Intel Xeon Phi accelerators: any potential for running audio DSP?

Configure and optimize you computer for Audio.
Post Reply New Topic
RELATED
PRODUCTS

Post

http://news.techeye.net/chips/intel-xeo ... out-in-may
Intel is planning to release two new Xeon Phi co-processor cards in May.

In addition to the already announced Xeon Phi 3100-series products due in the first half of the year, it seems that Chipzilla will also unveil new 5100- and 7100-series coprocessors starting this May.

There will be the Xeon Phi 5120D, 3120A, 3120P, 7120P and 7120X. It is not clear what the exact specs of co-processors will be but all of them are based on Knights Corner.

The Xeon Phi 3100-series was supposed to deliver over 1TFLOPS peak double precision performance, which is about as floppy as it gets at the moment. But the Xeon Phi 5120D, 7120P and 7120X coprocessors are even floppier. At present, Intel does not have exact shipment dates for the new products, but expects them to become generally available from May 1, 2013 to July 31, 2013.

The Intel Xeon Phi coprocessor 3100 family is designed for a boffin on a budget as it can run compute-bound workloads such as life science applications and financial simulations. The Intel Xeon Phi 3100 family supports up to 6GB memory at 240GB/s bandwidth. It also has a series of reliability features including memory error correction codes (ECC).

The whole family will run at less than 300W thermal design point (TDP) envelope. The Intel Xeon Phi coprocessor 3100 will cost about $2000.

The Intel Xeon Phi coprocessor 5110P features 60 cores with 4-way simultaneous multi-threading technology and 512KB L2 cache per core, provides additional performance at a lower power envelope.

It can manage 1.01TFLOPS double-precision performance with the wind behind it, and supports 8GB of GDDR5 memory at a higher 320 GB/sec memory bandwidth.

It needs 225 watts TDP, and is a passively cooled Intel Xeon Phi coprocessor 5110P. It is designed for denser computing environments such as the movie industry and energy research. You will be able to pick one of those up for $2,649.
I imagine plugins would probably have to be coded with these in mind - as is currently the case with Nebula and CUDA. It seems feasible that Intel - who supply compilers etc - could perhaps make this a lot simpler by making it possible for DSP code that runs on vector processing units on current CPUs to run more easily on these massively parallel Knights Corner chips.

At the price quoted, these aren't going to be for everybody but anyone who is currently using dual Xeons is already looking at a fairly expensive machine. Intel might not be interested in pricing these like "enthusiast" (gamers) graphics cards but that could be what is needed to find a market for video and audio types.

Post

Well, it may help a little but its parallel so wont be a huge advantage except for some FX rather like Nvidia CUDA on graphics cards. Audio is better served by serial processing.

Post

Running lots of Eq, compressor or synth instances is definitely helped by many cores in parallel. The practical limitations of CUDA have related to the latency that this system has compared to running the DSP code on the CPU. The other limits are the amount of memory and bandwidth available within and to the graphics sub-system. Nvidia are now trying to park large amounts of very fast access memory directly on top of their GPUs - they are talking Tb/s bandwidth to the local memory. The other approach is a move to a unified memory architecture where for example Intel are giving the Haswell graphics systems access to main memory.

Post

Fairlight has a system which uses a DSP accelerator card for basic stem effects like EQ and compression. Still, one core rules them all so to speak, and serial processes require more serial processing.

Post

Most DAWs are processing operations in a massively parallel fashion. Otherwise you'd only see CPU activity on one core. Only one effects chain needs to be serial, and serial operations like a master channel can be done after parallel operations like the separate tracks. It adds slightly to the latency, but this is how your machine is handling it anyway, as a bunch of serial + parallel+ serial operations.This is why 16 cores perform better than 8 cores perform better than 4 cores. If DAWs facored serial over parallel, then a 4.5 Ghz dual-core processor should outperform a quad-core 2.66 Ghz CPU...

it won't.

Most of the DSP processor cards have multiple DSPs. Some, like the Pro Tools DSP cards, have a *lot* of DSPs.These DSPs are processing different items as separate threads and some plugins can run multiple plugins to a DSP, based on their DSP load. Once again, serial+parallel operation.

The x86 cores on the Xeon Phi would require use of their custom API to access, but the x86 compilers are quite mature.

However... the Xeon Phi uses 512-bit wide registers (which would need to be converted up and down from 64-bit that most audio software uses), and uses GDDR5 memory so it is likely to suffer from the same latency problems that running VST effects on the GPU suffers. I actually discussed this with John Gustafson of AMD in an interview this week.

Post

Any new angle on this? Theres a 57 core phi here for CAD$140!!

Post

Nope, really been no movement in this area.

Post Reply

Return to “Computer Setup and System Configuration”