With Sandybridge Intel created a new set of instructions to replace/supplement SSE. There are some nice things about them that ought to mean that they should have higher performance. However, no existing software uses them so the AVX stuff isn't being used right now. In N months when software has code execution paths using AVX we might see these guys pull ahead of Nehalem processors. AMD will have its own processors using AVX coming out soon too.
On thing I didn't realize immediately. Since the larger registers are program state that has to be stored and restored between thread switches you need operating system support before you can enable AVX. Apparently this is coming to Windows in service pack 1 of Windows 7.
The AVX extension adds a set of new instructions that can operate on 256 bits of data at once (in the form of 4 64-bit floats, 8 32-bit floats, or various other data types), and the new registers to fit them in.
These new instructions promise to significantly increase throughput for a limited set of operations -- but to do that, the software has to be compiled to use them, and often entirely rewritten to use the wider vectors. Few will bother to do that until the new instructions have significant market penetration -- which will take years.
I am not quite as optimistic as Symmetry is about the usefulness of these latest extensions to x86 -- while many fringe tasks will benefit a lot, most of what you do on a computer just isn't limited by the width of the vectors or the amount of SIMD operations you can do per clock. IMHO, the insanely fast caches are a much more significant improvement.
They also allow the CPU register context to be saved/loaded with a single instruction as well, User-Mode Scheduler in Win7 SP1 uses it (the normal scheduler too afaik).
Desktop: Pentium 4 3.20GHz, 2 GB RAM (circa 2004)
Laptop: Pentium 4 3.0GHz, 1 GB RAM (circa 1Q2005) {It's a Sager (Clevo rebrand) and has been an absolute tank, but the screen is starting to dim)
Both run Ubuntu pretty well, for what I usually hack on (python/django) and the desktop used to run the Source engine games pretty well (under XP). I've recently started learning Android and Eclipse (and especially the AVD emulator) and both of those machines just thrash.
I'm looking to build a new box and pick up a Thinkpad when they update. Going to be an expensive year.
ha, my Sager 8xxx was a great laptop and lasted quite a while. Would you like to buy it for parts? When it died I switched to a thinkpad t42 from 2004 that I bought for $100.
It's amazing that you can get their top-of-the-line processor for only $294. The best Pentium 4 cost $637 when it launched in June 2003, which is $757 in today's dollars. This means the best Intel processor money can buy is 2.5x cheaper today as compared with only 7.5 ago.
This is only an artifact of market segmentation and release schedules.
LGA1155 is their budget line. 2600K is not the best processor they can make -- it is the best processor they can make give very stringent whole platform cost limits. In addition to having the desktop budget line, they also have a desktop high-end line. Last generation, these were the LGA1156 and LGA1366 respectively. The processors they just released are the budget line for this gen, and the high-end line fits a socket called LGA2011, which will be released later, and on which the processor cost will range from 300$ to about 1000$.
In every generation, new Intel processors gain roughly as much more power as is the difference between the budget and high-end product lines, so it's normal for todays budget processors to win over yesterday's high end. Usually Intel releases the high-end first for better margins and to avoid this kind of confusion -- this time, they started on the budget end, probably to pummel AMD, who previously only competed on the budget segment and now cannot really compete much at all.
Intel adding integrated graphics to their CPU's is basically spending ~20% of their chip on something that most mid to high end consumers are not going to use, and would be more cheaply produced as a separate chip on the low end. Intel's basic assumption is AMD is so far behind they can't exploit it, and they can exploit their monopoly to crush the ultra low end GPU market. But, it's still risky.
IMO Intel focusing on the low to mid range of this release cycle is vital or they are going to be eaten.
How is this more cheaply produced as a separate chip? If that's the case then AMD could exploit it. I thought the point was that producing this would be cheaper, passing on the cost savings to the consumer, while at the same time offering better perf than low-end GPUs.
Everything I've seen seems to indicate that this is true. Did I miss something?
The larger the chip the fewer chips you get from the same process. The larger the chip the higher risk a failure. The larger the number of transistors the lower your heat budget for the remaining chip.
But, far more importantly, producing a 32nm graphics chip is more expensive (for now) than a 45nm chip using the same number of transistors. Intel has a huge process advantage and is often on the lookout for useful things to do with their huge transistor budget, but integrated graphics has huge downsides, most notably memory bandwidth issues.
Regarding BW, the GPU now has access to onchip (L3) cache. Since low-end GPUs use shared memory anyways, this actually saves bandwidth to main memory, and also since it needn't go get main memory, also saves power (from the power hungry GPU).
I get that you decrease yield, but w/o knowing the numbers, it's hard to really see if its much impact on the bottomline. I suspect its a bigger win than loss (not hard to speculate, since Intel apparently had made the same bet).
8MB of Cache has far less value when dealing with modern 3D rendering than you might think. Granted, building HW to deal with HD Video decoding is a huge win from an energy use perspective, but 3D is a different beast that is not vary cache friendly.
To give some idea of cache's relative value a GT110 is a 3 billion transistor chip with less than 1mb of cache. A QX9650 is a 0.82 billion transistor chip with 12MB of cache.
PS: At a minimum it's a 1 - (1/1.2) = 17% loss but adding in bad chips it's probably closer to 25% until they start selling chips without video capabilities. Also, I am not saying it's a bad long term move, just that in the short term it's risky.
Mid to high end mobile devices can use these chips with a switchable external GPU that can be turned on for graphics-intensive applications and off when it's not needed in order to improve power consumption - Apple is doing this with their MB Pros, and as far as I know it has been an effective configuration.
I can imagine a similar setup in desktop machines to reduce their power consumption.
From the reading I've done lately, switchable graphics are just now coming into their own. Nvidia Optimus in notebooks (only) now apparently does very nice automatic seamless switching between CPU and dedicated GPU. There are some good articles out there benchmarking the much improved battery life. When the next round of SSD are released, one will be able to buy a pretty sweet laptop: powerful when you need it, but super efficient when you don't, and fast always.
I'm pretty certain I read (an article on anantech?) that graphics switching is explicitly not going to be supported on desktops until the next release of intel processors.
"Quick Sync with a Discrete GPU There’s just one hangup to all of this Quick Sync greatness: it only works if the processor’s GPU is enabled. In other words, on a desktop with a single monitor connected to a discrete GPU, you can’t use Quick Sync.
This isn’t a problem for mobile since Sandy Bridge notebooks should support switchable graphics, meaning you can use Quick Sync without waking up the discrete GPU. However there’s no standardized switchable graphics for desktops yet. Intel indicated that we may see some switchable solutions in the coming months on the desktop, but until then you either have to use the integrated GPU alone or run a multimonitor setup with one monitor connected to Intel’s GPU in order to use Quick Sync."
Eventually, high-end consumers are meant to buy Sandy Bridge-E on the LGA2011 platform.
The vast majority of LGA1155 users are going to use that integrated gpu. Producing it as a separate chip would be much more expensive -- on-die it can share expensive resources, notably the memory controller, and moving it away from the chipset means there only has to be one high-performing piece of silicon in the whole system.
Using a quarter more die does not mean that the cpu is a quarter more expensive -- the fabrication costs are a minuscule amount of the sale price, and if this move allows Intel to reduce the total amount of different high-performing chips made, it could well result in a reduction of total costs, potentially making the CPU less expensive -- even to people who do not use the integrated GPU.
Strictly for the purposes of (expensive) price comparisons of processor products from Intel (and yes, straying from the x86 architecture), the various versions of the Intel Itanium 9300 series processors were (are?) priced from $946 to $3838, and that's in quantities of a thousand.
Great progress, especially in the power consumption. However, last generation processors are already very fast. A better upgrade for most developers would be an SSD though. All of this speed is great, but useless as long as the processor is waiting on a mechanical hard disk.
Once you get used to your applications opening almost instantly, it's tough to go back to a mechanical HDD.
I bought a 60GB one last year, which is plenty for my OS and applications--I use a 1.5TB external for storage. It's the most tangible difference in my computing experience since 4 years ago when I bought 8GB of RAM and got rid of my swap file. If you're willing to spend money to make things faster, spend it on an SSD.
My "perspective" is that Sandy Bridge has nothing new on the trusted computing front and Microsoft isn't smart enough to use that hardware anyway (see Palladium).
30 comments
[ 3.2 ms ] story [ 56.4 ms ] threadThese new instructions promise to significantly increase throughput for a limited set of operations -- but to do that, the software has to be compiled to use them, and often entirely rewritten to use the wider vectors. Few will bother to do that until the new instructions have significant market penetration -- which will take years.
I am not quite as optimistic as Symmetry is about the usefulness of these latest extensions to x86 -- while many fringe tasks will benefit a lot, most of what you do on a computer just isn't limited by the width of the vectors or the amount of SIMD operations you can do per clock. IMHO, the insanely fast caches are a much more significant improvement.
Both run Ubuntu pretty well, for what I usually hack on (python/django) and the desktop used to run the Source engine games pretty well (under XP). I've recently started learning Android and Eclipse (and especially the AVD emulator) and both of those machines just thrash.
I'm looking to build a new box and pick up a Thinkpad when they update. Going to be an expensive year.
LGA1155 is their budget line. 2600K is not the best processor they can make -- it is the best processor they can make give very stringent whole platform cost limits. In addition to having the desktop budget line, they also have a desktop high-end line. Last generation, these were the LGA1156 and LGA1366 respectively. The processors they just released are the budget line for this gen, and the high-end line fits a socket called LGA2011, which will be released later, and on which the processor cost will range from 300$ to about 1000$.
In every generation, new Intel processors gain roughly as much more power as is the difference between the budget and high-end product lines, so it's normal for todays budget processors to win over yesterday's high end. Usually Intel releases the high-end first for better margins and to avoid this kind of confusion -- this time, they started on the budget end, probably to pummel AMD, who previously only competed on the budget segment and now cannot really compete much at all.
IMO Intel focusing on the low to mid range of this release cycle is vital or they are going to be eaten.
Everything I've seen seems to indicate that this is true. Did I miss something?
But, far more importantly, producing a 32nm graphics chip is more expensive (for now) than a 45nm chip using the same number of transistors. Intel has a huge process advantage and is often on the lookout for useful things to do with their huge transistor budget, but integrated graphics has huge downsides, most notably memory bandwidth issues.
I get that you decrease yield, but w/o knowing the numbers, it's hard to really see if its much impact on the bottomline. I suspect its a bigger win than loss (not hard to speculate, since Intel apparently had made the same bet).
To give some idea of cache's relative value a GT110 is a 3 billion transistor chip with less than 1mb of cache. A QX9650 is a 0.82 billion transistor chip with 12MB of cache.
PS: At a minimum it's a 1 - (1/1.2) = 17% loss but adding in bad chips it's probably closer to 25% until they start selling chips without video capabilities. Also, I am not saying it's a bad long term move, just that in the short term it's risky.
I can imagine a similar setup in desktop machines to reduce their power consumption.
I'm pretty certain I read (an article on anantech?) that graphics switching is explicitly not going to be supported on desktops until the next release of intel processors.
Here it is:
http://www.anandtech.com/show/4083/the-sandy-bridge-review-i...
"Quick Sync with a Discrete GPU There’s just one hangup to all of this Quick Sync greatness: it only works if the processor’s GPU is enabled. In other words, on a desktop with a single monitor connected to a discrete GPU, you can’t use Quick Sync.
This isn’t a problem for mobile since Sandy Bridge notebooks should support switchable graphics, meaning you can use Quick Sync without waking up the discrete GPU. However there’s no standardized switchable graphics for desktops yet. Intel indicated that we may see some switchable solutions in the coming months on the desktop, but until then you either have to use the integrated GPU alone or run a multimonitor setup with one monitor connected to Intel’s GPU in order to use Quick Sync."
The vast majority of LGA1155 users are going to use that integrated gpu. Producing it as a separate chip would be much more expensive -- on-die it can share expensive resources, notably the memory controller, and moving it away from the chipset means there only has to be one high-performing piece of silicon in the whole system.
Using a quarter more die does not mean that the cpu is a quarter more expensive -- the fabrication costs are a minuscule amount of the sale price, and if this move allows Intel to reduce the total amount of different high-performing chips made, it could well result in a reduction of total costs, potentially making the CPU less expensive -- even to people who do not use the integrated GPU.
The server ones are even more expensive.
Intel Xeon X5677 = $1663
Intel Xeon X5680 = $1570
Intel Xeon W3680 = $1057
Once you get used to your applications opening almost instantly, it's tough to go back to a mechanical HDD.