Robert Wessel <[email protected]> writes:
> I'd be careful making that claim...
>
> A resource is:
>
> http://www.tech-news.com/publib/index.html
>
> Anyway, approximate price for a 485 MIPS 9021-9X2 in 2Q94 is $1.3M,
> for a 78,426 MIPS 2827-7A1 in 3Q12 $33.1M.
>
> In terms of MIPS growth, that's 161-fold in 18 years, or about 33% per
> year.
>
> In terms of cost-per-MIPS, that's 9.85-fold over that interval (taking
> into account about 55% inflation over that interval), or about 14% per
> year.
>
> Both are well below the slower (doubling every two years) version of
> Moore's Law.  OTOH, Moore was talking about the number of transistors
> on a chip, not system performance or cost/performance, although those
> are common interpretations.

industry standard for MIPS is dhrystones which is actually count of
iterations compared to benchmark on 370/158 assumed to be 1mip
processor.

recent mainframes
z900, 16 processors, 2.5BIPS (156MIPS/proc), Dec2000
z990, 32 processors, 9BIPS, (281MIPS/proc), 2003
z9, 54 processors, 18BIPS (333MIPS/proc), July2005
z10, 64 processors, 30BIPS (469MIPS/proc), Feb2008
z196, 80 processors, 50BIPS (625MIPS/proc), Jul2010
EC12, 101 processors, 75BIPS (743MIPS/proc), Aug2012

max configured z196 @50BIPS is $28M or $560k/BIPS

max configured ec12 @75BIPS is $33M or $440k/BIPS and a processor is 743
times that of 370/158 from 40yrs ago.

moore's law has been about circuit size decreasing ... early on
benefited by getting more circuits on chip and performance having to go
off-chip less (3033 started out as 168-3 logic mapped to 20% faster
chips but had 10times circuits/chip, some last minute logic revision to
better leverage going off-chip less got 3033 up to 50% faster than
168-3).

note 2010 IBM had base list price for E5-2600 blade of $1815
that have been benchmarked at 527BIPS ... or $3.44/BIPS ... E5-2600 had
two 8core chips ... for 16proc or 33BIPS/proc. there has been one turn
of the process crank since then and they are just starting to benchmark
the next turn of the process crank .... 
http://www.anandtech.com/show/8046/pricing-and-details-for-intels-devils-canyon-and-unlocked-pentium-leaked-online-i7-4790k-i5-4690k-g3258

note that manufactures are saying they now ship more server chips
directly to large cloud operators (than to brand name server vendors
like HP, DELL, IBM, etc) and the large cloud operators claim they
assemble their own servers for 1/3rd the price of brand name vendors
(say around $1/BIPS compared to around half-mil/BIPS for mainframe).

one of the issues is that memory access latency hasn't kept up with
increase in processor cycle performance ... when the memory access
latency is counted in number of processor cycles ... it is about the
same as 1960s disk access latency measured in number of 1960s 360
processor cycles.

The increasing mismatch of memory access latency started to result in
hardware processor design more akin to 360 software multitasking
(overlapping execution while waiting for disk access) ... starting at
least two decades ago in risc & i86 cisc processor designs
... out-of-order instruction execution, branch prediction, speculative
execution, hyperthreading (continue executing other instructions while
waiting for memory access).

The claim is that similar features start to appear in z196, accounting
for over half the increase in z196 per processor MIP rate (compared to
z10 per processor MIP rate, at least half the 156MIPS increase from 469
to 625) because of starting to introduce similar features ... with some
further introduction in ec12 (part of the 118MIPS increase from 625 to
743)

-- 
virtualization experience starting Jan1968, online at home since Mar1970

----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN

Reply via email to