IBM reaches 350 GHz with silicon -germanium at room temperature...

[Jan Panteltje]

http://www.nytimes.com/2006/06/20/technology/20chip.html

This is silicon germanium tech, 500GHz when cooled with liquid helium
at 451 degrees below zero Fahrenheit.

They do not expect to make processors with this anytime soon, but it makes
me wonder if single core has more future then is supposed atm.

 

[George Macdonald]

http://www.nytimes.com/2006/06/20/technology/20chip.html

This is silicon germanium tech, 500GHz when cooled with liquid helium
at 451 degrees below zero Fahrenheit.

They do not expect to make processors with this anytime soon, but it makes
me wonder if single core has more future then is supposed atm.

I dunno about anything higher than dual but I expect single-core to be
dead, or at least relegated to budget systems, within a year. With the
workload presented by even home users now, yes that includes bloat and not
just from M$, dual core is a no-brainer and the effect is striking to
anybody.

 

[Jan Panteltje]

I dunno about anything higher than dual but I expect single-core to be
dead, or at least relegated to budget systems, within a year. With the
workload presented by even home users now, yes that includes bloat and not
just from M$, dual core is a no-brainer and the effect is striking to
anybody.

Oh well, in bussiness perhaps yes, I still run a single core server for web,
ftp, email, and video encoding.
Indeed is is MS bloat (and some Linux also) that replaces often the old
typewriter.
A conspiricy that sells hardware....

But if they make a 300GHZ processor, clearly that would be on par with
a 'hundred core 3GHz' almost.

So that would allow for some more years of bloat creation.

It is sort of hard to believe, 300GHZ in a procesor, must the memory then
be silicon-germanium too? PC traces will be too long, lightspeed is 3.3 nano
seconds per meter, slower in a conductor, 300 GHZ is about .001 ns, 1 pico
second period time, half a picosecond for half a clock, so the signal travels
only 1/3300 meter in that time, or 1/3 millimeter.... bit of a problem.

This math correct?

Optical chip connections, integrating memory on the same chip as the processor,
what could be done to solve the delay?

Think this is a Cray like problem.....

So we should not hold the breath and just simply go for the quad core when it is
available

 

[Evgenij Barsukov]

http://www.nytimes.com/2006/06/20/technology/20chip.html

This is silicon germanium tech, 500GHz when cooled with liquid helium
at 451 degrees below zero Fahrenheit.

They do not expect to make processors with this anytime soon, but it makes
me wonder if single core has more future then is supposed atm.

Strangely, the media is not talking at all about what kind of
process it is. 45nm? 30nm? What fab methods - far UV or water
immersion?

Regards,
Yevgen

 

[Del Cecchi]

Strangely, the media is not talking at all about what kind of
process it is. 45nm? 30nm? What fab methods - far UV or water
immersion?

Regards,
Yevgen

The paper will be published. As I recall from looking at advance copy,
the design rules are not especially aggressive, 90 or 130. The vertical
profile is an enhanced SiGe BiCMOS beyond the current offering from IBM.

 

[Alexander Grigoriev]

It's NOT clock of a working CPU. It's either 0dB frequency, where the
transistor doesn't amplify power anymore, or ring oscillator frequency
(which depends on a single gate delay).

I'd expect regular n-channel transistors in regular Pentium have cutoff
frequency at about 50-100 GHz. So 350 GHz is just next generation, nothing
extraordinary.

 

[Del Cecchi]

It's NOT clock of a working CPU. It's either 0dB frequency, where the
transistor doesn't amplify power anymore, or ring oscillator frequency
(which depends on a single gate delay).

I'd expect regular n-channel transistors in regular Pentium have cutoff
frequency at about 50-100 GHz. So 350 GHz is just next generation, nothing
extraordinary.

SiGe HBT bipolars have many advantages over FETs for some applications,
particularily communications oriented ones. And this represents about a
doubling of the previous record.

The paper doesn't seem to be on the IEEE database yet. But the authors
have many other papers discussing SiGe performance, fabrication, and
advantages.

It does seem to not be the sort of thing one would use to make a
microprocessor. Didn't BIT try that years ago?

 

[Johannes]

Oh well, in bussiness perhaps yes, I still run a single core server for web,
ftp, email, and video encoding.
Indeed is is MS bloat (and some Linux also) that replaces often the old
typewriter.
A conspiricy that sells hardware....

But if they make a 300GHZ processor, clearly that would be on par with
a 'hundred core 3GHz' almost.

So that would allow for some more years of bloat creation.

It is sort of hard to believe, 300GHZ in a procesor, must the memory then
be silicon-germanium too? PC traces will be too long, lightspeed is 3.3 nano
seconds per meter, slower in a conductor, 300 GHZ is about .001 ns, 1 pico
second period time, half a picosecond for half a clock, so the signal travels
only 1/3300 meter in that time, or 1/3 millimeter.... bit of a problem.

This math correct?

My starting point is that I know that the wavelength of 300Mhz is 1m. So the
wavelength of 300GHz must be 1mm. Signals coming from different processing
units must meet within maybe 1/10 of a wavelength, so the chip estate must me
incredibly small.

Optical chip connections, integrating memory on the same chip as the processor,
what could be done to solve the delay?

To connect to the motherboard, there would have to be several intermediate steps
of bus multipliers.

Think this is a Cray like problem.....

Yes, equalizing the signal travel distance.

So we should not hold the breath and just simply go for the quad core when it is
available

And we haven't reached 4Ghz yes, not officially anyway...

 

[Del Cecchi]

My starting point is that I know that the wavelength of 300Mhz is 1m. So the
wavelength of 300GHz must be 1mm. Signals coming from different processing
units must meet within maybe 1/10 of a wavelength, so the chip estate must me
incredibly small.




To connect to the motherboard, there would have to be several intermediate steps
of bus multipliers.




Yes, equalizing the signal travel distance.




And we haven't reached 4Ghz yes, not officially anyway...

Not for processor clocks. But there are 11 GHz serial interconnects and
wireless and other applications of high frequency circuits. Not
everything is a processor ya know.

 

[Jan Panteltje]

Not for processor clocks. But there are 11 GHz serial interconnects and
wireless and other applications of high frequency circuits. Not
everything is a processor ya know.

But this sort of begs the question, is the reason we are stuck below 4GHz for
processors only one of heat, or DOES the travel time of the signal become an
issue on the chip?

So say 10GHz = 100 ps, half a clock = 50pS, you need some setup times,
(say we are addressing CPU registers)...
1 ns = 33 cm, so 50 pS = 33/20 = 1 cm.
I see a barrier here that is not so easily taken.

This right?

I agree that for RF stuff these new transistors are great.
Several hundred GHz communication links can carry much more bandwidth,
but it will not go through anything! not through walls.

 

[Del Cecchi]

But this sort of begs the question, is the reason we are stuck below 4GHz for
processors only one of heat, or DOES the travel time of the signal become an
issue on the chip?

So say 10GHz = 100 ps, half a clock = 50pS, you need some setup times,
(say we are addressing CPU registers)...
1 ns = 33 cm, so 50 pS = 33/20 = 1 cm.
I see a barrier here that is not so easily taken.

This right?

I agree that for RF stuff these new transistors are great.
Several hundred GHz communication links can carry much more bandwidth,
but it will not go through anything! not through walls.

They support ECL gates with very short delays. The processor issue is a
combination of things. Power, travel time of signals, jitter, skew,
gate delay, all enter in. And chip wires are pretty high resistance, so
delay is proportional to RC which increases as the square of the length.

And not through walls can be a benefit.