Larrabee Details
Before
actually mentioning Larrabee, Intel briefly talked about a new
extension to the x86 IA architecture known as Advanced Vector Extension
(AVX). Called a “256-bit vector extension to SSE for FP intensive
applications” you can assume that this is at least partially for
upcoming graphics work on Larrabee.
In
an effort to discuss the benefits of Intel’s upcoming Larrabee
architecture, Intel put together a slide comparing the different
between the current mainstream graphics GPUs and the power of “visual
computing” that you’ll be able to get with their discrete GPU option.
It has some good points listed and basically calls out current
generation GPU architectures as being too rigid in their
programmability and inefficient for non-graphics computing; while
partially true at least the last few steps of GPU technology has slowly
been inching away from many of these issues and more towards a fully
programmable design.
Intel
is addressing Larrabee and its architecture at more than just graphics
and includes new user interfaces, HD audio and video and of course
computational modeling like physics in the realm of their new GPU
architecture.
As
far as the architecture of Larrabee goes, there really wasn’t anything
new discussed: it is still a many-core chip that supports a version of
IA that will obviously be limited when compared to standard general
purpose cores. Gelsinger did discuss the cache system to say that it
will be both effective for integer and floating point operations. In
the past, he says, traditional cache structures have not been very
effective in graphics workloads but that they are addressing that with
unique designs that will offer enough speed and locality of workloads.
Larrabee up close
The coherent cache architecture will allow programmers to dynamically partition it across many cores yet sustain coherency with them as well. This allows for the ability to work on shared data structures, say a large collection of fragments, across many cores in a more efficient manner. Another example Gelsinger describes is the “scatter-gather” feature that allows Larrabee to permute mass loads and stores in order to quickly pull in a large data structure from memory – a necessary feature of modern GPUs.
It’s obvious that Intel sees the problems that existing x86 architecture designs present for graphics models and they are attempting to address them.
Larrabee up close
The coherent cache architecture will allow programmers to dynamically partition it across many cores yet sustain coherency with them as well. This allows for the ability to work on shared data structures, say a large collection of fragments, across many cores in a more efficient manner. Another example Gelsinger describes is the “scatter-gather” feature that allows Larrabee to permute mass loads and stores in order to quickly pull in a large data structure from memory – a necessary feature of modern GPUs.
It’s obvious that Intel sees the problems that existing x86 architecture designs present for graphics models and they are attempting to address them.
If there is one thing that Intel has done very well in the past, it is working with developers and creating compilers and optimization software for their own hardware. We have commented quite a bit on Intel’s purchases of Havok and Project Offset and I think that we’ll find these technologies offered to developers as well in the near future.
Note also that at the bottom of this slide Intel clearly states that both DirectX and OpenGL will be supported in the upcoming software tools for Larrabee.
Closing Thoughts
We’ll have some more analysis from Josh on these topics later today, but for now, I thought I’d just offering a few quick points on the main topics of interest to most of our readers: Nehalem and Larrabee.
Nehalem details were somewhat muted in the fact that we still do not have any kinds of estimates on clock speeds or performance. As I have mentioned several times, technological leadership doesn’t always translate into performance leadership or market share as AMD and there K8 and K10 design can attest to. Intel needs to actually follow through with the new architecture behind Nehalem – it is a large shift from their current lineup and even this IS Intel, there is a chance for only modest results.
Larrabee didn’t change much in this presentation either from what we knew before: no detail on number of cores, speeds, implementations, etc. We did finally seeing the fruits of Intel’s software labor with the discussion on AVX, a new extension for SSE that will start to bring more graphics functionality to x86 IA architecture for Larrabee’s cores. New information on the cache system on Larrabee was interesting but again nothing to get overly excited about since these are problems that we knew Intel was going to have to address in order to even stay competitive with modern GPUs in current applications.
Stay tuned to PC Perspective for more discussion on these topics as IDF proceeds next month. Also, if you haven’t followed all of our Larrabee and Intel GPU discussion, you should check these articles out:
More Reading on Gaming and Ray Tracing:
- Podcast #25 – John Carmack speaks on ray tracing, future of PC gaming
- John Carmack on id Tech 6, Ray Tracing, Consoles, Physics and more
- NVIDIA Comments on Ray Tracing and Rasterization Debate
- Ray Tracing and Gaming – One Year Later
- Rendering Games with Raytracing Will Revolutionize Graphics
- Ray Tracing and Gaming – Quake 4: Ray Traced Project
- Intel buys Project Offset, makers of the Offset engine
- Intel demonstrates ray tracing on ultra-mobile PCs
- Playstation 3 Runs Real-time Ray tracing
- His future’s so bright … he’s gotta ray trace shade
