@Zebras
Probably not, unless reviews are oustanding. I may wait for a battery version.
Peter
2 Likes
@ziploc
Maybe they’ll produce a Roger Mic with the Sphere chip in? If it’s as good as they claim, it may be a big seller.
Peter
3 Likes
From what I can see, it should be able to be added as a selectable additional programme.
Peter
Edit: or maybe not if it’s “attached” to the Speech in Loud Noise programme
I’m trying to understand what he meant by this quote
For some reason, these products targeted for prescription by medical professionals continue to be a complicated maze of too many options and multiple layers of different technologies that, in the end, do not benefit the people who need hearing-aids.
Is he being critical of the complex landscape of hearing aid options, as some of the new Phonak products launched this week have AI based features and some don’t? In other words, are there too many hearing aid models and this complexity doesn’t benefit the people who need hearing aids?
Or
Is he saying the technology itself provided in modern hearing aids are not beneficial to people who need hearing aids?
If the former, he’s not wrong, however how much can a guy who hyphenates the term hearing-aids really know about them? 
1 Like
Anybody know process size they use in their CPUs? I know the Sword processor from a ways back was groundbreaking with a 28nm processor when most competitors were using 65nm. I’m guessing this is notably smaller.
1 Like
Probably, the laptop did not have the specialized NPU (such as DEEPSONIC in Audeo Infinio Sphere), and the AI processing was done by a less dedicated GPU, which requires more power (I assume; I have not read about it yet).
EDIT:
Some specification of laptop here:
The CPU processor is from Q2 2018:
https://www.intel.com/content/www/us/en/products/sku/134876/intel-core-i58300h-processor-8m-cache-up-to-4-00-ghz/specifications.html
The graphic card with GPU is from Q4 2016:
Yes, I read the line you cited in the Nature paper by Audatic and its research collaborators. Earlier in this thread where it was questioned whether HAs have enough computational power, I mentioned NPUs vs GPUs as you rightly point out could make DNN calculations much more power-efficient: Phonak Audéo Sphere - #71 by jim_lewis.
But just for laughs, I guesstimated the TFLOP processing power and TFLOP/TDP of the Sphere vs. the NVIDIA GTX 1050. TFLOP is teraflop, a trillion floating point operations per second. TDP is thermal design power, usually given in watts.
An Internet search shows the GTX 1050 can perform 1.6 TFLOPs and consumes 75 watts. Thus, its TFLOP/TDP is 1.6/75 = 0.021 TFLOP/Watt (whaddaya expect for 2016!).
Page 4 of the Sphere white paper cited by @bigaltavista (Phonak Audéo Sphere - #236 by bigaltavista) says the DeepSonic denoising chip performs 7,700 million operations per second (7.7 billion ops per second). That’s only 0.0077 TFLOP. We don’t know if those are equivalent floating point instructions to those in the GTX 1050, but, Hey, this is just guesstimation. Let’s say they are. A Li-ion battery is a 3.7V supply. Most HAs consume 1 to 2 mA. Let’s say the Sphere with two processors and an 18-hour battery life consumes 4 mA. 3.7V x 0.004A is 0.015W. 0.0077/0.015 is 0.51 TFLOP/Watt.
Perhaps there will be even better performance in future, smaller process size design versions of the DeepSonic chip. 0.51 TFLOP/Watt is nothing to write home about. The Raspberry Pi 5 AI kit has a dedicated NPU on an M.2 HAT+ board. It can do 13 TFLOPS. It’s TDP is < 2W. So, 13/2 = 6.5 TFLOP/Watt. And I’m sure it’s not even a state-of-the-art NPU.
So, maybe if Sonova starts selling lots of Sennheiser earbuds with the DeepSonic chip inside, the economics of scale would justify going to a much smaller process size, which might increase the TFLOPS the processor is capable of while reducing the TDP (as has happened with smartphones, tablets, computers, etc).
| Item | TFLOPS | TDP | TFLOPS/TDP*** |
|---|---|---|---|
| NVIDIA GTX 1050 | 1.6 | 75 | 0.021 |
| Audeo Sphere | 0.0077 | 0.015 | 0.51 |
| Raspberry Pi 5 | 13 | 2 | 6.5 |
| AI Board |
Edit_Update: ***Note, I am confusing TOPS and TFLOPS. TFLOPS is tera floating point operations per second. TOPS is tera operations per second, usually integer (uses a smaller instruction bit size). GPUs are typically scored in TFLOPS. NPUs in TOPS. So, I’m mixing apples and oranges a bit. So, the guesstimate for the Audeo Sphere DeepSonic chip is probably most comparable to the TOPS for the Raspberry Pi 5 Hailo AI (NPU) chip.
The NVIDIA GTX 1050, of 2016 vintage, is probably so old that no one is really bandying about a TOPS rating for it. But NVIDIA’s RTX 4090, one of its most powerful recent GPUs, has an 83 TFLOPS rating (floating point-based) and a 1300 TOPS rating (integer-based). The TDP of the RTX 4090 is 450W. So, 1300/450= 2.9 TOPS/Watt. That’s why folks want NPUs for TOPS without burning a lot of watts. 
2 Likes
I think that will be a long long wait
4 Likes
a few months, hopefully much less
Audeo Lumity was released about 2 years ago. The L90-312 is only about 6 months old.
Maybe? Look on the bright side?
WH
2 Likes
Anyone know how long before these will be available in the uk ? Just curious as from what i see they are not showing up on the phonak uk website yet
Yes, but I understood that they initially didn’t plan to bring a disposable batteries device, now they likely do. So I assume it will be sooner
1 Like
And if they are rechargeable, people want the possibility that these rechargeable batteries can be practically replaced with a new battery.
2 Likes
If phonak cannot meet these expectations. Then it’s great that there is competition. This year, Starkey released genesis hearing aids, which provide up to 50 hours of operation. When I buy hearing aids, it will be one of the characteristics that will be important when choosing.
1 Like
I think that if you have new hearing aids, you should skip these because. Let’s remember how it was with the Phonak Marvel series, it was the first phonak hearing aids with built-in bluetooth, a revolution compared to the old ones, but it also had some shortcomings that they fixed with the Paradise version.
I would like to see a comparison test of how advanced the phonak dnn is compared to Starkey and Oticon. Oticon has been at it longer and they have polished it better in the Intent version…
1 Like
End of August for the RIC models. Early September for the custom Virto:
3 Likes
Those are not high [Signia advertises 12,000Hz] and they’re not XMEMS. Firstly, it’s too early for that and secondly XMEMS have different response curve.
1 Like
An interesting comparison of the computing power of an NVIDIA GTX and the Sphere… The NVIDIA GPU in my PC is the size (and shape) of a healthy brick, with two dedicated cooling fans, burning up most of the wattage my PC uses. I don’t expect my HAs to engage in bitcoin mining in a dull moment anytime soon…
7 Likes
As I noted at the end of my previous post, I erred in only coming up with a TFLOPS index for the graphical computing power of the GTX 1050, not a TOPS index figure, which is more indicative of AI capability. I asked Google Gemini what the potential TOPS figure for a GTX 1050 would be. The interesting answer is that the GTX 1050 is not a very AI-capable NVIDIA GPU. It doesn’t have the tensor cores that newer NVIDIA GPUs, such as the different RTX series, do. It said that the GTX 1050 would probably have, at most, a TOPS figure in the low to mid-single digit range, i.e., it would be approximately equal to its TFLOPS rating.
So it’s interesting that Audatic and its research collaborators actually designed a neural network for denoising using a GPU that was pretty suboptimal for AI computations, and the rightmost column in my previous post, which really should be labeled TOPS/TDP is not as far off as I thought it would be for mistakenly using the GTX 1050 TFLOPS index instead of a TOPS index.
Gemini also cautioned me, though, that it’s difficult to compare TOPS figures across different architectures:
Different TOPS metrics: There are various ways to calculate TOPS, depending on the specific operations considered (FP32, FP16, INT8, etc.), memory bandwidth, and other factors. This makes comparisons between different architectures difficult.
So, whatever I wrote in my previous post really was a “guesstimate!”
1 Like
I think Phonak will likely release the infinio non-sphere devices with a traditional battery. I’m less sure about them releasing the sphere chip with a traditional battery any time soon. It’s an engineering problem that they haven’t yet solved. So, how long will it take them to figure that out, IF they figure it out, and then from there how long is the transition to manufacturing it? Not a few months.
4 Likes