AI
Wednesday, August 15, 2007 4:35:23 PM
Did you all miss the 3 GHz demo ?
1) A design and process combo with a center frequency of
1.9 GHz will yield three sigma fast process tail end parts
of about 2.5 GHz. In large quantities across many wafer
batches roughly 1 in 400 parts will run at a production
quality 2.5 GHz.
2) A production 2.5 GHz processor has to be guaranteed
to run in all motherboards in all systems with different
chipsets, VRMs, DIMMs etc whether the system is in a hot
office in Phoenix or a cold basement in Anchorage. This
requires some guard banding to ensure all combinations
of parts and systems run reliably anywhere. But if you hand
match a specific part to a specific motherboard, system,
chipset, memory etc you can probably coax a 2.5 GHz
into reliable operation in a benign environment to say 2.6
GHz at stock voltage and cooling without much problem.
3) An advanced CMOS integrated circuit has to operate
under a certain combination of supply voltage and junction
temperature to achieve an aceptable level of reliability
and lifetime. However it can be operated at significantly
higher temperatures and voltages at the cost of expon-
entially reduced lifetime. This is how you do accelerated
aging studies to establish those acceptable levels in the
first place.
According to this IBM paper.
http://www.ibm.com/developerworks/library/pa-powerenv/
running a 970FX MPU in IBM's 90 nm SOI process at 1.3V
instead of 1.2 V reduces its expected lifetime from 100k
hours to 50k hours.
So how much could AMD raise the voltage (and junction
temperature, given standard cooling) of a demo sample
that only needed a lifetime of say 1000 hours? Or even
100 hours? A great deal. More than the 200 or 300 mV
boost that would likely raise our three sigma golden
Phenom sample in carefully matched motherboard and
system to a nice round 3.0 GHz to show off to credulous
analysts and fan-boys everywhere.
But AMD is still left trying to raise their 1.9 GHz center
frequency so they don't end up producing as many parts
binning out below 1.9 GHz as above.
1) A design and process combo with a center frequency of
1.9 GHz will yield three sigma fast process tail end parts
of about 2.5 GHz. In large quantities across many wafer
batches roughly 1 in 400 parts will run at a production
quality 2.5 GHz.
2) A production 2.5 GHz processor has to be guaranteed
to run in all motherboards in all systems with different
chipsets, VRMs, DIMMs etc whether the system is in a hot
office in Phoenix or a cold basement in Anchorage. This
requires some guard banding to ensure all combinations
of parts and systems run reliably anywhere. But if you hand
match a specific part to a specific motherboard, system,
chipset, memory etc you can probably coax a 2.5 GHz
into reliable operation in a benign environment to say 2.6
GHz at stock voltage and cooling without much problem.
3) An advanced CMOS integrated circuit has to operate
under a certain combination of supply voltage and junction
temperature to achieve an aceptable level of reliability
and lifetime. However it can be operated at significantly
higher temperatures and voltages at the cost of expon-
entially reduced lifetime. This is how you do accelerated
aging studies to establish those acceptable levels in the
first place.
According to this IBM paper.
http://www.ibm.com/developerworks/library/pa-powerenv/
running a 970FX MPU in IBM's 90 nm SOI process at 1.3V
instead of 1.2 V reduces its expected lifetime from 100k
hours to 50k hours.
So how much could AMD raise the voltage (and junction
temperature, given standard cooling) of a demo sample
that only needed a lifetime of say 1000 hours? Or even
100 hours? A great deal. More than the 200 or 300 mV
boost that would likely raise our three sigma golden
Phenom sample in carefully matched motherboard and
system to a nice round 3.0 GHz to show off to credulous
analysts and fan-boys everywhere.
But AMD is still left trying to raise their 1.9 GHz center
frequency so they don't end up producing as many parts
binning out below 1.9 GHz as above.
