Article: AMD's Mobile Strategy
By: anon (anon.delete@this.anon.com), December 21, 2011 5:21 am
Room: Moderated Discussions
anon (anon@anon.com) on 12/21/11 wrote:
---------------------------
>EduardoS (no@spam.com) on 12/20/11 wrote:
>---------------------------
>>Seni (seniike@hotmail.com) on 12/20/11 wrote:
>>---------------------------
>>>This already exists, but hybrid drives are neither as cheap, nor as fast as you
>>>expect. The Momentus XT, for example, is costing like $100 more than a normal drive,
>>>and performs more like a slightly faster normal drive, but is nowhere close to the speed of the SSDs.
>>
>>It costs so much because is the only in the market so Seagate put a higher margin...
>>
>>It doesn't actually costs (to Seagate) that much.
>>
>>>If you only have 1-2 NAND chips to keep it cheap, the cache miss rate will be too
>>>high and peformance will be mostly HD-like.
>>
>>Momentus XT isn't slow because of high miss rate, it is slow because only features
>>a read cache, nowadays PCs have a big amount of RAM wich is already used as a read
>>cache, writes (wich is the most important for the cache to speed up) aren't served
>>by this drive, that's what I said in my first post, the lack of a decent low-end hardware implementation.
>>
>>>And with a hybrid drive, not only do you have to pay extra on for the NAND chips,
>>>you also have to pay extra for the flash-facing side of the controller, which now
>>>needs its own cores and high-pincount bus. So there's an unavoidable floor cost
>>>there too. It's a bad deal to pay both floor costs instead of one.
>>
>>This cost is near zero per device, there is a quite big development costs in both
>>the firmware and controller but can be diluted by a high volume of sales.
>>
>
>I still think you overestimate these things. It will be very difficult to "solve the problem".
>
>Claims that random reads don't matter and random writes are the main thing that
>needs speeding up is no truer now than decades ago when log structured filesystems wanted to take over the world!
>
>For transaction procesing it may be true, but not for most client stuff. Actually
>random writes quite often benefit more from a cache, because your apps aren't waiting
>for the data. Exception in the case of data integrity operations, which can be quite
>important even on client, I won't deny, but well coded apps should often be able
>to avoid blocking all processing to wait for syncs.
>
Meant to say: writes can benefit more from caches in RAM.
---------------------------
>EduardoS (no@spam.com) on 12/20/11 wrote:
>---------------------------
>>Seni (seniike@hotmail.com) on 12/20/11 wrote:
>>---------------------------
>>>This already exists, but hybrid drives are neither as cheap, nor as fast as you
>>>expect. The Momentus XT, for example, is costing like $100 more than a normal drive,
>>>and performs more like a slightly faster normal drive, but is nowhere close to the speed of the SSDs.
>>
>>It costs so much because is the only in the market so Seagate put a higher margin...
>>
>>It doesn't actually costs (to Seagate) that much.
>>
>>>If you only have 1-2 NAND chips to keep it cheap, the cache miss rate will be too
>>>high and peformance will be mostly HD-like.
>>
>>Momentus XT isn't slow because of high miss rate, it is slow because only features
>>a read cache, nowadays PCs have a big amount of RAM wich is already used as a read
>>cache, writes (wich is the most important for the cache to speed up) aren't served
>>by this drive, that's what I said in my first post, the lack of a decent low-end hardware implementation.
>>
>>>And with a hybrid drive, not only do you have to pay extra on for the NAND chips,
>>>you also have to pay extra for the flash-facing side of the controller, which now
>>>needs its own cores and high-pincount bus. So there's an unavoidable floor cost
>>>there too. It's a bad deal to pay both floor costs instead of one.
>>
>>This cost is near zero per device, there is a quite big development costs in both
>>the firmware and controller but can be diluted by a high volume of sales.
>>
>
>I still think you overestimate these things. It will be very difficult to "solve the problem".
>
>Claims that random reads don't matter and random writes are the main thing that
>needs speeding up is no truer now than decades ago when log structured filesystems wanted to take over the world!
>
>For transaction procesing it may be true, but not for most client stuff. Actually
>random writes quite often benefit more from a cache, because your apps aren't waiting
>for the data. Exception in the case of data integrity operations, which can be quite
>important even on client, I won't deny, but well coded apps should often be able
>to avoid blocking all processing to wait for syncs.
>
Meant to say: writes can benefit more from caches in RAM.