This looks like they are referring to that techradar article: http://tech.uk.msn.com/news/article.aspx?cp-documentid=9896153
Looks like just rumours.
Looks like just rumours.
Predict: The Next Generation Console Tech
- Thread starter Acert93
- Start date
- Status
TimothyFarrar
Regular
First page before the article got pulled,
"Intel's upcoming Larrabee chip will pack nearly two billion transistors and 64 tiny x86 processor cores. That's what sources close to Intel's intriguing co-processor project have told TechRadar today.
Factor in an expected operating frequency of at least 2GHz and it's likely Larrabee will have four to five times the raw computational grunt of the fastest current graphics chips.
Now, you might be wondering whether this is just another throw away rumour, a headline grabbing stab in the dark. Certainly Larrabee remains a long way out. It won't appear until the end of next year at the absolute earliest. Much can change in that time.
Bright future
But if you cast your web browser back to last summer, you may remember a little story we broke. We suggested that Larrabee's core architecture would be based on the Pentium MMX CPU of the late 1990s. And whaddya know, that turned out to be true.
So, what else do we know about Larrabee? We're told the first generation chip will be available in both 64 and 32 core flavours (though we're not sure if both will be go at launch). The chip itself is said to be big, really big.
In fact, it's on a similar scale to Intel's recently released six-core Xeon server chip, codenamed Dunnington. Indeed, Dunnington is built using essentially the same 45nm silicon process as Larrabee will use and packs 1.9 billion transistors. In other words, Intel is already making a Larrabee-scale x86 chip with existing technology.
We also hear the board that Larrabee will be strapped to looks very much like a conventional graphics card complete with a large but not ludicrous cooling solution.
Of course, what we all really want to know is how powerful Larrabee will be. Well, that's something we can also shed some light on, at least in theoretical throughput terms.
Intel has revealed that each Larrabee core will pack a supercharged ALU capable of 16 vector operations per cycle. With 64 cores that works out to no less than 1,024 vector ops per cycle. By comparison, AMD's top GPU, the Radeon HD 4870, can do 800 per cycle while NVIDIA's GeForce GTX 280 knocks out a relatively modest 240."
"Intel's upcoming Larrabee chip will pack nearly two billion transistors and 64 tiny x86 processor cores. That's what sources close to Intel's intriguing co-processor project have told TechRadar today.
Factor in an expected operating frequency of at least 2GHz and it's likely Larrabee will have four to five times the raw computational grunt of the fastest current graphics chips.
Now, you might be wondering whether this is just another throw away rumour, a headline grabbing stab in the dark. Certainly Larrabee remains a long way out. It won't appear until the end of next year at the absolute earliest. Much can change in that time.
Bright future
But if you cast your web browser back to last summer, you may remember a little story we broke. We suggested that Larrabee's core architecture would be based on the Pentium MMX CPU of the late 1990s. And whaddya know, that turned out to be true.
So, what else do we know about Larrabee? We're told the first generation chip will be available in both 64 and 32 core flavours (though we're not sure if both will be go at launch). The chip itself is said to be big, really big.
In fact, it's on a similar scale to Intel's recently released six-core Xeon server chip, codenamed Dunnington. Indeed, Dunnington is built using essentially the same 45nm silicon process as Larrabee will use and packs 1.9 billion transistors. In other words, Intel is already making a Larrabee-scale x86 chip with existing technology.
We also hear the board that Larrabee will be strapped to looks very much like a conventional graphics card complete with a large but not ludicrous cooling solution.
Of course, what we all really want to know is how powerful Larrabee will be. Well, that's something we can also shed some light on, at least in theoretical throughput terms.
Intel has revealed that each Larrabee core will pack a supercharged ALU capable of 16 vector operations per cycle. With 64 cores that works out to no less than 1,024 vector ops per cycle. By comparison, AMD's top GPU, the Radeon HD 4870, can do 800 per cycle while NVIDIA's GeForce GTX 280 knocks out a relatively modest 240."
According to http://www.anandtech.com/showdoc.aspx?i=3414 The die size of Dunnington is 503mm^2. That would make the 64 core version far too much for any console at 45nm.
My quickie calculation implies that a 64 core might creep down to about 300mm^2 for a 32nm version which *Might* just be enough to squeeze one in.
My quickie calculation implies that a 64 core might creep down to about 300mm^2 for a 32nm version which *Might* just be enough to squeeze one in.
I'm not getting how they figure 4-5x the raw computational power of todays most powerful GPU's.
So the're saying its going to pack between 4 and 6 TFLOPs?
Does that work out for 64 cores at 2Ghz??
So the're saying its going to pack between 4 and 6 TFLOPs?
Does that work out for 64 cores at 2Ghz??
Intel starts 32nm mass production next fall and 22nm should be ready in time for the next Xbox but it probably wouldn't be used right away. If they manage to make games run on just Larrabee then having only one processor in the console will benefit costs significantly.
I take it that's at least part of the reason why they temporarily pulled the article? Or is my admittedly weak grip on GPU terminology slipping?
So we've gone from 32 cores at 1GHz in their SIGGRAPH paper to 64 at 2GHz. 64 cores at 16 scalar [sic?] MADs per core is I guess 1024 * 2flops * 2GHz = 4Tflops (single precision)? Which I guess is ballpark "four to five times" G280 and 4870. I forget how many (or how capable are the) texture units LRB's packing per core.
Edit: I see the tech.uk.msn link, which seems to be the same article and includes page two, uses much the same numbers.
I remember reading about 16 cores @1GHz prototype. It's messy...
Anyway,so if there any thruth to these rumors a LRB core should resolve around 30 millions of transistors.
If it's true it's always impressive to see how good are Intel process.
A cell @45 is 115mm² a SPE is ~15 millions transistors
Intel can pack 64 twice as big cores in ~500mm², that's evil
Anyway,so if there any thruth to these rumors a LRB core should resolve around 30 millions of transistors.
If it's true it's always impressive to see how good are Intel process.
A cell @45 is 115mm² a SPE is ~15 millions transistors
Intel can pack 64 twice as big cores in ~500mm², that's evil
Last edited by a moderator:
TimothyFarrar
Regular
I think they are saying that a 64core at 2GHz will be 4 tflops. I keep in mind that every clock which Larrabee spends emulating the GPU's fixed function hardware is a drop in 2048 flops. Hard to tell what percentage of Larrabee clocks will be in fixed function vs running actual shaders. Then for possible Larrabee as GPU and CPU console, another percentage of clocks are bound to be running badly coded OO C++ code....
Wouldn't the only way Larrabee would be used in future consoles, is if either MS, Sony or Nintendo bought the chips directly from Intel? I doubt Intel would license out the design and even if they did, who could manufacture the chip at Intel specs without Intel's semiconductor technology.
Buying the Larrabee chips directly is in no way going to be equal to or cheaper than licensing the design and outsourcing manufacturing of that GPU, which while maybe not as powerful as Larrabee would be more economical and and designed to accomodate a small form factor.
Buying the Larrabee chips directly is in no way going to be equal to or cheaper than licensing the design and outsourcing manufacturing of that GPU, which while maybe not as powerful as Larrabee would be more economical and and designed to accomodate a small form factor.
Last edited by a moderator:
How are you so sure? This isn't the celeron from Xbox 1, that was a mainstream cpu that Intel was selling just fine regardless of the console deal. Intel might feel compelled to offer Larrabee to MS at barely above cost in order to make sure their design gets widespread use and support. Landing a next-gen console deal is the best and perhaps the only way to do that.
Another question for the experts. Is there anything preventing intel from throwing in a couple more powerful general purpose cores (say using the space of 2 or 3 larrabee cores), attaching them to the ringbus, and providing a one-chip solution for a next gen console?
Last edited by a moderator:
I'm curious if anyone believes that AMD is capable of creating a competitive platform for the next generation of consoles.
I looked for an article I remember reading detailing AMD's future direction after their ATI acquisition. Here's the relevant page.
It would seem to me that their work on interconnects and integration would (assuming they actually realize their goals) put them in a position to offer a complete solution of CPU/GPU/Platform with a promise of future cost reduction over the life of the console as they follow their discreet component->package integration->die integration strategy.
I looked for an article I remember reading detailing AMD's future direction after their ATI acquisition. Here's the relevant page.
It would seem to me that their work on interconnects and integration would (assuming they actually realize their goals) put them in a position to offer a complete solution of CPU/GPU/Platform with a promise of future cost reduction over the life of the console as they follow their discreet component->package integration->die integration strategy.
Success on a closed system meant for gaming isn't going to ensure or help widespread use in the PC market. Larrabee is Intel's response to GPGPU for the high performance computing market and a way to keep AMD from operating in the graphics market by itself.
Intel is going to have all the traction it needs to make sure larrabee gets ample sales by making a less powerful, cheaper and intergrated mobile form of larrabee a standard component of Centrino. All Intel has to do after that is make sure that its higher end version of Larrabee can compete against whatever Nvidia or AMD is offering. Intel doesn't need to sell MS or Sony Larrabee at a dirt cheap priceswhen its has better alternatives to market adoption.
StreamIt: the next-gen software technology?
If the future is going to be one of huge data-flows through many cores, a new language designed around streaming might prove attractive. Have any developers here heard of the MIT language project StreamIt?
http://www.cag.lcs.mit.edu/streamit/
If the future is going to be one of huge data-flows through many cores, a new language designed around streaming might prove attractive. Have any developers here heard of the MIT language project StreamIt?
http://www.cag.lcs.mit.edu/streamit/
unless its a bitch to program for . The current tools out there may not work well for this design.
We also don't know how this will perform against newer model ati adn nvidia gpus.
If the performance is there and heat output can be lowered, the devs will have no choice as LRB is more than likely to be used in Intel plans for its hetergenous multi-core chips. Even if LRB can't perform as a discrete card it slated to take the place of GMA, which is more important for Intel than the discrete market.
There are other chip makers than intel. Labree can suck and gaming shifts to amd.
AMD and Nvidia already own the discrete card market and GMA already sucks, regardless Intel is the biggest graphics provider in the market. The biggest push now for discrete cards are GPGPU, which is a effort to diversify the market of discrete cards to segment of the PC market outside of gaming. LRB is Intel strategy to limit that push, compete with AMD fusion and provide competition to AMD in the discrete card market. Intergrating low end LRB with Intel cpu accomplishes the first two facets strategy while pratically ensuring wide adoption by devs.
Ultimately from a gpu perspective, LRB in its integrated form only has to compete against the low end discrete market for LRB to be deemed an gaming success. Devs code the lowest common denominator and if a lowend LRB performs like a low end discrete card, LRB will see widespread adoption by devs because it will be in almost all new CPU/GPU hybrids.
Intel only has to worry about performance with LRB as a discrete card because the lowend intergrated version will ensure widespread adoption from a coding point of view.
From a timing standpoint, I think the recent Halo 3 recon announcement puts a likely firmer date on the Xbox 3.
Some time ago rumors had stated MS had a alternating years Gears/Halo tentpole holiday title strategy. As such, it would now look like this:
06 Gears
07 Halo 3
08 Gears 2
09 Halo 3: Recon
10 Gears 3
Which means more than likely imo the current plan is Xbox 3 in fall 2011, launching with Halo 4 currently under dev by the MS Halo team.
Some time ago rumors had stated MS had a alternating years Gears/Halo tentpole holiday title strategy. As such, it would now look like this:
06 Gears
07 Halo 3
08 Gears 2
09 Halo 3: Recon
10 Gears 3
Which means more than likely imo the current plan is Xbox 3 in fall 2011, launching with Halo 4 currently under dev by the MS Halo team.
I like your thinking there. It sounds like the perfect timing for the console to me.
- Status
Similar threads
- Locked
