Tilera preps many-cored Gx chips for March launch

Upstart multicore RISC chip maker Tilera is timing the launch of its third generation of Tile processors to rain a little on Intel's forthcoming parade, and to try to blunt all of the excitement that is building for ARM-based alternatives for servers. Tilera will today begin sampling of its Tile-Gx series of processors. As El …

[Robert Heffernan]

Serious Contender?

If Tilera have really thought out their architecture and it performs well, this could be a serious contender in the server stakes. It takes a very game company these days to release a new architecture, especially considering they are responsible for the linux kernel support, but they are also ruling out Windows as a server OS as well. I wonder if it supports virtualization through the linux kernel. I can see this doing well as a host for small budget virtual servers, say two cores and a gig of ram, on the 100 core model with 64gig ram will give you 50 virtual machines and ram to spare.

[P. Lee]

Actually this is good for ARM

More heterogeneous systems means more linux. More linux mindshare and applications mean more scope for ARM.

[Anonymous Coward]

Two thumbs for innovation and competition!

Hip-hip hooray!

[bazza]

Big hurdles

Tilera have some big hurdles to overcome if they're going to significantly expand their market. The biggest problem is that no-one (relatively speaking) has heard of them. In contrast, everyone knows who Intel, AMD and ARM are.

A significant advantage that ARM has over Tilera is that ARM don't actually make their chips. The very large number of ARM licensees out there works in ARM's favour, because the licensees are the ones who put all the effort into working out what will actually sell, developing a whole ecosystem and marketing it. That strategy worked in the mobile space. And with the likes of Microsoft, HP, and loads of others all sniffing around the ARM server market, you'd have to say that ARM have some *very* big partners helping out. Tilera don't have that benefit.

Also I'd have to say that Tilera may have mistaken raw compute performance for being the key market driver when it comes to servers. Sure, some servers are very busy but a very large number are primarily driving disks and network interfaces. ARM look like they are offering just the right amount of compute power and features for many people's server needs (just like they did in the mobile phone market). Again, ARM can leave that sizing issue largely up to their licensees, whereas Tilera have to work it out themselves. I'm not convinced a large number of cores with presumably not much network/storage I/O will do it for a lot of server operators. And if ARM do decide that some architectural change is necessary to compete with Tilera they could easily develop the architecture and then let their many licensees do the rest.

[Michael H.F. Wilkinson]

Interesting system

The addition of 64bit and floating point is key. I would like to see how these perform on our gigapixel image processing workloads.

[something]

Core is MIPS

Correct me if I am wrong but I think that the core implements some version of the MIPS instruction set. I'm not sure if they have added any special, custom or whatever instructions, but I believe that they didnt go as far as "invent" a totally new ISA.

Not that this bad, just for the correctness of things.

[introiboad]

I believe you are correct, the ISA is derived from MIPS. However, since they had to submit it to the kernel as a full new arch (see arch/tile/ in the source) I assume they changed it enough to be not compatible at all.

[Paul A. Clayton]

Tile-Gx is (V)LIW

Quoting from the ISA manual: "Each instruction bundle is 64-bits wide and can encode either two or three instructions." and "either all of the instructions in the bundle are executed or none of the instructions in a bundle are executed."

[Joe Burmeister]

Hopefully more competition

This is why I think open and free will win. New processor designs come along and we just recompile the repository and use it. Instantly thousands of native applications. The closed world must either get important players to port to try and create critical mass on the platform, or do byte code and take the speed hit compared with native open/free software.

Wintel are pooing their pants about ARM, imagine if it's not just ARM they have to worry about.

Competition is good, and we can just use what's best for the job not worrying about legacy instruction set support.

[Sean Baggaley 1]

I don't think ARM have too much to worry about yet.

The Tilera chips seem to require a lot more power than ARM cores. Fine for servers, which will hurt AMD and Intel's business, but ARM's core (sorry!) business strength is in mobile and embedded markets, not servers. Mobile and embedded customers consider power / watt pretty bloody important.

Even when you add up all the CPUs in all the data centres in the world, it's still peanuts compared to the *billions* of mobile and embedded chips out there. That's why Tilera's pricing starts at $many and rises very quickly to $lots.

ARM do need to get their 64-bit architecture out, but I don't think they're too worried about the server market; compared to mobile and embedded, it's unlikely to become a primary source of revenue for them.

Remember, too, that ARM's philosophy is that their cores can be coupled easily to other _specialist_ processors, such as GPUs and DSPs. Tilera's approach is to assume everyone wants the exact same core lots and lots of times. Both have their advantages and disadvantages, so ARM aren't necessarily going to lose out here. If most of your processing is better done on a DSP or GPU, you don't need lots of 64-bit cores as well: a bunch of 32-bit cores should be plenty to parcel out the jobs to the specialists.

[Richard Lloyd]

CentOS-based...hmmm...

It's interesting that the Tilera CPUs will run a CentOS-based distro, because CentOS itself has never actually supported the Tilera processor architecture of course. I'm not sure I'd want to run a server OS that was a fork of a clone with only Tilera's engineers to run to if there were any issues.

So will Tilera try to push their ports upstream to RHEL and/or CentOS? Perhaps they would willing to becoming the official maintainers of CentOS on Tilera chips (CentOS has actually lost platforms over time, so it's ripe for a new platform to be added)? If neither is true, then I think we're talking about an OS dead-end here.

[Will's]

Single supplier ISA, single supplier complier, single maintainer Linux

I'd not want to bet my company on this, I don't think ARM have much to worry about.

Although, the price point looks enticing I don't think I'd want to be the guinea pig for an unknown compiler and a largely untested Linux port.

ARM have outlived a large number of alternative ISAs, and its not clear that they have always been "better" than those that fell from exhaustion trying to compete.