Whichever of Intel's VPU block or AMD's AI block allow the closest to the metal programming with open documentation will win. None of this gimmicky OS specific or vendor specific AI "enhancements" will do. If both firms don't do this, they'll cede the entire market to Nvidia when they eventually enter the consumer market with silicon for Windows on Arm. Mark my words.
You are just talking shit. Closest to the Metal requires vendor lock in. That's literally what means. It tailors to one specific microarchitecture or ISA.
>Closest to the Metal requires vendor lock in. That's literally what means. No, it also means being able to implement back-ends for otherwise standardized interfaces and optimizing the same.
"Intel Meteor Lake is now expected to launch by the end of this year for laptops. The company hasn’t commented on reports that Meteor Lake-S was canceled, but it’s possible that such processors will be released. What may also be important to note is that VPU will be available on all Meteor Lake CPUs, including the supposed desktop “F” variants which usually lack integrated graphics support."
There's been some confusion over whether or not Meteor Lake will come to desktop. But getting the AI accelerators into everything needs to be done sooner rather than later, and is probably being demanded behind the scenes by the likes of Microsoft and Adobe.
Even if it's not as important as it is for mobile, it would still be nice to pick up any Intel CPU made in 2024+ and get modest AI acceleration (outperforming the iGPU if nothing else). Same goes for AMD. Putting an iGPU in every AM5 CPU released so far was a nice move that makes their CPUs more versatile. Hopefully they put XDNA in all Zen 5 desktop CPUs.
Until we know more about the architecture, meaningful comparisons are going to be hard to make. Useless ALUs are easy to throw at things, so FLOPS alone is not a very helpful metric.
from what I can see in the Linux patches (found'em via Phoronix), the API seems to be NN-only; looks a bit like a subset of what was available for the Keem Bay SoC's, with none of the ISP/CV APIs (available for M2/MX/KB) being made available.
so it doesn't look like just old IP on a new node (then again, the Movidius stuff were all SoC's, so maybe Intel just picked the parts they wanted); I'd love to see some benchmarks, but I guess we'll have to wait a bit for those.
it's a bit confusing, as the V in VPU apparently now stands "Versatile" as opposed to "Vision"; so while not mentioning "vision" matches the lack of a Computer Vision-related API, it's hard to imagine how having less makes it "versatile" :)
This seems misguided on Intel’s part. I’m a huge believer in AI and the benefits to society, putting dedicated silicon for AI inference into laptops and desktops seems like a waste of transistors. For data center, I get the use case for high power chips that are dedicated for that purpose. Likewise, in very power constrained environments (such as mobile phones operating <5w), I can see the benefit in power savings and battery longevity. But on the desktop I’m really struggling to find a use case where generic CPU cores running at 100w couldn’t get the job done ‘good enough’ for consumers. What applications are out there today where AI inference is being done locally on the desktop? I can’t think of a single example of where this would even be used.
Processing of audio and video. Look at NVIDIA Broadcast software. Great enhanced mic recording and background blur/replacement that exceed the capabilities of any "meeting" software I've used (like Teams or Cisco Meetings.) Surely, that's a limited field, but for people working from home a really nice enhancement.
Or for example, ask excel to build a worksheet based on a natural language description. No more use of the insane amount of buttons. It will allow users to be a lot more producfive.
I really hate to disagree with you on that because it's a great use, but I think cloud services are going to want their cut of that transaction. Yeah, Windows will record your voice, but it'll be sent to a MS cloud service, and the solution sent to you. For a price.
missing the point - to move some of these capabilities TO the edge devices. ALL edge devices. Keeping them in the DC means high latencies to access and means unrealistic response times. We expect our PCs to do the same things as our cell phones, but much better and faster since they are bigger and more powerful. The use cases are literally detailed in the slides above. Your avatar's face in a game can show real-time expressions from your camera. The possibilities are endless, once the hardware is in place people will find a cool way to use it. Also nobody wants their laptop running at 100W, the battery life is not acceptable. Dedicated accelerators are more power efficient
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.
17 Comments
Back to Article
GhostOfAnand - Monday, May 29, 2023 - link
Whichever of Intel's VPU block or AMD's AI block allow the closest to the metal programming with open documentation will win. None of this gimmicky OS specific or vendor specific AI "enhancements" will do. If both firms don't do this, they'll cede the entire market to Nvidia when they eventually enter the consumer market with silicon for Windows on Arm. Mark my words.dotjaz - Monday, May 29, 2023 - link
You are just talking shit. Closest to the Metal requires vendor lock in. That's literally what means. It tailors to one specific microarchitecture or ISA.Dolda2000 - Sunday, June 4, 2023 - link
>Closest to the Metal requires vendor lock in. That's literally what means.No, it also means being able to implement back-ends for otherwise standardized interfaces and optimizing the same.
TristanSDX - Monday, May 29, 2023 - link
So it is baseline feature but only for mobile in order to save energy. On desktop AI workload will be routed to GPU / CPUnandnandnand - Tuesday, May 30, 2023 - link
Not what I've heard:https://videocardz.com/newz/intel-shows-off-16-cor...
"Intel Meteor Lake is now expected to launch by the end of this year for laptops. The company hasn’t commented on reports that Meteor Lake-S was canceled, but it’s possible that such processors will be released. What may also be important to note is that VPU will be available on all Meteor Lake CPUs, including the supposed desktop “F” variants which usually lack integrated graphics support."
There's been some confusion over whether or not Meteor Lake will come to desktop. But getting the AI accelerators into everything needs to be done sooner rather than later, and is probably being demanded behind the scenes by the likes of Microsoft and Adobe.
Even if it's not as important as it is for mobile, it would still be nice to pick up any Intel CPU made in 2024+ and get modest AI acceleration (outperforming the iGPU if nothing else). Same goes for AMD. Putting an iGPU in every AM5 CPU released so far was a nice move that makes their CPUs more versatile. Hopefully they put XDNA in all Zen 5 desktop CPUs.
tipoo - Monday, May 29, 2023 - link
I gather this VPU ported to a new node and higher clocked might be around 4TOPS? The M2 is like 16? Or is this not comparable?Ryan Smith - Monday, May 29, 2023 - link
Until we know more about the architecture, meaningful comparisons are going to be hard to make. Useless ALUs are easy to throw at things, so FLOPS alone is not a very helpful metric.aiculedssul - Tuesday, May 30, 2023 - link
from what I can see in the Linux patches (found'em via Phoronix), the API seems to be NN-only; looks a bit like a subset of what was available for the Keem Bay SoC's, with none of the ISP/CV APIs (available for M2/MX/KB) being made available.so it doesn't look like just old IP on a new node (then again, the Movidius stuff were all SoC's, so maybe Intel just picked the parts they wanted); I'd love to see some benchmarks, but I guess we'll have to wait a bit for those.
it's a bit confusing, as the V in VPU apparently now stands "Versatile" as opposed to "Vision"; so while not mentioning "vision" matches the lack of a Computer Vision-related API, it's hard to imagine how having less makes it "versatile" :)
nandnandnand - Monday, May 29, 2023 - link
Stable Diffusion benchmarks now!noobmaster69 - Tuesday, May 30, 2023 - link
This seems misguided on Intel’s part. I’m a huge believer in AI and the benefits to society, putting dedicated silicon for AI inference into laptops and desktops seems like a waste of transistors. For data center, I get the use case for high power chips that are dedicated for that purpose. Likewise, in very power constrained environments (such as mobile phones operating <5w), I can see the benefit in power savings and battery longevity. But on the desktop I’m really struggling to find a use case where generic CPU cores running at 100w couldn’t get the job done ‘good enough’ for consumers. What applications are out there today where AI inference is being done locally on the desktop? I can’t think of a single example of where this would even be used.hansmuff - Wednesday, May 31, 2023 - link
Processing of audio and video. Look at NVIDIA Broadcast software. Great enhanced mic recording and background blur/replacement that exceed the capabilities of any "meeting" software I've used (like Teams or Cisco Meetings.)Surely, that's a limited field, but for people working from home a really nice enhancement.
jmlocatelli - Wednesday, May 31, 2023 - link
Or for example, ask excel to build a worksheet based on a natural language description. No more use of the insane amount of buttons. It will allow users to be a lot more producfive.hansmuff - Wednesday, May 31, 2023 - link
I really hate to disagree with you on that because it's a great use, but I think cloud services are going to want their cut of that transaction. Yeah, Windows will record your voice, but it'll be sent to a MS cloud service, and the solution sent to you. For a price.nandnandnand - Thursday, June 1, 2023 - link
Wait until we see how many square millimeters are devoted to the VPU. Damn near nothing, I expect.Laptops have power constraints just like phones.
As far as the applications go, they exist. More will exist after the inference hardware is widespread for a few years.
tipoo - Friday, June 2, 2023 - link
Where you don't want to peg your CPU cores to 100% when dedicated silicon could trivially handle it at low power?jjjag - Friday, June 2, 2023 - link
missing the point - to move some of these capabilities TO the edge devices. ALL edge devices. Keeping them in the DC means high latencies to access and means unrealistic response times. We expect our PCs to do the same things as our cell phones, but much better and faster since they are bigger and more powerful. The use cases are literally detailed in the slides above. Your avatar's face in a game can show real-time expressions from your camera. The possibilities are endless, once the hardware is in place people will find a cool way to use it. Also nobody wants their laptop running at 100W, the battery life is not acceptable. Dedicated accelerators are more power efficient