bit-tech.net

Larrabee more efficient than GPUs at rasterisation

Larrabee more efficient than GPUs at rasterisation

Intel's forthcoming Larrabee architecture is based on x86 cores, but is apparently still efficient at rasterisation tasks.

It’s the product that’s either going to transform the graphics chip industry beyond recognition or simply become another footnote in PC gaming history.

Larrabee, Intel’s first foray into the world of discrete graphics since the abysmal i740, has been a hot talking point in the PC gaming business, not just because it’s another competitor in the GPU business, but because it also relies on software rendering via x86 CPUs. However, while Intel has given us very few details about the architecture, little is known about how the chip will actually perform, which is what makes games programmer Mike Abrash’s recent session on Larrabee at GDC particularly interesting.

VentureBeat attended the lecture, in which Abrash revealed some of his experiences of the performance of Larrabee in comparison to traditional GPUs. According to VentureBeat, Abrash claimed that, despite using software rendering, Larrabee can in fact be more efficient than traditional GPUs when it comes to performing rasterisation tasks.

The site also says that Abrash claimed that Larrabee’s performance will be above 1TFLOPS, although this computing milestone has already been passed by current traditional GPUs such as the Radeon HD 4870. Interestingly, Abrash also noted that the “raw graphics performance” of Larrabee isn’t likely to be as fast as other graphics chips, but praised Intel’s chip for its power-efficiency and flexibility.

The latter appears to be particularly appealing to programmers such as Abrash, who worked on the original Doom, Quake and Half-Life games, before moving to RAD Game Tools. According to VentureBeat, Abrash described Larrabee as the “most fascinating graphics architecture” that he’d seen in 15 years, with the chip’s cores making it more flexible than a standard GPU pipeline.

By using multiple x86 processors, many people have noted that Larrabee could also be used for ray tracing as well as rasterisation, although it would need a tremendous amount of computing power to do this effectively in real time.

However, VentureBeat also quotes a memo apparently sent by Nvidia CEO Jen-Hsun Huang to Nvidia employees, which says that Intel could potentially breach the cross-license between the two companies by releasing Larrabaee.

“In our countersuit,” says the memo, “we assert our belief that we are licensed to build chipsets for Intel processors. In a pair of agreements signed in 2004, we negotiated for rights to build chipsets. In exchange, Intel obtained a cross license to our valuable patents. Today, Intel is using technologies that we invented in their integrated graphics chips. And they will soon integrate Nvidia patented technologies into their CPUs and upcoming Larrabee processors.”

Could flexibility become more important than raw graphics performance in the future of graphics chips? Let us know your thoughts about Intel’s Larrabee graphics chip in the forums.

34 Comments

Discuss in the forums Reply
Xtrafresh 30th March 2009, 14:48 Quote
pics or it didn't happen! :D
perplekks45 30th March 2009, 15:58 Quote
What about a counter-countersuit?

On topic:
I just can't wait to see how Larrabee performs. Although I think it'll be designed to sell in the mainstream market rather than in the high-end segment. Still... interesting.

//edit: Just a quick question: Who writes the articles at Venturebeat? 10 year olds? :| Or is just aimed at people who have hardly any detailed knowledge of the technology behind graphics cards?
Goty 30th March 2009, 16:08 Quote
The biggest hurdle that Larrabee is going to have to overcome is that of drivers. As anyone that's been around long enough to remember the Radeon 8500 can attest to, drivers can make or break a video card, and driver support has never been Intel's forte.
wuyanxu 30th March 2009, 16:19 Quote
"can be", that's the most important part.

i don't see how Intel, who is a new player can out-perform nVidia/ATI in their own games.
bowman 30th March 2009, 16:30 Quote
Apparently the low raw performance is due to the frequency of the part being at the performance/watt 'sweet spot'. So those of us who couldn't care less about energy efficiency should potentially be able to overclock it quite a bit.
Skiddywinks 30th March 2009, 16:32 Quote
Quote:
Originally Posted by wuyanxu
i don't see how Intel, who is a new player can out-perform nVidia/ATI in their own games.

It has already been said they won't be. People need to stop assuming Intel plan on usurping ATI and nVidia in one fell swoop.

Give Intel some credit, what they have seems very interesting, and has a lot of potential.
Turbotab 30th March 2009, 17:59 Quote
By the time Larrrabee launches, we will have the high-end 40nm DirectX 11 cards pushing out possibly over 4 Teraflops in dual GPU cards. The biggest problem for Intel will be, mainstream 40 nm cards, retailing at the £100 mark, offering well over 1 TFLOPS. Hell, even ATI's current 40nm card the RV740 series, produces nearly a TFLOP, and according to Fudzilla, is only rated at 80 watts. IIRC correctly the TDP for Larry is rumoured to be quite high, so for what market is Intel aiming.
Furymouse 30th March 2009, 18:08 Quote
Agreed with Wuyanxu. Maybe this is why they are in suits against both AMD and Nvidia. It " can be " the best if we get the others money mired in lawsuits instead of R&D :D
perplekks45 30th March 2009, 18:29 Quote
The whole point of Larrabee is flexibility and power efficiency, as mentioned before.
I don't think Intel will fail on the latter. They might not be as effective as ARM as we've learned but they're pretty good at that.
nicae 30th March 2009, 19:23 Quote
We must remember that Larrabee will face an even greater challenge than PhysX. NVIDIA has ~65% of the market and still struggles to make developers adopt PhysX. How many developers will program their rasterisation for x86 Larrabee if it has 0% of the market?

I'm sure Intel can really push Larrabee by integrating these cards just like they push their horrible IGPs, but how fast can they do that and will they get far enough among the hardcore crowd to convince major developers into going through all the development hassle? Development is already a very costly process and the financial crisis is making many studios close, suggesting they don't have spare cash to spend on such risky projects.
perplekks45 30th March 2009, 22:12 Quote
nVidia has what?! 65% of what market?

In 2007 it looked like this:

http://media.arstechnica.com/news.media/graphics.png

Source: ArsTechnica

And according to JPR's Q208 report:
Quote:
Intel held the top spot for overall graphics market with 47.3% market share. NVIDIA has 31.4% of the market and AMD held 18.1% of the market.

Source: i4u.com
Coldon 30th March 2009, 22:43 Quote
the discrete gfx card market... intel has 0% in that department...

If you remove all the crappy intel IGPs that the shipped with the bulk of the office machines out there. I think that a geforce 4 mx is maybe on par with the best IGP intel has.

The situation at the moment is quite scary with the current intel lawsuits, it almost looks like intel is trying to block the competition from using its technologies while still keeping hold of their tech ( SLI on x58, AMD64 extensions on C2,I7)...

I'm holding thumbs that intel loses those suits and is made to pay the consequences for their anti-trust actions.

On the larrabee issue, larrabee might be almost as efficient at rasterization as newer GFX cards, but what will it matter, by the time larrabee launches (assuming a delayed launch) we'll have GFX cards with hardware tessellation units and DX11.

Nevermind the insane power requirements a 32 core larrabee card will need, estimations of this if intel uses a 45nm process, is around 300w just for the cores, once you include memory and other subsystems, you'll be looking at a 400w power draw. And from all account those cores will be kept pretty busy even for basic tasks.

Personally I'm holding thumbs for larrabee to be a epic failure but then Intel will just strong arm OEMs into supplying the cards to their customers, just like with their atrocious IGP chipsets, that 45% of people are using today!
nicae 30th March 2009, 22:45 Quote
Quote:
Originally Posted by perplekks45
nVidia has what?! 65% of what market?

In 2007 it looked like this:

http://media.arstechnica.com/news.media/graphics.png

Source: ArsTechnica

And according to JPR's Q208 report

Source: i4u.com

I was speaking of discrete graphics, where Intel has no presence and S3 can be disconsidered.

Therefore, we're talking about a market of 31.4% + 18.1% = 49.5%.
As such, NVIDIA has 31.4% / 49.5% = 63.4343434343434% of that market, with the rest left to ATI. My memory of Valve's hardware surveys ain't that bad, eh? =)

But that's still approximate as we would also need to split NV and ATI IGPs, but that won't change my point.
Adnoctum 31st March 2009, 06:35 Quote
Just a thought I had (sorry if this has been asked elsewhere)....

As Larrabee is based on x86 architecture and instructions, and a part of the agreement with AMD is to share x86 and developments from it (e.g. x86-64, otherwise known as AMD64), doesn't this mean that AMD could make their own Larrabee compatible part if this whole architecture takes off?

And if it does, where does this leave nVidia, who doesn't have an x86 licence?
Adnoctum 31st March 2009, 06:52 Quote
Double post, f*#@ing wireless.
wuyanxu 31st March 2009, 11:16 Quote
x86 is not as good as it sounds, it's got a lot of unused instructions since the days of i386. it's power hungry and inefficient comparing to RISC processors.

although they say it's going to be x86 compatible, i highly doubt that, it's probably just part of the x86 instruction set, just enough for ray-trace and current generation.
nicae 31st March 2009, 12:38 Quote
Quote:
Originally Posted by Adnoctum
Just a thought I had (sorry if this has been asked elsewhere)....

As Larrabee is based on x86 architecture and instructions, and a part of the agreement with AMD is to share x86 and developments from it (e.g. x86-64, otherwise known as AMD64), doesn't this mean that AMD could make their own Larrabee compatible part if this whole architecture takes off?

And if it does, where does this leave nVidia, who doesn't have an x86 licence?

The question is: Will they want to make x86 graphics cards? It's a very risky change and there's a lot of market inertia that should hold back the architecture from industry-wide adoption. If it were a two-man race (out of the three men NV, AMD and Intel) or if x86 was guaranteed to be better, than they would be scrambling to get it done as well. Yet, if only Intel tries the push, AMD and NV could remain comfortably earning cash on current rasterisation cards and leave the costs of change to Intel. Once it does change the industry (if it does), they can jump in. With a bit of delay, for sure, but a tiny fraction of the costs and almost no risk. Sounds like a more reasonable plan considering you have a multibillion-dollar company at stake.

Assuming x86 turns out to be better, yes, AMD could hop in, as well as VIA, I'd guess. NVIDIA would be left out as they kind of already are, justifying their seeking of x86 with VIA and internal developments.
DXR_13KE 31st March 2009, 12:52 Quote
i feel that this graphic card will be utter epic fail...
gagaga 31st March 2009, 13:13 Quote
Quote:
Originally Posted by Coldon
Nevermind the insane power requirements a 32 core larrabee card will need, estimations of this if intel uses a 45nm process, is around 300w just for the cores, once you include memory and other subsystems, you'll be looking at a 400w power draw. And from all account those cores will be kept pretty busy even for basic tasks.

I disagree - that's nearly 10w a core. You can run a full Core 2 with cache at that level etc at that power level. Remember these are dumb simple cores with in-order execution and a lot of the power hungry bits missing, just like Atom. I'm guessing they'll be looking at around 2w a core giving a total near to 80-90w all up. Still a lot, but it depends on what the load is like and what amount of useful work you're getting out of the other end.
gagaga 31st March 2009, 13:13 Quote
Quote:
Originally Posted by gagaga
Quote:
Originally Posted by Coldon
Nevermind the insane power requirements a 32 core larrabee card will need, estimations of this if intel uses a 45nm process, is around 300w just for the cores, once you include memory and other subsystems, you'll be looking at a 400w power draw. And from all account those cores will be kept pretty busy even for basic tasks.

I disagree - that's nearly 10w a core. You can run a full Core 2 with cache at that level etc at that power level. Remember these are dumb simple cores with in-order execution and a lot of the power hungry bits missing, just like Atom. I'm guessing they'll be looking at around 2w a core giving a total near to 80-90w all up. Still a lot, but it depends on what the load is like and what amount of useful work you're getting out of the other end.

Ooops, I meant a single Core 2 core.
azrael- 31st March 2009, 14:38 Quote
Quote:
Originally Posted by wuyanxu
although they say it's going to be x86 compatible, i highly doubt that, it's probably just part of the x86 instruction set, just enough for ray-trace and current generation.
The Larrabee cores are supposed to be modified (and heavily shrunken) P54 cores. Yup, that's P54 as in 2nd generation Pentium processors. :)
Coldon 31st March 2009, 15:26 Quote
the 300w power draw is pretty much the common estimate, on the web.

here's some links:

http://news.cnet.com/8301-13512_3-10024280-23.html
http://en.wikipedia.org/wiki/Larrabee_(GPU)

I'm sorry if i misunderstood but did you just saw that a single C2 core runs at around 10w? so that would mean a quad core will only use 40w? Something doesn't sound right there...

Aslo the larrabee cores are as azreal pointed out p54 cores (which are pretty old). There has also been major speculation that the card will have up to 48 or more cores for the high end cards. Those cards are gonna be monstrosities.

Seeing as at 1ghz, it takes 24 of those core to run FEAR (and old game) at 60FPS. If larrabee launches with 24 cores at lets say 3ghz, which is another estimation of the operating speed, we might be looking at around 100FPS. (don't tell me its gonna be at 180fps, since by now people should realize nothing scales linearly) which isn't exactly revolutionary since my 8800gtx could already pull that off.

I'm also curious to see MultiAA performance, Anistropy performance, HDR performance on these cards, till now intel has been awfully quiet...
perplekks45 31st March 2009, 15:30 Quote
They never said the wanted to be King of the Hill...
Coldon 31st March 2009, 15:36 Quote
no that's exactly what they said with their whole: "The GPU is dead" statements, I'm too lazy to find links right now.
perplekks45 31st March 2009, 17:05 Quote
Well, in the future, yes, but defintely not with their first generation of Larrabee.
Turbotab 31st March 2009, 17:14 Quote
I wonder what came first the Atom or the Larrabee?, both are based on a modified P3, in order architecture.
Turbotab 31st March 2009, 17:16 Quote
Quote:
Originally Posted by Coldon
the 300w power draw is pretty much the common estimate, on the web.

here's some links:

http://news.cnet.com/8301-13512_3-10024280-23.html
http://en.wikipedia.org/wiki/Larrabee_(GPU)

I'm sorry if i misunderstood but did you just saw that a single C2 core runs at around 10w? so that would mean a quad core will only use 40w? Something doesn't sound right there...

Aslo the larrabee cores are as azreal pointed out p54 cores (which are pretty old). There has also been major speculation that the card will have up to 48 or more cores for the high end cards. Those cards are gonna be monstrosities.

Seeing as at 1ghz, it takes 24 of those core to run FEAR (and old game) at 60FPS. If larrabee launches with 24 cores at lets say 3ghz, which is another estimation of the operating speed, we might be looking at around 100FPS. (don't tell me its gonna be at 180fps, since by now people should realize nothing scales linearly) which isn't exactly revolutionary since my 8800gtx could already pull that off.

I'm also curious to see MultiAA performance, Anistropy performance, HDR performance on these cards, till now intel has been awfully quiet...

For power consumption the Larrabee cores will be more akin to the Atom, rather than a Core 2, from what I can gather.
Adnoctum 1st April 2009, 03:49 Quote
Quote:
Originally Posted by nicae
The question is: Will they want to make x86 graphics cards?

I agree with your argument. Only if it takes off, and there is every chance that it won't. You only have to look at PhysX to see that it is a big IF.

Intel has only gone down the multi-core x86 route because it doesn't know how to do anything else. Interviews I've read with game developers would indicate to me that they aren't too thrilled by Larrabee, and Intel would surely need to partner with a developer to create that must-have halo product (not a pun, but like Halo did for the Xbox, and PhysX doesn't have) to drive adoption.

And it has to be a great product too, pretty and crap won't sell your mega-$$$ graphics card.
Coldon 1st April 2009, 08:39 Quote
@turbotab:

lets assume each core @ 3ghz has a TDP of around 10w (seeing as a 1.6ghz atom has a TDP of around 4w-8w depending on the family)...

we have 32 of these cores on a larrabee card: 32 * 10 = 320W, now add voltage regulator circuitry and RAM power and other components - lets guess and say around 50w (just a guess seeing as a gtx260 core is rated at 180w TDP and uses around 230w full load)

we're now sitting on 370w for a 32 core card...

lets be optimistic and say it'll use 7w per core (unlikely since they want speeds of 3ghz) - we're left with a card that uses 274w, which is the same as a gtx285...

It's all speculation and guess work on my end tho, so chances are I'll be proved wrong... (I hope not, I want to see intel fail! intel740 style!! )
nicae 1st April 2009, 12:53 Quote
Quote:
Originally Posted by Coldon
@turbotab:

lets assume each core @ 3ghz has a TDP of around 10w (seeing as a 1.6ghz atom has a TDP of around 4w-8w depending on the family)...

we have 32 of these cores on a larrabee card: 32 * 10 = 320W, now add voltage regulator circuitry and RAM power and other components - lets guess and say around 50w (just a guess seeing as a gtx260 core is rated at 180w TDP and uses around 230w full load)

we're now sitting on 370w for a 32 core card...

lets be optimistic and say it'll use 7w per core (unlikely since they want speeds of 3ghz) - we're left with a card that uses 274w, which is the same as a gtx285...

It's all speculation and guess work on my end tho, so chances are I'll be proved wrong... (I hope not, I want to see intel fail! intel740 style!! )
Remember that the Atom is comparable to a Celeron M, capable of running a PC on it's own. Therefore, I would guess lower consumption than Atom. Also, they could use Atom's Z family lower consumption model. Add to that possibly lower clocks and 32nm lithography and we could see something reasonable even for their high-end goal of 48 cores.
Turbotab 1st April 2009, 13:52 Quote
Quote:
Originally Posted by Coldon
@turbotab:

lets assume each core @ 3ghz has a TDP of around 10w (seeing as a 1.6ghz atom has a TDP of around 4w-8w depending on the family)...

we have 32 of these cores on a larrabee card: 32 * 10 = 320W, now add voltage regulator circuitry and RAM power and other components - lets guess and say around 50w (just a guess seeing as a gtx260 core is rated at 180w TDP and uses around 230w full load)

we're now sitting on 370w for a 32 core card...

lets be optimistic and say it'll use 7w per core (unlikely since they want speeds of 3ghz) - we're left with a card that uses 274w, which is the same as a gtx285...

It's all speculation and guess work on my end tho, so chances are I'll be proved wrong... (I hope not, I want to see intel fail! intel740 style!! )


The Larrabee, what a stupid name, was 'benchmarked' at 1 GHz and IIRC, Intel according to the rumours, are aiming for 2GHz. This makes sense given that the broadly similar Pentium M architecture never run much above 2Ghz.
Intel could score big with Larry, with its combo GPU/CPU Clarkdale, especially if they are able to design a chip that offers enough grunt for mainstream gamers and some fancy EIST to turn the GPU off in desktop mode to reduce power consumption.
The Wii demonstrated that if there are ulterior benefits, people are willing to overlook the lack of outright power, well maybe not Bit-Techers:D
steveo_mcg 1st April 2009, 14:09 Quote
Could this be used general purpose processing? Couple of these would make a fantastic code compiling box.
djgandy 9th April 2009, 18:42 Quote
People speculating also need to add in performance to their speculations.

A 3 GHz Larrabee with 32 cores can issue 16 instructions per clock, plus can do fused multiply add, doubling GFLOPS.

3 * 32 * 16 * 2 = 3072 GFLOPS

370W for 3072 GFLOPS. Yes please.

If it could do Fused Multiply Add and another instruction such as Add/Mul (like nvidia's cards) they'd be able to claim another 50% on top of that. This is extremely unlikely however.

A 16 core Larrabee @ 2 GHz will break the TFLOP barrier though, and even the pessimists will agree that a 16 core version at 2 GHz is very likely.


Larrabee what a stupid name? As opposed to going through ever number combination in existence like nvidia? It's a codename btw, its not meant for consumption by the ignorant.

Get over the x86 cores too. They are going to be something like Atom. They aren't meant to be fast, they provide an easy to program platform. The speed comes from the 16 issue SIMD AVX cores (LRB equivalent of SSE). SIMD is no different to all the graphics cards on the market today. All nvidia's cards are SIMD, and ATI's too.

I'm not sure why people get their knickers in a twist so badly and say how Nvidia/ATI are pro's. There's not a lot to be pro about. GPU's are relatively basic processors compared to CPU's. The art is feeding all the cores with the required data.
wuyanxu 9th April 2009, 19:09 Quote
Quote:
Originally Posted by djgandy
People speculating also need to add in performance to their speculations.

A 3 GHz Larrabee with 32 cores can issue 16 instructions per clock, plus can do fused multiply add, doubling GFLOPS.

3 * 32 * 16 * 2 = 3072 GFLOPS

370W for 3072 GFLOPS. Yes please.

If it could do Fused Multiply Add and another instruction such as Add/Mul (like nvidia's cards) they'd be able to claim another 50% on top of that. This is extremely unlikely however.

A 16 core Larrabee @ 2 GHz will break the TFLOP barrier though, and even the pessimists will agree that a 16 core version at 2 GHz is very likely.
sorry to say this, but your calculations are a bit....... well, too simple.

FLOP means floating point operations, and 16 instructions per clock doesn't mean 16 FLOPs per clock. because floating points are hard to do, it may need to break into micro-code and executed in many steps.

the nVidia shader pipelines, however, are optimised for calculating floating point units with all the necessary bits and bobs to do it in one clock cycle, that's the advantage of RICS architecture, and that's why x86 is not the answer to everything.
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums