[Info-vax] Chinese Alpha?

[Arne Vajhøj]

On 5/3/2012 10:23 PM, David Froble wrote:
> Arne Vajhøj wrote:
>> On 5/3/2012 12:16 AM, David Froble wrote:
>>> Arne Vajhøj wrote:
>>>> On 5/2/2012 12:35 AM, David Froble wrote:
>>>>> When it was being competitively developed, nothing was faster than
>>>>> Alpha.
>>>>>
>>>>> As has been mentioned elsewhere, some people will pay for performance.
>>>>>
>>>>> To get to your question, you need to ask, "why was it fast?" Perhaps
>>>>> because the design was better than the competition? If so, and if
>>>>> there
>>>>> has not been any new designs that might be better, then a
>>>>> competitively
>>>>> developed Alpha might still be the fastest single processor available.
>>>>> When talking multiple cores, the "glue" on the chip, memory
>>>>> controllers,
>>>>> interprocessor communications, and such in EV7 and EV7z, even with the
>>>>> larger die size was doing things that out performed competing CPUs
>>>>> that
>>>>> had continued development and smaller die size.
>>>>>
>>>>> So, if the Alpha could be successfully shrunk down to say 35 nm, or
>>>>> even
>>>>> 22 nm which Intel has just released, and with a large on chip cache as
>>>>> the IA-64 has, perhaps it might still out-perform anything else
>>>>> available.
>>>>>
>>>>> Now, that is a mighty big "if", and what it would cost I have no idea,
>>>>> other than it wouldn't be cheap. So, commercially, might it be
>>>>> viable? I
>>>>> really have no idea. It would depend upon how many might pay for such
>>>>> performance.
>>>>
>>>> A shrink to current technology could make it run at 3-4 GHz.
>>>>
>>>> Or about 3-5 times faster than back then.
>>>>
>>>> That is not enough.
>>>>
>>>> But modify it from 1 to 8 cores per socket.
>>>>
>>>> Then it will be 24-40 times faster than back then
>>>> (for multi thread/process usage).
>>>>
>>>> It is starting to look like something.
>>>>
>>>> Double all 3 levels of cache and do some optimizations
>>>> based on what is learned since then to double performance.
>>>>
>>>> Then it will be 48-80 times faster than back then.
>>>>
>>>> But the first item require a B$ investment in production
>>>> facility.
>>>>
>>>> And the two last items require significant design work that
>>>> will take a lot of time (years).
>>>
>>> The real problem is, the cost per chip must pay for the development and
>>> production. Here is where numbers matter. If you cannot sell enough
>>> chips, then the cost will be too high.
>>>
>>> I was very disappointed to see the end of Alpha. But I'm not blind to
>>> the economic realities. The CPUs going into tablets, phones, and such
>>> are a much better economic investment than what might power a much
>>> smaller number of very powerful servers. Such is reality.
>>
>> The fixed costs is very large and increasing relative to the
>> variable cost for CPU's.
>>
>> This will inevitable lead to fewer CPU producers.
>>
>> Intel and AMD will survive because they sell desktop
>> CPU's in the 2 or 3 digit millions per design.
>>
>> I do not have much confidence in Power or SPARC long term.
>
> Do be careful when betting against IBM.
>
> Got some IBM stock that was purchased at $104 and 2 days ago it was at $206
>
> The thing with IBM is that they are not just a semiconductor mfg. They
> can afford to take a loss in one department, if that allows other
> departments to do well. If Power gives them an edge in sales, perhaps
> that's better for the company than just being another "me too".

IBM is a business.

If they make money on Power CPU's then they will make them. If not
then they will drop them.

IBM will still make a ton of money on all the other things
they do.

Arne