuvia has it doesn't hurt to name your organization after them, and it likely helps they have no presence in the USA, but what happens when NUVIA (ARM engineering) enters the Canadian market where Nuvia (nuclear engineering) has offices?Lord of the Bored - Saturday, September 26, 2020 - link
I'm more surprised it isn't a pharmaceutical name."Ask your doctor if Nuvia is right for you."
quorm - Thursday, September 24, 2020 - link
So, Amazon, Google, and Microsoft. Are there any other potential customers for this? At least 2 of those are already involved with designing their own chips.deil - Thursday, September 24, 2020 - link
it have one big gain, they offset risk to someone else. so even if they have their own chip and this unrelated team manages to make something incredible, those 3 will merge nuvia chip magic with their own designs gaining performance. "so alll they do is branched out chip design that might bring unexpected gains that would normally die inside corpo design procedures.
its not a high cost if they really get to something useful.
Samus - Friday, September 25, 2020 - link
I don't think Amazon is going to be a large customer. They have invested billions in Annapurna which makes hyperscale ARM designs, while NUVIA doesn't appear in the short term to be targetting that segment...but who knows, there is little information here on actual material products because they don't have one yet.There is definitely room to fill a void though, that is, HPC and nearline server rackspace, where there are few legitimate options.
dotjaz - Saturday, September 26, 2020 - link
But Annapurna doesn't design microarchitectures. NUVIA could still be a viable option to replace ARM's Neoverse core.Samus - Saturday, September 26, 2020 - link
That's true, but getting adoption for an entirely new IP is going to take some huge reach. In the short term I suspect they will make ARM-compatible designs, unless they have some amazing devkit that can port software or emulate instructions via hardware translation (ie Transmeta Crusoe.)They won't be able to break into a flooded market with an incompatible product. ARM hasn't even been able to pull that off in most markets outside of mobile and they've been trying for 30 years.
PandaBear - Friday, October 30, 2020 - link
You wouldn't want to replace ARM, the risk of something goes wrong is too high. You would want to just pay a license fee for new cores that everyone else use anyways, so even if there's a problem everyone will have a problem or a solution for everyone.Ian Cutress - Thursday, September 24, 2020 - link
They're also going after the general Arm server market, not just the hyperscalers. There's also Baidu, Tencent, Alibaba, Facebook, Oracle, etcname99 - Thursday, September 24, 2020 - link
The question that matters is: why did GW3 do this?The naive answer is "Apple wouldn't let him create server cores". To me that sounds insane; obviously Apple will soon be using the same tech they'll use for iMac Pro and Mac Pro in data centers.
So the answer must be something else, in other words:
- get very rich or
- create an empire.
If THOSE two are the goals, then what helps further them? One obvious answer is acquisition. Obviously helps with the very rich, and done properly allows empire creation.
So who might engage in such an acquisition? It's easy to suggest possibilities, without being able to say anything certain until both the nVidia/ARM deal and the future course of ARM servers are clear (the latter perhaps by this time next year, once we see what Amazon does with V1, and what the "just you wait for AMD-next and Intel-next" camp have delivered. But some possibilities that seem plausible include
- nVidia (especially if the ARM deal does not go through)
- Microsoft (we can do anything Apple can do, only better)
- AMD (we can see the twilight of x86 as well as anyone, and we won't fumble it this time)
Less plausible (but not insane, IMHO) possibilities
- Cisco (full service data center provider...)
- IBM (POWER had a good run, but what if we couple our memory/IO to ARM's raw compute performance?)
- Intel (look, we screwed with x86, OK, we admit it. But we're the company you trust for your enterprise IT, and boy do we have a good range of new server chips for you. And your autos will also love them.)
Ian Cutress - Thursday, September 24, 2020 - link
"Apple wouldn't let him create server cores" is exactly the reason. I've heard from multiple primary sources. Now that may change in the future. Apple doesn't sell the chips it makes anywhere else - if it made its own server processors, it would only use them themselves, and they'd be rebuilding their entire back-end infrastructure from x86 to Arm. You have to wonder what benefit for Apple that would be.Zeratul56 - Thursday, September 24, 2020 - link
I think it would be interesting if one of these companies released an arm workstation running Windows on ARm. Server back ends have had different architectures off on but if one of these new ARM startups could break the biggest rock in the industry, wintel, It would get people talking. Right now, Microsoft is trying and failing to get off x86 due to the anemic performance of Qualcomm when compared with other modern intel and arm cpu's.name99 - Friday, September 25, 2020 - link
I see Apple at some point launching the equivalent of AWS for Apple developers -- a way to provide remote compute service for non-huge companies that provides Apple-like APIs, security, and hooks into the Apple ecosystem. Something like that is much more desirable if it's running the same ISA as the rest of Apple.More generally, if ARM servers are overall a better value proposition (cost, performance, security) for everyone, the same is true for Apple. So they can buy ARM servers from Ampere, or they can use their own chips. Using their own chips seems cheaper and gives more control.
Ultimately few people are going to say "the reason I'm leaving my company to start a new one is because I want to become Silicon Valley Rich" or "start my own empire". I mean, that's just human nature. You clothe your exit (maybe even your tell yourself) it's somehow for the benefit of humanity.
webdoctors - Friday, September 25, 2020 - link
You don't make crazy money in HW. Especially in licensable CPU cores, look at ARM considering their size they're not making crazy money on CPU licensing. Spending a few years making wattsapp or byteDance or SurveillancePhoneApp2.0 will get you a far higher return in SV than designing HW that requires another company to integrate it into their products.PandaBear - Friday, October 30, 2020 - link
Amazon already bought Anapurna for $400M years ago. They won't buy it.Peskarik - Thursday, September 24, 2020 - link
Good luck to them. Innovation is good.DukeN - Thursday, September 24, 2020 - link
My gf loves their rings!Samus - Friday, September 25, 2020 - link
But do you? Tell us how you really feel!GuyOnABudget - Thursday, September 24, 2020 - link
For Ian: "two the final two acting as investor partners." (second paragraph)Excited to see NUVIA back their bold claims of "+40-50% IPC increases over Zen2 for only a third of the power," which we've seen before.
name99 - Friday, September 25, 2020 - link
Which part of this graph do you not understand?https://semiaccurate.com/assets/uploads/2020/09/Nu...
The primary functionality that matters for servers but is not captured by GB5 (or SPEC...) is large instruction footprints. BUT
- Apple clearly cares about this. Witness the 128K I-cache.
- Better I prefetching is a mostly solved problem over the past two years or so. A nimble company (Apple and presumably Nuvia, right away, ARM a year later, Intel some time around 2030) will implement those better prefetch algorithms.
There's simply no reason, on technical grounds, to expect that Nuvia can't deliver. Now maybe they can't deliver on execution grounds -- none of us know yet how well they can execute, though obviously the top management have all executed will in other contexts.