Tim Sweeney likes Larrabee

Posted Aug 13, 2008 5:19 am GMT

What you all failed to realise is that Intel is not aiming at the graphics market with Larrabee. They just understood AMDs strategy, and it got them scared. Think AMD Fusion. Just to make it clear:

Fusion is the union of CPU and GPU power, but since they cannot make it right away (lots of research to go), they divided it into three phases: one, CPU and GPU working together; two, CPU/GPU integrated; three, "GCPU", or CPU and GPU on one chip design.

Now look at phase one and think: AMD launched the 780G chipset with a Radeon HD3200 integrated, waay more than they need to take on Intel's GMA kingdom. Also, they're advertising the Spider platform, a CPU+GPU+Chipset concept that just doesn't offer anything new YET.

That said, AMD is on phase one of Fusion, a very real concept that will make ANY traditional x86 CPU obsolete. Intel has the best x86 processor on the market, and can push it a lot more if needed, but it still is a x86 processor, and they have NO graphics/stream processing experience.

Now THAT scares Intel. THAT is why they're going for graphics now. They know it's just not a profitable market, but they have to enter it NOW to be competitive in the future. Who cares about nVidia? They make and will keep making great graphics card solutions, and probably remain competitive against AMD/ATI.

In my humble opinion, it would a good time to buy AMD stocks...

• agree 1
• disagree 0
• = +1
• Sign in to rate this comment

Posted Aug 11, 2008 3:05 pm GMT

At the end of the day the pc & all that is integrated into it is limited by it's architecture. It's still based on the dated IBM. Surely if someone took the initiative to design a whole new cpu the possibilities would be amazing. Then they could really go to town on graphics.
Yes, everyone would have to upgrade, all new hardware & software, etc, but look at the business this would generate for all the hardware & software companies. How many "cores" are you prepared to keep shelling out for? 32? 128? 512??? This multiple core BS will have to stop, if only for cost reasons. Due to heat limitations te core speeds can't really get any faster.
Fix this problem first & the rest would be much easier. After all, all "Windows" really does is mask these problems & associate drivers so we don't have to mess about with xms memory, irq assignments, mem' access, etc'.
Instead, we all complain about Microsoft & the monopoly they have but because of these design flaws we really have no choice & MS will just keep releasing the same program / OS every few years which we will all end up having to buy.
At the end of the day you can paint a 3 legged chair any colour you like but it will still only have 3 legs.

• agree 2
• disagree 2
• = 0
• Sign in to rate this comment

Posted Aug 11, 2008 11:25 am GMT

I'm not completely sold, myself. While there are great things you can do, Larrabee will need a massive memory pipeline and high speed memory to work quickly because everything but cache is on mainboard memory. That is GOOD if you want raytracing which needs entire scene access, but not not terribly useful for polygon rendering (because you don't need the whole scene).

The main optimization in ditching the fixed function pipeline is you can parallel-ize a lot more, but that is essentially what shaders do, and the more functionality that goes into shaders, the less the fixed function pipeline is doing. Things like geometrical culling can vastly simplify a scene, while culling (at least in the backface way) is not practical for ray tracing (since surfaces can be curved).

The way Larabee seems to be approaches scalability is to break the scene down and throw the pieces on different processors (for example, if you have 4 Larabee GPUs, it divides the monitor in 4 pieces). The sections then sort the scene and render individually (using its cache to form sort-order of in-memory scene objects). Carmack's criticism of this is that compared to rasterization, it's terribly inefficient and they make up for it by throwing more processors and a higher frequency rate. If GPUs gain the ability to quickly tessellate primitives (something Carmack is also skeptical of), they could have ray-trace like objects like spheres that scale to processor.

• agree 1
• disagree 0
• = +1
• Sign in to rate this comment

Posted Aug 9, 2008 9:52 pm GMT

I think the future of GPU and or CPU is that CPU, GPU, and Sound Card Processors will be put in one small chip.

• agree 0
• disagree 0
• = 0
• Sign in to rate this comment

Posted Aug 8, 2008 10:52 pm GMT

amazing, some coherent comments on something i am very interested in now. cool!

• agree 0
• disagree 0
• = 0
• Sign in to rate this comment

Posted Aug 6, 2008 9:33 pm GMT

Larrabee is a x86 capable part for pci-express, meant to compete against ati and nvidia, but it aint a graphic card only, unlike nvidia and ati cards which sport a rigid (soldered) graphic logic in the core + some narrow space for programming (the foundation of all gpus), Larrabee can run costume drivers for the whole render pipeline ( the order in which things are processed before display).

The reason why some sli or crossfire games dont scale well, or some games run better with nvidia or ati parts, its because, the developer cant really decide what to render first, so, at best, nvidia/ati have to wait for a game to be near release and have their scientific monkeys profiling each game, so most of the tiame game profiles appears months after release.

See larrabee as one of those chips for college students with empty logic, it has the best of a software renderer but it doesnt run in windows, the software layer runs in the chip itself and uses opengl/dx10 to show results. With such a feature for example you could update a PS3 or xbox 360 to shader model 5.0 and so on. It can be an adobe accelerator a folding home for the masses, why? it uses actual x86 code, the same we use to make apps, so no CUDA learning needed.

If you had read tech sites like ars or theinquirer for the past 3 years, you would have expected this move, Nvidia is no longer targeting ATI, they're going for Intel, not for their market share, but for the relevance the market gives to cpus, lets say, if you sell a gpu (next 55nm GTX280) at the same price of an upcoming eight core mid range , and concentrate in super accelerating a few key applications like photoshop or HD video encoders, then the consumer from a not so distant future driven by real life earning needs would think "why a $400 32 core and a GTX 240*, when I mostly encode and post 30min design tutorials in youtube and make money with adobe at my uncles print shop"; then instead, he buys a $200 16 core and a $400* GTX 280.

So the main stream conservative market no longer sees Nvidia as company for games or watching blu ray movies, the commodity becomes mainstream necessity (Intels playground). Nvidias worst nightmare isnt AMD which sports a market cap of $3.8 billion with 5 billions in debt after they bought ATI. Nvidia already enjoys a $10 billion market cap and only need discounts rebranded stuff and broader "the way is meant to be played" support to slow down ATIs earnings

And Yes, Nvidia will win Larrabee`s 1st round, since larrabee is a proof of concept, on the other hand theres nehalem (and 8 to 84 core with a larrabee inside), Intel with 180 billion market cap and 9 times AMDs net income; this crappy vga maker worth 18 times nvidia not only sells cheap P4s united by an awesome architecture (coreduo) at higher prices than AMD, its also has been the world biggest gpu maker for a long time, yes, they make crap gpus, these are for everyone not some like us.

How about if you give the same gpu tech chained monkeys 10 times the funding and permission to use 8 or more cores from the next cpu in order to create a gpu, well, with 32mb+ cache access at 45nm and up to (84) 3.8ghz cores then, after 5 years of paper and wafer cooking (nehalem dev cycle), theyll give you a reason to think "Hey! Nvidia didnt had alien technology applied in their cards and drivers, It was just about size, design and gigahertz, why would I buy an nvidia or an ati card?, when I can spend $400 in a 32 core cgpu that can later "slicross" in my new multi socket $130 mobo, yes some cores go for Havok physics". This my friend its THE WORST NIGHTMARE for Nvidias CEO and co-founder Jen-Hsun Huang.

I have almost none consumer innocence for this or that brand due to my marketing career and geeky habits. My point, Nvidia is doing what its left for them to do, play hard and show no sweat, with their outstanding PR and marketing dep, everyone thinks they re in safe harbor. AMD made some investors go suicide and went red numbers, knowing this Intel plan, they bought a 6 billion "out of Jail" ticket. Now they can make the same as Intel and extend their fight 10 more years, the most brilliant hard decision.

Nvidia, the respected Nvidia, needs a really rigid 0 error, road map and strategy to survive, they ve been doing this with near perfection for the last 2 years, until they made a $1200 million notebook chip fiasco blamed their provider (TSMC) for a design flaw which was their fault, and made HP to pay 50% for their own mistake. Interesting times to come, never the less my 8800 still deserves a life achievement award.

• agree 8
• disagree 0
• = +8
• Sign in to rate this comment

Posted Aug 6, 2008 3:32 am GMT

From my extensive reading about Larrabee it's not an integrated graphics solution. It's actually a powerful GPU made up of multiple Pentium x86 processors that have been slightly modified to handle OGL/DX and software based rendering. It's designed as a Graphics Card solution.

It's interesting. Intel hired Michael Abrash - the guy responsible for optimising code for Doom and Quake as well as working on the X-Box.

The software based renderer could open up all sorts of features in games that we've never seen before. It will give game designers more freedom.

Ultimately it will give consumers more choice. If they manage to get it into a dominant games console such as the next X-Box then I think that could make it a success. If not then they'll have a harder job of getting game developers on-side with the exception being the wide-spread familiarity with x86 based architecture.

Good luck Intel.

• agree 5
• disagree 0
• = +5
• Sign in to rate this comment

Posted Aug 5, 2008 10:12 pm GMT

oh, and by the way, heres a good look at larabee for those that want to see exactly whats going on (thats assuming that everybody here isnt a stupid git that doesent know what integrated graphics are):

http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3367

• agree 1
• disagree 2
• = -1
• Sign in to rate this comment

Posted Aug 5, 2008 6:40 pm GMT

having read an article from another tech website, I'm personally impressed with what intel is developing.

Larrabee could be the NEXT way forward in GPU, especially when thinking fully programmable multi GPU compared to the x2 or crossfire/SLI ideas that ATI/nVidia provide.

I also agree with a-sa and think intel actually sell more IGPs than ati and nvidia put together.

• agree 1
• disagree 0
• = +1
• Sign in to rate this comment

Posted Aug 5, 2008 4:11 pm GMT

To the guy talking about a good integrated gfx solution. Have you not noticed that ALL consoles to date have had integrated cards, "and its worked out pretty well so far."

• agree 0
• disagree 0
• = 0
• Sign in to rate this comment

Posted Aug 5, 2008 4:02 pm GMT

This was the shortest interview I've ever read. Wow...

You only get three questions with Sweeney, and they have to be delivered as Haikus, like this:

Do you think that Larrabee
would make sense for next-
generation game consoles?

He's a strange, strange man who makes brilliant games.