https://www.thehindu.com/sci-tech/technology/new-nvidia-ai-chips-face-issue-with-overheating-servers-report/article68880681.ece

Nvidia's new Blackwell AI chips, which have already faced
delays, have encountered problems with accompanying servers
that overheat, causing some customers to worry they will not
have enough time to get new data centres up and running,
the Information reported on Sunday.

The Blackwell graphics processing units overheat when
connected together in the customised server racks the
company has designed, the report said, citing sources
familiar with the issue.

.. . .

Oh oh ... Nvidia kinda bet the bank on these
newest chips ......... and there are massive
downstream stock-market implications.

--
033-33

On Sun, 17 Nov 2024 23:11:25 -0500
"186282@ud0s4.net" <186283@ud0s4.net> wrote:

> Nvidia's new Blackwell AI chips, which have already faced
> delays, have encountered problems with accompanying servers
> that overheat, causing some customers to worry they will not
> have enough time to get new data centres up and running,
> the Information reported on Sunday.
>
> The Blackwell graphics processing units overheat when
> connected together in the customised server racks the
> company has designed, the report said, citing sources
> familiar with the issue.

Gee, if only they hadn't been too busy chasing the ELIZA bubble to
bother testing properly...! Hope you enjoyed that month on the Dow,
guys!

186282@ud0s4.net wrote:
> https://www.thehindu.com/sci-tech/technology/new-nvidia-ai-chips-face-issue-with-overheating-servers-report/article68880681.ece
>
>
> Nvidia's new Blackwell AI chips, which have already faced
> delays, have encountered problems with accompanying servers
> that overheat, causing some customers to worry they will not
> have enough time to get new data centres up and running,
> the Information reported on Sunday.
>
> The Blackwell graphics processing units overheat when
> connected together in the customised server racks the
> company has designed, the report said, citing sources
> familiar with the issue.
>
> . . .
>
>   Oh oh ... Nvidia kinda bet the bank on these
>   newest chips ......... and there are massive
>   downstream stock-market implications.
>

1000 watts TDP?
Seems just making more powerful chips by brute force has to end soon.

On Thu, 21 Nov 2024 07:35:47 +0800
Woozy Song <suzyw0ng@outlook.com> wrote:

> 1000 watts TDP?
> Seems just making more powerful chips by brute force has to end
> soon.

We can't stop *now,* we've poured too much money into it...!

On Thu, 21 Nov 2024 07:35:47 +0800, Woozy Song wrote:

> 1000 watts TDP?
> Seems just making more powerful chips by brute force has to end soon.

Faster == hotter. I remember a seminar about 40 years ago when CMOS was
coming into its own and touted as low power. The dirty little secret was
the power required to switch capacitive loads is a function of the voltage
and frequency so the dynamic power dissipation exceeded that of 74LS parts
when you drove them hard.

A lot has changed since then but I'm not waiting for a miracle.

On 11/20/24 6:35 PM, Woozy Song wrote:
> 186282@ud0s4.net wrote:
>> https://www.thehindu.com/sci-tech/technology/new-nvidia-ai-chips-face-issue-with-overheating-servers-report/article68880681.ece
>>
>>
>> Nvidia's new Blackwell AI chips, which have already faced
>> delays, have encountered problems with accompanying servers
>> that overheat, causing some customers to worry they will not
>> have enough time to get new data centres up and running,
>> the Information reported on Sunday.
>>
>> The Blackwell graphics processing units overheat when
>> connected together in the customised server racks the
>> company has designed, the report said, citing sources
>> familiar with the issue.
>>
>> . . .
>>
>>    Oh oh ... Nvidia kinda bet the bank on these
>>    newest chips ......... and there are massive
>>    downstream stock-market implications.
>>
>
> 1000 watts TDP?
>  Seems just making more powerful chips by brute force has to end soon.

Agreed ... some new underlying tech - like the switch
from bipolar transistors to CMOS - seems necessary
for the "next step".

The question of course is "WHAT tech" ?

Some fundamental re-thinking of digital and how
to do processors might also be useful. But again,
The Question ... still wondering if efficient
non-binary logic is possible these days but
haven't seen much on that.

On 11/20/24 11:45 PM, rbowman wrote:
> On Thu, 21 Nov 2024 07:35:47 +0800, Woozy Song wrote:
>
>> 1000 watts TDP?
>> Seems just making more powerful chips by brute force has to end soon.
>
> Faster == hotter. I remember a seminar about 40 years ago when CMOS was
> coming into its own and touted as low power. The dirty little secret was
> the power required to switch capacitive loads is a function of the voltage
> and frequency so the dynamic power dissipation exceeded that of 74LS parts
> when you drove them hard.
>
> A lot has changed since then but I'm not waiting for a miracle.

SOME "miracle" will be necessary ... we're about
at the limit.

On Wed, 20 Nov 2024, John Ames wrote:

> On Thu, 21 Nov 2024 07:35:47 +0800
> Woozy Song <suzyw0ng@outlook.com> wrote:
>
>> 1000 watts TDP?
>> Seems just making more powerful chips by brute force has to end
>> soon.
>
> We can't stop *now,* we've poured too much money into it...!
>
>

Anyone here shorting nvidia? Seems like a nice little bubble is building
up there.

On 2024-11-21 05:45, rbowman wrote:
> On Thu, 21 Nov 2024 07:35:47 +0800, Woozy Song wrote:
>
>> 1000 watts TDP?
>> Seems just making more powerful chips by brute force has to end soon.
>
> Faster == hotter. I remember a seminar about 40 years ago when CMOS was
> coming into its own and touted as low power. The dirty little secret was
> the power required to switch capacitive loads is a function of the voltage
> and frequency so the dynamic power dissipation exceeded that of 74LS parts
> when you drove them hard.

Still, they are fantastic for things like a battery powered clock. Or
processors that stay a lot of the time waiting.

>
> A lot has changed since then but I'm not waiting for a miracle.

--
Cheers, Carlos.

On Thu, 21 Nov 2024 02:06:42 -0500
"186282@ud0s4.net" <186283@ud0s4.net> wrote:

> > A lot has changed since then but I'm not waiting for a miracle.
>
> SOME "miracle" will be necessary ... we're about
> at the limit.

If the long-term viability of your industry requires the delivery of
miracles on schedule...it *miiight* be time to find a new industry.

On Thu, 21 Nov 2024 09:52:27 +0100, D wrote:

> On Wed, 20 Nov 2024, John Ames wrote:
>
>> On Thu, 21 Nov 2024 07:35:47 +0800 Woozy Song <suzyw0ng@outlook.com>
>> wrote:
>>
>>> 1000 watts TDP?
>>> Seems just making more powerful chips by brute force has to end
>>> soon.
>>
>> We can't stop *now,* we've poured too much money into it...!
>>
>>
>>
> Anyone here shorting nvidia? Seems like a nice little bubble is building
> up there.

I don't do stocks but that seems like a good bet. AMD and others are
coming from behind and may be able to supplant CUDA.

On Thu, 21 Nov 2024 12:04:12 +0100, Carlos E.R. wrote:

> On 2024-11-21 05:45, rbowman wrote:
>> On Thu, 21 Nov 2024 07:35:47 +0800, Woozy Song wrote:
>>
>>> 1000 watts TDP?
>>> Seems just making more powerful chips by brute force has to end
>>> soon.
>>
>> Faster == hotter. I remember a seminar about 40 years ago when CMOS was
>> coming into its own and touted as low power. The dirty little secret
>> was the power required to switch capacitive loads is a function of the
>> voltage and frequency so the dynamic power dissipation exceeded that of
>> 74LS parts when you drove them hard.
>
> Still, they are fantastic for things like a battery powered clock. Or
> processors that stay a lot of the time waiting.

Definitely. Like all design decisions they have their place. To really
date myself I sent my Osborne 1 back to the factory for the 100 column and
DS/DD floppy upgrade. The new video would work for a while and then die. I
isolated the problem chip with a can of cooler. It was CMOS and when I
replaced it with the equivalent 74LS problem solved.

On Thu, 21 Nov 2024, rbowman wrote:

> On Thu, 21 Nov 2024 09:52:27 +0100, D wrote:
>
>> On Wed, 20 Nov 2024, John Ames wrote:
>>
>>> On Thu, 21 Nov 2024 07:35:47 +0800 Woozy Song <suzyw0ng@outlook.com>
>>> wrote:
>>>
>>>> 1000 watts TDP?
>>>> Seems just making more powerful chips by brute force has to end
>>>> soon.
>>>
>>> We can't stop *now,* we've poured too much money into it...!
>>>
>>>
>>>
>> Anyone here shorting nvidia? Seems like a nice little bubble is building
>> up there.
>
> I don't do stocks but that seems like a good bet. AMD and others are
> coming from behind and may be able to supplant CUDA.
>

I think once the LLM:s are exhausted and we no longer see significant
gains, might be a good time to start.

On 2024-11-21 20:59, rbowman wrote:
> On Thu, 21 Nov 2024 12:04:12 +0100, Carlos E.R. wrote:
>
>> On 2024-11-21 05:45, rbowman wrote:
>>> On Thu, 21 Nov 2024 07:35:47 +0800, Woozy Song wrote:
>>>
>>>> 1000 watts TDP?
>>>> Seems just making more powerful chips by brute force has to end
>>>> soon.
>>>
>>> Faster == hotter. I remember a seminar about 40 years ago when CMOS was
>>> coming into its own and touted as low power. The dirty little secret
>>> was the power required to switch capacitive loads is a function of the
>>> voltage and frequency so the dynamic power dissipation exceeded that of
>>> 74LS parts when you drove them hard.
>>
>> Still, they are fantastic for things like a battery powered clock. Or
>> processors that stay a lot of the time waiting.
>
> Definitely. Like all design decisions they have their place. To really
> date myself I sent my Osborne 1 back to the factory for the 100 column and
> DS/DD floppy upgrade. The new video would work for a while and then die. I
> isolated the problem chip with a can of cooler. It was CMOS and when I
> replaced it with the equivalent 74LS problem solved.

That's a curious one. :-)

--
Cheers, Carlos.

On 27/11/2024 12:33, Carlos E.R. wrote:
> On 2024-11-21 20:59, rbowman wrote:

>> Definitely. Like all design decisions they have their place. To really
>> date myself I sent my Osborne 1 back to the factory for the 100 column
>> and
>> DS/DD floppy upgrade. The new video would work for a while and then
>> die. I
>> isolated the problem chip with a can of cooler. It was CMOS and when I
>> replaced it with the equivalent 74LS problem solved.
>
> That's a curious one. :-)
>

Actually it isn't. Ive been peripherally involved with hardware for
years, because that was where I started so any hardware problems came my
way.

Where I started in military spec hardware we endeavoured to guarantee
that if *every single chip* was 'edge of spec' between -40°C and +125°C,
the bloody thing still worked.
Imagine my surprise at encountering 'commercial design' where it was
sufficient if the thing worked between +10°C and +40°C and the customer
did the product testing, in terms of returned items.

But then a PC doesn't cost $25,000 does it?

--
WOKE is an acronym... Without Originality, Knowledge or Education.

On Wed, 27 Nov 2024 12:46:12 +0000, The Natural Philosopher wrote:

> Imagine my surprise at encountering 'commercial design' where it was
> sufficient if the thing worked between +10°C and +40°C and the customer
> did the product testing, in terms of returned items.

My youthful idealism took a hit with my engineering statistics course.
Much time was devoted to how many widgets have to be tested to insure only
X% are crap, X being determined by the cost of replacing defective widgets
versus the cost of doing it right.

On 27/11/2024 19:16, rbowman wrote:
> On Wed, 27 Nov 2024 12:46:12 +0000, The Natural Philosopher wrote:
>
>> Imagine my surprise at encountering 'commercial design' where it was
>> sufficient if the thing worked between +10°C and +40°C and the customer
>> did the product testing, in terms of returned items.
>
> My youthful idealism took a hit with my engineering statistics course.
> Much time was devoted to how many widgets have to be tested to insure only
> X% are crap, X being determined by the cost of replacing defective widgets
> versus the cost of doing it right.
>
It's a very important part of engineering. I once tested 1000
phototransistors to get an idea of the spreads.

It was a perfect bell curve with one end truncated. Except for two.
Those were obvious rejects that had slipped through.

Another time I got a panicky call from my boss. PA amplifiers were
popping like flies at a very important and well paid gig.

It turned out that on switch-on, the output hit the rails, and a relay
had been fitted to stop the loudspeakers blowing up. Shit design. But
anyway that put more voltage across the output transistors than they
were rated for. 'But it always worked before' I tested all the
transistors in stock. The whole batch were only marginally in spec.

(We left the amps powered up for 3 days solid to avoid switching them
on: once connected to a load the voltage sagged enough to be safe)

--
"First, find out who are the people you can not criticise. They are your
oppressors."
- George Orwell

On 2024-11-27, rbowman <bowman@montana.com> wrote:

> On Wed, 27 Nov 2024 12:46:12 +0000, The Natural Philosopher wrote:
>
>> Imagine my surprise at encountering 'commercial design' where it was
>> sufficient if the thing worked between +10°C and +40°C and the customer
>> did the product testing, in terms of returned items.
>
> My youthful idealism took a hit with my engineering statistics course.
> Much time was devoted to how many widgets have to be tested to insure only
> X% are crap, X being determined by the cost of replacing defective widgets
> versus the cost of doing it right.

I guess it's time to revive that joke about the American firm that
ordered a crate of widgets from a Japanese firm. When they opened
the crate, they found a small bag containing several widgets lying
on top of the rest of the shipment. When they asked the supplier
what the bag was for, the supplier replied, "Well, you specified a
2% defect rate, so for your convenience we packaged them separately."

--
/~\ Charlie Gibbs | Growth for the sake of
\ / <cgibbs@kltpzyxm.invalid> | growth is the ideology
X I'm really at ac.dekanfrus | of the cancer cell.
/ \ if you read it the right way. | -- Edward Abbey

On 11/21/24 3:38 PM, D wrote:
>
>
> On Thu, 21 Nov 2024, rbowman wrote:
>
>> On Thu, 21 Nov 2024 09:52:27 +0100, D wrote:
>>
>>> On Wed, 20 Nov 2024, John Ames wrote:
>>>
>>>> On Thu, 21 Nov 2024 07:35:47 +0800 Woozy Song <suzyw0ng@outlook.com>
>>>> wrote:
>>>>
>>>>> 1000 watts TDP?
>>>>>   Seems just making more powerful chips by brute force has to end
>>>>> soon.
>>>>
>>>> We can't stop *now,* we've poured too much money into it...!
>>>>
>>>>
>>>>
>>> Anyone here shorting nvidia? Seems like a nice little bubble is building
>>> up there.
>>
>> I don't do stocks but that seems like a good bet. AMD and others are
>> coming from behind and may be able to supplant CUDA.
>>
>
> I think once the LLM:s are exhausted and we no longer see significant
> gains, might be a good time to start.

LLMs can become very good - at least for many, human-replacement,
purposes. This is why Big Biz puts so much money into them, to
get rid of the annoying expensive humans.

Forgetting that disemployed humans can't BUY their stuff ...

(I suspect a "can't get there from here" point)

However I do think there's something beyond LLMs. Alas
I'm not 100% sure what that is. NN research continues,
but it doesn't yield as much as originally promised.
Might be even better ways of faking what NNs do without
trying to literally emulate the biological product -
at some point it's gonna be a kinda tight functional
equation. Actually real NNs - NOT just neurons/synapses
but lots of input from OTHER kinds of brain cells and
chemistry. Took a billion years of trial and error.

Proper 'consciousness' - don't know if it's the best
idea actually, won't be remotely as controllable as
optimists believe. However I suspect some kind of
'factal' infinite-regression/mirroring thing is involved.
It IS some kind of pattern/equation though, you can
smell it ... we see the paradigm not only in humans but
well down the tree. With varying degrees of sophistication
the "I AM" thing is widely seen.

TRUE "AI" ... we'd be building "aliens" - not at ALL
like we are.

On Tue, 3 Dec 2024, 186282@ud0s4.net wrote:

>> I think once the LLM:s are exhausted and we no longer see significant
>> gains, might be a good time to start.
>
>
> LLMs can become very good - at least for many, human-replacement,
> purposes. This is why Big Biz puts so much money into them, to
> get rid of the annoying expensive humans.

So far I've found the AI:s I've encountered at companies to be very bad and
annoying. I also read that more and more people support companies that provide
them with humans, over chat bots, and that the AI label is scaring customers
away.

> Forgetting that disemployed humans can't BUY their stuff ...

Like every technology shift, they will find other jobs, and our goods and
services will become cheaper thanks to automation and AI, which will benefit
everyone. There is no impending mass unemployment, regardless of how much
politicians would like for that to be true, so they could scare people into
accepting new "AI-taxes".

> (I suspect a "can't get there from here" point)
>
> However I do think there's something beyond LLMs. Alas

This is obvious. Of course, fast forward, 10-20 years, there will be more break
throughs. I believe LLMs have hit their limits and that we won't see _dramatic_
gains as long as LLMs are the underlying technology. We will still see gains,
but they are starting to taper off.

We will also see gains in pruning and making models smaller and more efficient,
while providing the same performance. This I also think is a given.

But LLMs will not lead us to consciousness. They lack will, volition and
motivations. These things, for me, are important ingredients if you want to
argue that you have created a consciousness.

> I'm not 100% sure what that is. NN research continues,
> but it doesn't yield as much as originally promised.
> Might be even better ways of faking what NNs do without
> trying to literally emulate the biological product -
> at some point it's gonna be a kinda tight functional
> equation. Actually real NNs - NOT just neurons/synapses
> but lots of input from OTHER kinds of brain cells and
> chemistry. Took a billion years of trial and error.
>
> Proper 'consciousness' - don't know if it's the best
> idea actually, won't be remotely as controllable as
> optimists believe. However I suspect some kind of
> 'factal' infinite-regression/mirroring thing is involved.
> It IS some kind of pattern/equation though, you can
> smell it ... we see the paradigm not only in humans but
> well down the tree. With varying degrees of sophistication
> the "I AM" thing is widely seen.
>
> TRUE "AI" ... we'd be building "aliens" - not at ALL
> like we are.

Maybe. We don't even know what we are aiming at, so could be that we create
something brand new and alien. Could also be that we run into some "laws of
thought" that hit once the complexity of the system reaches a certain level, and
we end up replicating ourselves.

I believe I will be alive when it happens and I'm looking forward to it! =)

On 03/12/2024 06:14, 186282@ud0s4.net wrote:
> It IS some kind of pattern/equation though, you can
>   smell it ... we see the paradigm not only in humans but
>   well down the tree. With varying degrees of sophistication
>   the "I AM" thing is widely seen.

If you think about it, consciousness has to be separate from the thing
it is conscious of.

This leads to the 'problem of consciousness' in quantum physics.

--
"When one man dies it's a tragedy. When thousands die it's statistics."

Josef Stalin

On 12/3/24 4:19 AM, D wrote:
>
>
> On Tue, 3 Dec 2024, 186282@ud0s4.net wrote:
>
>>> I think once the LLM:s are exhausted and we no longer see significant
>>> gains, might be a good time to start.
>>
>>
>>  LLMs can become very good - at least for many, human-replacement,
>>  purposes. This is why Big Biz puts so much money into them, to
>>  get rid of the annoying expensive humans.
>
> So far I've found the AI:s I've encountered at companies to be very bad and
> annoying. I also read that more and more people support companies that
> provide
> them with humans, over chat bots, and that the AI label is scaring
> customers
> away.

Just don't SAY they're "AI"s ... with Chat5 now hitting
the market the average consumer WON'T KNOW.

>>  Forgetting that disemployed humans can't BUY their stuff ...
>
> Like every technology shift, they will find other jobs, and our goods and
> services will become cheaper thanks to automation and AI, which will
> benefit
> everyone. There is no impending mass unemployment, regardless of how much
> politicians would like for that to be true, so they could scare people into
> accepting new "AI-taxes".

Alas, THIS time, I don't think it's all gonna work out.
LLMs or similar are being tuned to completely replace
the average Joe/Jane in essentially every venue they'd
be qualified to pursue. The Saboteurs were right - just
a little premature. WTF *would* Joe DO that's worth
money ? Jane can at least hang out on street corners ...

>>  (I suspect a "can't get there from here" point)
>>
>>  However I do think there's something beyond LLMs. Alas
>
> This is obvious. Of course, fast forward, 10-20 years, there will be
> more break
> throughs. I believe LLMs have hit their limits and that we won't see
> _dramatic_
> gains as long as LLMs are the underlying technology. We will still see
> gains,
> but they are starting to taper off.
>
> We will also see gains in pruning and making models smaller and more
> efficient,
> while providing the same performance. This I also think is a given.
>
> But LLMs will not lead us to consciousness. They lack will, volition and
> motivations. These things, for me, are important ingredients if you want to
> argue that you have created a consciousness.

Taken far enough, LLMs will become indistinguishable
from humans, or at least 'intelligent organisms'. Yea
they kinda fake it - but fake something WELL enough
and it's not really fake anymore ... just 'whatever
by an alternate means'.

>>  I'm not 100% sure what that is. NN research continues,
>>  but it doesn't yield as much as originally promised.
>>  Might be even better ways of faking what NNs do without
>>  trying to literally emulate the biological product -
>>  at some point it's gonna be a kinda tight functional
>>  equation. Actually real NNs - NOT just neurons/synapses
>>  but lots of input from OTHER kinds of brain cells and
>>  chemistry. Took a billion years of trial and error.
>>
>>  Proper 'consciousness' - don't know if it's the best
>>  idea actually, won't be remotely as controllable as
>>  optimists believe.  However I suspect some kind of
>>  'factal' infinite-regression/mirroring thing is involved.
>>  It IS some kind of pattern/equation though, you can
>>  smell it ... we see the paradigm not only in humans but
>>  well down the tree. With varying degrees of sophistication
>>  the "I AM" thing is widely seen.
>>
>>  TRUE "AI" ... we'd be building "aliens" - not at ALL
>>  like we are.
>
> Maybe. We don't even know what we are aiming at, so could be that we create
> something brand new and alien. Could also be that we run into some "laws of
> thought" that hit once the complexity of the system reaches a certain
> level, and
> we end up replicating ourselves.

They'll try - the 'new slaves', what everyone craves - but
there'd be too many diffs. We'll get 'aliens'.

> I believe I will be alive when it happens and I'm looking forward to it! =)

Being tech-based, ONCE "they" reach "the point" they
could self-evolve VERY VERY rapidly.

Our best outcome - they evolve WAY past the point where
they give a shit about humans or planets real quick and
zip off to their own little 'god dimension'.

On 12/3/24 5:57 AM, The Natural Philosopher wrote:
> On 03/12/2024 06:14, 186282@ud0s4.net wrote:
>> It IS some kind of pattern/equation though, you can
>>    smell it ... we see the paradigm not only in humans but
>>    well down the tree. With varying degrees of sophistication
>>    the "I AM" thing is widely seen.
>
> If you think about it, consciousness has to be separate from the thing
> it is conscious of.

Um ... I'm a materialist.

> This leads to the 'problem of consciousness'  in quantum physics.

The BEHAVIOR of a quantum-influenced NN system can be emulated.
It's all just math.

We'll figure it out eventually ... but may not LIKE
the results.