September 28, 2006 4:34 AM PDT

Processor, memory may marry in future computers

Related Stories

Intel opens up chip connections

September 27, 2006

Intel pledges 80 cores in five years

September 26, 2006

Intel to unveil new way to connect chips

March 2, 2005
A correction was made to this story. Read below for details.
SAN FRANCISCO--Will Intel get back into the business of selling computer memory? It might, if one of the ideas in a recent chip prototype moves into manufacturing.

On Tuesday, Intel showed off an 80-core processor at its developer forum taking place in San Francisco this week and one of the prominent features of the chip is that each core is connected directly to a 256KB memory chip through a technology called Through Silicon Vias, or TSV.

The memory wedded to the processor cores could constitute the entire memory needed for a computer, Intel CTO Justin Rattner told News.com in an interview during the Intel Developer Forum. TSV could be used in a variety of chips, not just the 80-core monster. As a result, computer makers, when building a system, would get their memory when they bought their processors from Intel. They would not have to obtain memory chips separately from other companies like they do now.

"You could buy it as a block," he said.

Back in the 1980s, Intel was one of the largest manufacturers of computer memory, or DRAM, but it got out of the market because of punishing competition from Japan. (The company, however, continues to produce NOR flash memory and a few other types of memory. Typically, these chips are not used in the same fashion as DRAM.)

Wedding memory directly to the processor would have huge performance benefits. Currently, memory and the processor in Intel-based computers exchange data through a memory controller, which moves at a far slower rate than the processor. It's one of the big bottlenecks in computer performance. TSV, which displaces the memory controller, would shuttle data far quicker.

Potential competitive edge
Data coming out of memory also squeezes through an overcrowded port. TSV would effectively open up thousands of ports. Overall, Rattner said, TSV is by far a more notable accomplishment than putting 80 cores on the same piece of silicon.

The processor cores are not restricted to getting memory from the chip wedded to it, added Rattner. The cores are connected to each other through high-speed links controlled by a router integrated into each core. Overall, the prototype 80-core chip has an aggregate memory bandwidth of 1 terabyte per second, meaning that it can shift a trillion bytes per second.

Conceivably, TSV would also put AMD on the hot seat. A good portion of the performance gains AMD achieved with the Opteron chip came from Opteron's integrated memory controller. Intel does not put integrated memory controllers on its chips.

In all probability, Intel would not get back into the business of manufacturing DRAM--it remains one of the most difficult chip markets in which to turn a profit. Nonetheless, TSV would put it back into the business of selling it.

Rattner, however, noted that implementing TSV will take time. The memory chips attached to the 80-core processor are SRAM, a relatively expensive memory that Intel still makes. The next step is to see how well DRAM works with TSV.

Engineers would also have to devise packages that would let the processor and memory live together. The processor typically generates more heat than the memory, which is one of the factors that would have to be considered. Although it doesn't get many headlines, packaging design is a huge challenge for chipmakers.

"It's still in the research stage," he said. "We will do a lot of work with it in the next several years."

 

Correction: In his keynote, Intel CTO Justin Rattner said that 256MB of SRAM existed for each core. He later clarified his comments to mean that the chip has 256KB to each core.

See more CNET content tagged:
memory chip, SRAM, DRAM, AMD Opteron, computer memory

27 comments

Join the conversation!
Add your comment
I hate to ask but..
I really hate to ask but would this mean in order to add more RAM one would have to upgrade the processor? I mean I would hope this would not be another 640k should be enough for anybody move.

Though I would guess we would end up with 2 banks of RAM one integrated and one like we have now.
Posted by pmfjoe (196 comments )
Reply Link Flag
No not like that...
Its not like that...the memory and the chip would be one their would be no slot. To add more RAM to a computer that specific computer manufacture would have to use a motherboard, one I would hope Intel would make, that would also include a memory chip controller and standard memory slots.

These would be used to upgrade the RAM but to be completely honest if this setup works the amount of RAM needed might be lessened because the processor would be moving data at a much faster rate and would not need to have huge amounts of information sitting in RAM waiting to be executed, moved and used....

You get what I am saying?
Posted by lpbruop (7 comments )
Link Flag
Remember the Commodore Amiga?
I am a novice in the areas of processors and memory chips....but if my memory serves me correctly, wasn't this idea originally done by the Commodore Amiga? It had 3 separate processors and built in "chip RAM" plus a separate bank of RAM you could add to. I think that is why the Amiga was so good back in the day at multitasking. Can anyone verify or add to my post?

swokjr
Posted by swokjr (8 comments )
Link Flag
Is it really 1 terabyte/sec or is it terabit?
Posted by nb2000nb (26 comments )
Link Flag
As well as graphics and other subsystems.
The technology exists now, whether it's on the same die or whether it is a hybrid like the SECC PII and PIII cartridges.
This would be primarily for locked down systems or laptops where pricepoint is more important the capability.
Memory is evolving so fast that virtually anything is possible given enough resources and time.
System memory speeds have always been a bottle neck to computers and unless you design the memory in the core at the processors' internal clock which would probably require GaAs type static type memory on a low capacitance substrate you're still going to have wait states and latencys just not like we do now.

Systems on a chip (SOC) will be here sooner than most realize particularly given AMD just bought ATI. Integration is the name of the game since the war now is price not so much technology. So first markets are laptops and price concious demographics (third world countries).

One caveat here is that there will ALWAYS be room for an expandable architecture so don't look at this as being "I'll have to switch processors to increase my system memory" exclusively since most SOC applications being cost sensitive will not have a socket anyway.
Posted by fred dunn (793 comments )
Reply Link Flag
Think Bigger...
You are saying for low end systems. This is an 80 Core 3GHz+ system that could be out in the next 5 years. That's a lot of processing power over what is in current low end systems.

Also the memory on these chips is to allow a vastly improved memory bandwidth. Not something a low end system will need. The example model shows 20GB of RAM. More than enough RAM and CPU power in one "chip" to make the most demanding gamer or 3D CAD designer VERY happy to own a system using this chip.

Oh, and it won't be in a laptop anytime soon. :)
Posted by zaznet (1138 comments )
Link Flag
Anti-Trust Suit on the way?
If this is successful, as with Microsoft, anti-capitalists are going to have a fit. Big government folks in the U.S. and the EU will sue intel and other chip makers for having an unfair advantage. Companies who make memory will fight this very hard. I hope the prototype becomes a reality because it could really make computing very efficient.
Posted by ZeekJones77 (8 comments )
Reply Link Flag
Apple is a bigger threat.
Providing chips to Apple is more anti-competitive than providing memory on a chip which really has been done for many years.

Heck, the memory per-core on this sample is not as large as the memory per core in some chips currently provided by Intel and AMD. It is just that with 80 cores you now have 80 times the memory and finally enough to run an application entirely on the chip.
Posted by zaznet (1138 comments )
Link Flag
History repeats
I am a novice in the areas of processors and memory chips....but if my memory serves me correctly, wasn't this idea originally done by the Commodore Amiga? It had 3 separate processors and built in "chip RAM" plus a separate bank of RAM you could add to. I think that is why the Amiga was so good back in the day at multitasking. Can anyone verify or add to my post?

swokjr
Posted by swokjr (8 comments )
Reply Link Flag
Different Designs...
Very different designs here. You are talking about an archetechture that was decentralized (the Amiga) and one that is becoming more centralized (Intel). Memory on a chip has existed in the Intel market for a long time and increases processor performance.

Amiga put work in multiple places. Intel is moving more functionality into the chip. There are advantages to both designs and that is to reduce the bottleneck on the bus.
Posted by zaznet (1138 comments )
Link Flag
Different Designs...
Very different designs here. You are talking about an archetechture that was decentralized (the Amiga) and one that is becoming more centralized (Intel). Memory on a chip has existed in the Intel market for a long time and increases processor performance.

Amiga put work in multiple places. Intel is moving more functionality into the chip. There are advantages to both designs and that is to reduce the bottleneck on the bus.
Posted by zaznet (1138 comments )
Link Flag
256 per core?
On Tuesday, Intel showed off an 80-core processor at its developer forum taking place in San Francisco this week and one of the prominent features of the chip is that EACH CORE IS CONNECTED DIRECTLY to 256 MB memory chip through a technology called Through Silicon Vias, or TSV.

I have read the article, and it still looks like 20 gig to me.
Posted by dragonfly8610 (49 comments )
Reply Link Flag
Reading Comprehension 101
and if you had read the IDF, it states,
"Intel's prototype uses 80 floating-point cores, each running at 3.16GHz, Justin Rattner, Intel's chief technology officer, said in a speech following Otellini's address. In order to move data in between individual cores and into memory, the company plans to use an on-chip interconnect fabric and stacked SRAM (static RAM) chips attached directly to the bottom of the chip, he said."

Sure looks like it says 1 256 meg chip per core to me.

Maybe some people should try a reading comprehension course.
Posted by dragonfly8610 (49 comments )
Link Flag
This is a band-aid fix
I can't remember who said this, it might be Tom Halfhill:

Adding more memory in the processor is good like a vaccine is good.
Posted by qwerty75 (1164 comments )
Reply Link Flag
3rd layer...
While this looks like L1, L2 or L3 cache it is enough to fill the RAM needs of the system. This places main RAM in the chip and thus provides a much faster bus. It's not just a small step, it's pretty big and a huge improvement in performance.

The next step would be to make it modular where a RAM chip plugs into a Processor chip and the combined item fits into a socket on the system board.
Posted by zaznet (1138 comments )
Link Flag
Doesn't Amiga have this technology?
As far as I know classic Amigas have a similar technology. Processor has direct access to fastram, and custom chipsets to chipram, thus allowing the computer to operate at twice the speed it could achieve.
Posted by ancalimon (1 comment )
Reply Link Flag
 

Join the conversation

Add your comment

The posting of advertisements, profanity, or personal attacks is prohibited. Click here to review our Terms of Use.

What's Hot

Discussions

Shared

RSS Feeds

Add headlines from CNET News to your homepage or feedreader.