Anker's new 'Thus' chip claims 150x more AI compute for earbuds using compute-in-memory architecture -- interesting approach but where are the real-world benchmarks?
Posted by xxxxiceyyyyyy@reddit | hardware | View on Reddit | 31 comments
https://www.theverge.com/tech/916463/anker-thus-chip-announcement
azorsenpai@reddit
Not that stupid everyone is getting tires of the AI but remember there are audio models that are also a way to improve noise cancellation or on the fly eqing of the buds. That might be what makes their buds punch way above their price range in terms of anc
clon3man@reddit
BT headphone manufactures will really do EVERYTHING in the world, except the 2 things that make the most difference for calling
anyway, still curious to see what they've cooked up for their AI demo
Meal_Elegant@reddit
if they could do music upscaling on the fly it with this would be pretty neat. Its all bell and whistles until they improve the actual audio.
pixelpoet_nz@reddit
You really want the audio equivalent of DLSS sloshing directly in your ears? Rather than just normal ass 44.1khz 16bit in use since forever?
Every day we stray further...
Toastti@reddit
You can losslessly compress the music a larger amount and with the increased computing of the headset have enough performance to decompress it back to the original size without a delay.
nanonan@reddit
Is bandwidth even an issue? However fast you're compressing and decompressing, it will still introduce unwanted delays.
Meal_Elegant@reddit
I mean if they gonna put the engine between my ears they might as well start it.
AnechoidalChamber@reddit
I assume the usage would probably be some for of noise cancellation instead.
Seanspeed@reddit
Nobody reads articles anymore, do they?
AnechoidalChamber@reddit
Not when it's pay-walled, no they don't expect that they will.
Seanspeed@reddit
Ah yea, already forgot that. Read it this morning by just constantly reloading and reading a paragraph down at a time before the paywall notice popped up. lol
Rippthrough@reddit
And contextual voice command.
VenditatioDelendaEst@reddit
I don't think anything but wakeword detection should be run on the buds, because the device has much more power and silicon area budget.
But wakeword detection on the buds could be a great UX improvement, because bluetooth otherwise has to run in HSP (which has poor sound quality) all the time for two-way audio, or rely on the device mic for wakeword detection (which could be on the other side of a wall, in a strong wind, etc).
Seanspeed@reddit
Audio quality is not being limited by a standard digital format these days, especially in an IEM.
UpsetKoalaBear@reddit
From a cable manufacturer to a chip maker.
It’s kinda crazy seeing Anker’s rise.
lukfi89@reddit
Big W for Anker
whispous@reddit
I see what you did there
imaginary_num6er@reddit
They’re an “AI company” now
imKaku@reddit
I assume the answer is they are benchmarking something this product have been made for, the other wasn’t. If you have one piece of hardware that can do multiplication and the other can only do addition. Ofc you will be 150x quicker if you are trying to multiply by 150.
In case of products like these I only care about one thing, and that is audible performance. I’m still trying to find a piece of headphones (those are not made for headphones I believe) that I would rather use than my AirPod Pro 2s.
Seanspeed@reddit
Why? What's wrong with the Airpods?
imKaku@reddit
Nothing. (Good) Headphones are just better sounding. Small in-ears with Dac, bluetooth, noise cancellation chips, can only do so much.
Thankssomuchfort@reddit
Could check out the Sennheiser HDB 630.
It just came out but it seems to be considered the current best wireless headphones for those who care about audio quality
dbcoopernz@reddit
After the disasterous Momentum 4G headphones, I'm very reluctant to go back to Sennheiser.
imKaku@reddit
Yah I know of them. It’s on my consideration list. With auditable performance I also take anc into consideration. So waiting for a sale, just for my own sanity.
certainlystormy@reddit
oh shit. not the other person but ill look into this as well lol
Shoddy_Cup_9306@reddit
The NOR Flash angle actually makes sense for this use case -- faster reads than NAND, and if you are doing repeated inference on the same model weights, keeping compute co-located with storage is a legit efficiency win. What I want to know: what is the process node? Dresden has GF doing 22FDX and X-Fab doing various NOR processes. The answer changes how believable the power numbers are.
HatchetHand@reddit
Earbuds that use AI?
What in tarnation?
bubblesort33@reddit
So from like 1 megaflops to 150 megaflops or something? From very, very, very useless, to just very useless?
saltyboi6704@reddit
Wonder if these will compete with what Nordic like to cook up every couple years
OsgoodSlaughters@reddit
“AI compute for earbuds”
Please go jump off a bridge
Framed-Photo@reddit
I guess we'll get the real world benchmarks when products start releasing with this chip in it. I'm not expecting them to make a press release with benchmark charts like a PC hardware manufacturer, even if that would be pretty neat.