So i obviously don’t have a good understanding of this but how are they getting an advantage with using general purpose cores? where are they making up the difference in parallelism is it just easier for software?
Hi Ian, Im a retail real estate professional for the last 12 years, but I originally started out in IT, been following you since Anandtech days and was thrilled when you launched TTP. love the content your bringing out. I kinda want to get back to IT, I was thinking perhaps I can combine the retail/corporate side of things and be a front for an AI company that needs exposure in Dubai as a hub for the MENA region.
Sounds like a good plan, the MENA region has a lot of latent potential in conventional and renewable energy and datacenter business. Unfortunately it can be very risky because technology products can depreciate earlier than expected sometimes.
@@owlmostdead9492 I get that for 16 bit, but for 24 bit the snr is 144db, if you set the levels so that 130db is the Max spl, then there is no audible sound that will not be perfectly clear. There is truly no purpose for 32 bit in recording. 24 bit yes, 32 bit no.
@@varno -I can't seem to find the video I have in mind- If you record in 32bit you can recover from clipping, every time. With 24bit depth you can still cause unrecoverable clipping, that's fact and there is nothing to debate about this either. I'm only arguing that 32bit depth would have saved a lot of recordings on this channel. This is the video I had in mind ua-cam.com/video/5uyBYnoZB5k/v-deo.html only the first 20 seconds are relevant
@@owlmostdead9492 And I am saying that the problem is an improperly set Max spl, and not a bit depth issue. From a technical point of view, 32 bits is not nessicary for interview recording. As the average spl is unlikely to have been 100dBA, and levels can be set too hot for a 32bit converter just as easily as for a 24bit one. All 32 bit converters on the market have between 22 and 26 real bits at audio sample rates, the 32 bit number is pure marketing.
![[24] Jim Keller Interview 2: AI Hardware on PCIe Cards](http://i.ytimg.com/vi/fOUB73dZXEo/mqdefault.jpg)
![[24] Jim Keller Interview 2: AI Hardware on PCIe Cards](/img/tr.png)
Wow, his description of chiplet architecture was fascinating, in conveying the base idea
Great stuff. CPUs really are coming along with AI acceleration. Must be a fun position.
Love this kind of interviews. Very interesting!
I know it's tricky being on-site, but we gotta be able to get better audio for these interviews.
It’s up to them what level of technology they allow in their offices. He was lucky to get an interview.
Any audio of this renowned gentleman is a gift.
Outstanding interview!
As usual love the video even if i have no idea what is going on in detail lol
So i obviously don’t have a good understanding of this but how are they getting an advantage with using general purpose cores? where are they making up the difference in parallelism is it just easier for software?
Hi Ian,
Im a retail real estate professional for the last 12 years, but I originally started out in IT, been following you since Anandtech days and was thrilled when you launched TTP. love the content your bringing out.
I kinda want to get back to IT, I was thinking perhaps I can combine the retail/corporate side of things and be a front for an AI company that needs exposure in Dubai as a hub for the MENA region.
Sounds like a good plan, the MENA region has a lot of latent potential in conventional and renewable energy and datacenter business. Unfortunately it can be very risky because technology products can depreciate earlier than expected sometimes.
I hope Wei-Han, Jim and Chris Lattner have coffee on the regular 🙏
What does he mean, ‘8-wide’? Like, the Itanium chip, several streams of instructions at once?
It can be consider as a vector processor or SIMD
So a torrented security architecture with heuristics by default?
Anyone know what the unusual floating point format is, that he’s talking about? Sounds like fixed point.
Unusual floating point format means non standard floating point format (IEEE 754) not only the format also the handling of the operator
You should move to 32bit audio recording, you're always recording too hot. With 32bit depth, audio clipping is basically impossible.
You don't need 32 bit bit depth, but you do need to set the levels lower. Frankly even 24 bit audio has sufficient dynamic range.
@@varno For someone who doesn't know how to set levels 32bit is a must-have. I could record with 16bit but that's because I know how to set levels.
@@owlmostdead9492 I get that for 16 bit, but for 24 bit the snr is 144db, if you set the levels so that 130db is the Max spl, then there is no audible sound that will not be perfectly clear. There is truly no purpose for 32 bit in recording. 24 bit yes, 32 bit no.
@@varno
-I can't seem to find the video I have in mind-
If you record in 32bit you can recover from clipping, every time. With 24bit depth you can still cause unrecoverable clipping, that's fact and there is nothing to debate about this either. I'm only arguing that 32bit depth would have saved a lot of recordings on this channel.
This is the video I had in mind ua-cam.com/video/5uyBYnoZB5k/v-deo.html only the first 20 seconds are relevant
@@owlmostdead9492 And I am saying that the problem is an improperly set Max spl, and not a bit depth issue. From a technical point of view, 32 bits is not nessicary for interview recording. As the average spl is unlikely to have been 100dBA, and levels can be set too hot for a 32bit converter just as easily as for a 24bit one. All 32 bit converters on the market have between 22 and 26 real bits at audio sample rates, the 32 bit number is pure marketing.
Ian: *breathes*
Wei-han: Uh-huh. Right. Uh-huh.
Tents to rent.
Your voice has an echo