I find it ironic that anyone would question the "images recognized per minute" method when even Apple used it, as seen in Daniel's article. This is a method to test the performance of hardware designed for machine learning.
As for a lack of understanding with regards to the hardware:
The Neural Processing Unit (NPU) in the Kirin 970 is using IP from Cambricon Technology (thanks to jjj for the tip, we confirmed it). In speaking with Eric Zhou, Platform Manager for HiSilicon, we learned that the licensing for the IP is different to the licensing agreements in place with, say ARM. Huawei uses ARM core licenses for their chips, which restricts what Huawei can change in the core design: essentially you pay to use ARM’s silicon floorplan / RTL and the option is only one of placement on the die (along with voltage/frequency). With Cambricon, the agreement around the NPU IP is a more of a joint collaboration – both sides helped progress the IP beyond the paper stage with updates and enhancements all the way to final 10nm TSMC silicon.
We learned that the IP is scalable, but at this time is only going to be limited to Huawei devices. The configuration of the NPU internally is based on multiple matrix multiply units, similar to that shown in Google’s TPU and NVIDIA’s Tensor core, found in Volta. In Google’s first TPU, designed for neural network training, there was a single 256x256 matrix multiply unit doing the heavy lifting. For the TPUv2, as detailed back at the Hot Chips conference a couple of weeks ago, Google has moved to dual 128x128 matrix multiply units. In NVIDIA’s biggest Volta chip, the V100, they have placed 640 tensor cores each capable of a 4x4 matrix multiply. The Kirin 970 TPU by contrast, as we were told, uses 3x3 matrix multiply units and a number of them, although that number was not provided.
At the keynote, and confirmed in our discussions after, Huawei stated that the API to use the NPU will be available for developers. The unit as a whole will support the TensorFlow and TensorFlow Lite frameworks, as well as Caffe and Caffe2. The NPU can be accessed via Huawei’s own Kirin AI API, or Android’s NN API, relying on Kirin’s AI Heterogeneous Resource Management tools to split the workloads between CPU, GPU, DSP and NPU. I suspect we’ll understand more about this nearer to the launch. Huawei did specifically state that this will be an ‘open architecture’, but failed to mention exactly what that meant in this context.
It is amazing what Apple and others have been able to do with camera phones, but as someone who used to make their living as a photographer during the film era I have a real problem with the idea of investing so much into something that will be traded or given away in about 2-3 years. Maybe it’s generational, but $1,000 for a throwaway point and shoot camera- even nice ones like the iPhone 8 and X are- is just not something I have any interest in.
I would love to see what Apple could do if they tried to make a serious stand alone camera that could work with and be controlled by a Mac or iOS device. If they built a camera on the Micro 4/3rds platform there would be a whole universe of great glass that could find new life.
I'm 47 and I grew up with film SLR cameras. It took many years for digital cameras to surpass the image quality of a good film camera but now even the cheaper ones leave our old film cameras in the dust. Now we are approaching the era when phone cameras take sharper and less noisy/grainy photos that our old film SLR cameras. Is $1000 a lot? Maybe compared to my pocket SONY camera but an iPhone is even more convenient as it fits in my jeans pocket. But you aren't just buying a camera, right? $1000 gets you a 4k video camera as well in your pocket. And a computer more powerful than desktop computers before 2013. And a photo viewer that contains every single photo and video I have ever taken since I started scanning photos or shooting with digital cameras. (I have well over 40,000 photos in my pocket) Add a GPS navigator for your car. I even use my iPhone for my motorcycle GPS. Add a date planner, a jukebox of every song I ever bought in my life, a personal assistant, an encyclopedia of every fact known to mankind, a pocket store to order anything I could ever want with free 2 day shipping, an instant messaging device and oh yeah, a phone and a video phone to call anyone on the entire planet for free. If we paid $10,000 that would still be cheaper that it should be. And, in 3 years you can sell that iPhone X for $400.
And add to all that, and referring to the last paragraphs of the article, all without needing to use your data for advertising, and sharing as little as possible with Apple, and when, in an anonymized way. That you can do all the above with privacy as a leading focus, is worth the money one puts in the devices. And no, one doesn't need 1000$, there are 3 other series of phones from Apple that do just the same and very capably too. And there is no need either to have GBs without end on a phone, if one has decent enough coverage and wi-fi access, to store everything in the Cloud for few bucks a month, and access in an intelligent way ("optimize storage" and iTunes Match) on a "as needed" basis.
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
I don't think so that Kirin 970 will have a better GPU than A11. If anything, Huawei's Kirin SoCs (launched in Oct) lag behind even Qualcomm's SoC of the same year (launched in Jan/Feb), leave alone Apple's latest SoCs (launched in Sep). They are catching up slowly, but I don't expect Kirin 970 to beat even SD 835 in GPU scores. Comparison with A11 - Forget about it, it is NOT worth it.
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
I don't think so that Kirin 970 will have a better GPU than A11. If anything, Huawei's Kirin SoCs (launched in Oct) lag behind even Qualcomm's SoC of the same year (launched in Jan/Feb), leave alone Apple's latest SoCs (launched in Sep). They are catching up slowly, but I don't expect Kirin 970 to beat even SD 835 in GPU scores. Comparison with A11 - Forget about it, it is NOT worth it.
It just doesn't matter.
I get that Huawei has to do this to stand out, and they are going to have to compete on SOC's with Qualcomm and Samsung in the Android OS world, but Apple is always going to have an advantage against an Android OEM in SOC's; that's just the result of Google's Android OS business model.
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
First, stop describing things as "your kirin" or "blowout that you obviously are hoping for". That's just wrong, I have no sworn allegiance or investment in these technologies, more so, an interest in new technology. I know you want to paint me as some "troll" or "fanboy", but again, that's just not the case.
Second, the CPU and GPU of the Kirin 970 are not going to beat Apple. We knew that from the moment the specifications for the Kirin 970 had been announced.
Finally, the only aspect of this SoC that makes it interesting is the NPU. Hence, I was curious as to what Apple's current performance level is with the A11. AI and machine learning are going to vastly change the way we use our devices, and this is something Huawei seems to have put a lot of effort into developing for their SoC. Tasks that would have once been taxing and/or power hungry on a CPU + GPU can be mitigated considerably using dedicated hardware. That's where my interest is.
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
First, stop describing things as "your kirin" or "blowout that you obviously are hoping for". That's just wrong, I have no sworn allegiance or investment in these technologies, more so, an interest in new technology. I know you want to paint me as some "troll" or "fanboy", but again, that's just not the case.
Second, the CPU and GPU of the Kirin 970 are not going to beat Apple. We knew that from the moment the specifications for the Kirin 970 had been announced.
Finally, the only aspect of this SoC that makes it interesting is the NPU. Hence, I was curious as to what Apple's current performance level is with the A11. AI and machine learning are going to vastly change the way we use our devices, and this is something Huawei seems to have put a lot of effort into developing for their SoC. Tasks that would have once been taxing and/or power hungry on a CPU + GPU can be mitigated considerably using dedicated hardware. That's where my interest is.
I'll take you at your word, but you really need to re-read what you have written in the past to get an idea on how you come across. You seem to be pushing an agenda biased against Apple, and on a Apple site no less.
You could have stated something like; "AI and machine learning are going to vastly change the way we use our devices, and this is something Apple and Huawei seem to have put a lot of effort into developing their SoC", and that would be as true as your own statement, but without the implied anti-Apple bias.
It is amazing what Apple and others have been able to do with camera phones, but as someone who used to make their living as a photographer during the film era I have a real problem with the idea of investing so much into something that will be traded or given away in about 2-3 years. Maybe it’s generational, but $1,000 for a throwaway point and shoot camera- even nice ones like the iPhone 8 and X are- is just not something I have any interest in.
I would love to see what Apple could do if they tried to make a serious stand alone camera that could work with and be controlled by a Mac or iOS device. If they built a camera on the Micro 4/3rds platform there would be a whole universe of great glass that could find new life.
The phones still have a value after 3 years and has 100 different other uses unlike the camera. A desktop computer of 2000 had very little different uses compared to.a current smart phone yet people paid as much as that. People don’t use inflation adjusted prices... when you factor that in, a cell phone is a miracle of usefulness
a cell phone is a miracle of usefulness
Yeah, and the best cell phone is the one on the camera you have with you...
That didn't answer my question. It also didn't discuss the Kirin 970, a chip that has a 1.92 TFLOPS neural processing unit.
There aren't any comparative benchmarks available today from independent third parties for the Kirin 970. Maybe there will be shortly after the Kirin 970 is released in a product. In the meantime, I find that Anandtech does the most thorough and descriptive benchmarking procedures, so I'll wait for their result; you should as well.
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
First, stop describing things as "your kirin" or "blowout that you obviously are hoping for". That's just wrong, I have no sworn allegiance or investment in these technologies, more so, an interest in new technology. I know you want to paint me as some "troll" or "fanboy", but again, that's just not the case.
Second, the CPU and GPU of the Kirin 970 are not going to beat Apple. We knew that from the moment the specifications for the Kirin 970 had been announced.
Finally, the only aspect of this SoC that makes it interesting is the NPU. Hence, I was curious as to what Apple's current performance level is with the A11. AI and machine learning are going to vastly change the way we use our devices, and this is something Huawei seems to have put a lot of effort into developing for their SoC. Tasks that would have once been taxing and/or power hungry on a CPU + GPU can be mitigated considerably using dedicated hardware. That's where my interest is.
I'll take you at your word, but you really need to re-read what you have written in the past to get an idea on how you come across. You seem to be pushing an agenda biased against Apple, and on a Apple site no less.
You could have stated something like; "AI and machine learning are going to vastly change the way we use our devices, and this is something Apple and Huawei seem to have put a lot of effort into developing their SoC", and that would be as true as your own statement, but without the implied anti-Apple bias.
Contrary, I was talking in context to the Kirin itself.
Apple has pushed for machine learning, maybe in this case not to the extent of Huawei, but the dedicated hardware in the A11 exists. I'm sure we'll see it developed more and more in future A-chips as it continues to grow in importance.
I often post more when I notice mis-information, unfortunately when it comes to things non-Apple, it's rampant around here. That, however, doesn't deter from my interest in Apple, or the articles and content on this site. The consequence, all too often, is an "us vs them" situation seems to arise. Too black and white of a view for my tastes.
Comments
As for a lack of understanding with regards to the hardware:
Source: Anandtech
Iit's almost a given that the A11 is going to be faster than any other SOC in single thread and multiple thread performance. That leaves you with the possibility of your Kirin 970 having better GPU performance, and a better benchmark score for "images recognized per minute" exceeding the A11.
The problem I have with the "images recognized per minute" benchmark, is that there is a 25% variation between what Apple tested, at 630 for the iPhone 7, and what Huawei tested for the iPhone 7 Plus, at 487; I'm thinking that they should test out the same. Given that, I'd rather let someone test both SOC's under the same conditions and procedures. Even then, I'm not seeing the blowout that you obviously are hoping for, now that Apple has two Bionic Cores for ML.
I do find it interesting that Huawei is willing to budget 5.5 million transistors and 100 mm2 of die for this. That makes it an expensive SOC, and actually larger than the A11's 90 mm2 and 4.3 m transistors, an SOC that should see about 225 million units shipped in FY 2018 alone. It appears on the surface to be more of a "brute force" approach to performance compared to Apple's extremely evolved A Series architecture.
I get that Huawei has to do this to stand out, and they are going to have to compete on SOC's with Qualcomm and Samsung in the Android OS world, but Apple is always going to have an advantage against an Android OEM in SOC's; that's just the result of Google's Android OS business model.
Second, the CPU and GPU of the Kirin 970 are not going to beat Apple. We knew that from the moment the specifications for the Kirin 970 had been announced.
Finally, the only aspect of this SoC that makes it interesting is the NPU. Hence, I was curious as to what Apple's current performance level is with the A11. AI and machine learning are going to vastly change the way we use our devices, and this is something Huawei seems to have put a lot of effort into developing for their SoC. Tasks that would have once been taxing and/or power hungry on a CPU + GPU can be mitigated considerably using dedicated hardware. That's where my interest is.
You could have stated something like; "AI and machine learning are going to vastly change the way we use our devices, and this is something Apple and Huawei seem to have put a lot of effort into developing their SoC", and that would be as true as your own statement, but without the implied anti-Apple bias.
a cell phone is a miracle of usefulness
Yeah, and the best cell phone is the one on the camera you have with you...
Apple has pushed for machine learning, maybe in this case not to the extent of Huawei, but the dedicated hardware in the A11 exists. I'm sure we'll see it developed more and more in future A-chips as it continues to grow in importance.
I often post more when I notice mis-information, unfortunately when it comes to things non-Apple, it's rampant around here. That, however, doesn't deter from my interest in Apple, or the articles and content on this site. The consequence, all too often, is an "us vs them" situation seems to arise. Too black and white of a view for my tastes.