Brian's Brain

Sep 28 2006 12:18PM | Permalink | Email this | Comments (0) |
Blog This! using:  Blogger.com | LiveJournal |
Digg This | Slashdot This | add to Del.icio.us

Ever since the publication of my 'Instigating a Platform Tug of War' article and blog series a year ago, I've been closely following developments in the embryonic CPU co-processor field. And there've been plenty of developments to track, most recently at this week's Intel Developer Forum.

My October 15, 2005 article focused much of its wordcount on broader applications for modern graphics accelerators, beyond graphics. Writeups on this topic in online forums such as AnandTech, Ars Technica and Slashdot include the following:
ATI eyes audio acceleration on the GPU
High performance FFT on GPUs
NVIDIA's PureVideo HD: HD-DVD Playback on the PC
PS3 Client for Folding@Home Debuts, ATI GPU Version Soon
PeakStream unveils multicore and CPU/GPU programming solution

ATI Technologies is unveiling news in this area tomorrow; unfortunately a schedule conflict precludes me from attending their event:
ATI's Stream Computing on the Way

And indicative of the tug-of-war between the CPU and GPU, exemplified by AMD and Intel's coming-soon quad-core projects:
Add Another Core for Faster Graphics

What about physics processing? AnandTech has been regularly revisiting AGEIA's PhysX chip as its drivers, along with the applications that allegedly exploit its capabilities, get fine-tuned. Here's the latest updates, from mid-May:
BFG PhysX and Performance Updates
and early September:
PhysX Performance Update: City of Villains
along with something from Extremetech:
How Well Does Ageia's PhysX Accelerate Game Physics?

I've already provided you a link to John Carmack's (id Software) opinions on dedicated physics acceleration; now another titan of the graphics industry, from Epic Software, weighs in with his thoughts:
Tim Sweeney ponders the future of physics cards

AI (artificial intelligence) is another MIPS-consuming function performed by modern games, so it's perhaps not surprising that, following in AGEIA's footsteps, a silicon startup has emerged that aspires to steal this task away from the CPU:
New chip promises better AI performance in games
Chip Promises AI Performance in Games

And not even packet processing is immune from potential dedicated hardware handoff. Witness Bigfoot Network's 'Killer' NIC, with a dedicated processor onboard:
Network Card for Gamers - Uses Linux to Reduce Lag
Killer NIC Hands-On Testing
BigFoot Networks Killer NIC: Killer Marketing or Killer Product?

AMD and Intel have uneasy coopetitive relationships with companies like those mentioned above; on the one hand they're supportive of anything that'll advance the capabilities of the PC platform, on the other, they don't want to back anything that'll significantly slow or stall the long-term CPU upgrade treadmill. Consider this tightrope walk in the context of recent weeks' news.

Back in early August, AMD announced that it was amenable to licensing the cache-coherent variant of its HyperTransport core-to-core and CPU-to-CPU interconnect bus (coverage from Ars Technica and Slashdot). Recall that, by virtue of their integrated DRAM controllers (among other factors), modern AMD CPUs don't provide outside-world access to their front-side buses. Still, access to the full-featured HyperTransport bus is performance- and functionally-preferable to secondary and tertiary alternatives like PCI, AGP and PCI Express. Potential beneficiaries of the Torrenza initiative include not only the dedicated-function chips mentioned earlier in this writeup but also general-purpose FPGA suppliers such as Altera and Xilinx, along with intermediary-function chip providers like ClearSpeed.

Intel's response, unveiled earlier this week, is a two-prong approach. Generically (but not cache-coherently), the company will work with the PCI Express standards body, in an effort code-named Geneseo, to add spec extensions that are amenable to application accelerators. And more specifically (and speedily), Intel has granted Altera and Xilinx highly-prized, rarely-allocated licensing access to its front-side bus patents.

This'll be fun to watch! How are you placing your bets?

Followup: Based on these AnandTech and Extremetech reports, it looks like ATI rehashed information (Folding@Home and PeakStream) that I've already shared with you at their Stream Computing press event on Friday; no new news to report.

Followup II: The ATI GPU-cogniscent Folding@Home client is released.