Archive for August, 2007

Tilera – Let Loose the Press Releases!

August 21, 2007 at 7:28 am · Filed under New Architectures

There’s been so much being said about Tilera over the last couple of days. Almost None of it dealt with the juicy facts we would like to hear. I’ve been doing some digging and managed to find out some details about the company. Before I get on to that though, I’ll do an in-depth review of the architecture:

It’s a homogeneous array of lightweight microprocessor cores with nearest neighbor interconnect. Oh, and a great programming environment too. Probably. And customers are begging for it. Oh god they are, why I heard that when they announced the chip was shipping, a crowd surrounded Tilera HQ in Menlo Park. The staff had to lock themselves in while desperate purchasers clawed at windows and doors until their fingers bled. They were chanting slogans like “The microprocessor is the new transistor!”, “Double our cores!” and “Freebird!”. They say a travelling salesman in Texas gave his teenage daughter in exchange for just ten of these wonderful devices.

Hmm… Enough of that craziness, let’s get down to brass tacks:

This article tells us that Tilera have $40m in funding. From this press release and this one, we learn that Tilera have managed to round up virtually all of this $40m in the last eight months. It seems though that they got some money in 2004.

Rob Chandra, a managing partner at “Bessemer Venture Partners”, invested in 2004, juding by what’s here. I guess this must’ve been the funding they used just to get funding.

Columbia Capital apparently led the Series B financing, that’s the most recent one that landed them $20m.

The archicture comes from RAW processor work done at MIT starting over a decade ago.

So what you have is a ten-year old academic project to which $40m, eight months and a marketing team have been added.

My questions to you regarding Tilera are these:

  • Do you believe that a start-up company with $40m can get to where they claim to be in eight months?
  • Do you believe that a startup can emerge with an architecture that has never been proven in battle and succeed, using only a $40m war chest?
  • How much of this $40m do you think is being spent on marketing, given the hype that is being generated?
  • Do you believe that this company has been able to develop a massively-multicore software development environment that is robust and productive, when that is generally recognised to be the grand challenge of our age?
  • Do you think it might be possible that the investors are hoping to pump Tilera up, then transfer ownership to the smal investor via an IPO, à la Mathstar?

I mean, I know that TSMC, Tilera’s fab, are involved in the investment somewhere along the line, so that might lead to a saving in fab costs. But can you even fab a chip for $40m these days, never mind run a company and fab a chip? Barry West of AMI semiconductor says

Estimates for the total cost of developing a complex 90nm ASIC are between $20-$30 million.

And they could be expected to be higher for a 65nm process! So they’ve just about got the money to squeeze out a chip, but what about all their other expenses. What about software development?

To cut a long story short, Tilera are selling us a very implausible tale. It’s too early to conclude of course. Nevertheless, if I were a betting man (and the Lord knows I am) I would bet on Tilera’s eventual failure.

Update:

It wouldn’t be easy to disagree with the points Joe from scalability makes here. He’s right that not being IA isn’t as critical in Tilera’s targeted markets.

There’s a bit more information here:

According to this there are presently 10 customers.

On the subject of the funding, I think it was noted with surprise how soon it was after series A funding that they secured series B funding, something like a month. Anyway, I’m working from press releases here, so if anyone knows more, then please share :-)

I’ll hold my tongue for now though till I see more new information

Update:

This article from Jon Stokes at arstechnica is a good overview, highlighting the fact that the chips are 90nm.

Comments (6)

The Cell Processor – Where’s The Beef?

August 17, 2007 at 7:58 am · Filed under The Cell

The amount of hot air blowing about on the subject of the Cell Processor is incredible. To echo Walter Mondale: “Where’s the Beef?”

(As an aside, that’s got to be one of my favorite Simpsons quotes, one where Homer suddenly has a deep and inexplicable knowedge of the ins and outs of the 1984 democratic primaries:

Marge: I saved this newspaper from the day Lisa was born.
Lisa: “Mondale to Hart: [confused] Where’s the beef?”
Bart: “Where’s the beef?” What the hell that’s supposed to mean?
Homer: Heh heh heh heh heh. “Where’s the beef”…No wonder he won
Minnesota.

)

OK, so the Cell Processor is in the Playstation 3. We all know that.

OK, so the Los Alamos’ Roadrunner supercomputer was announced as the first cell-accelerated supercomputer. A lot of us will know that. A question for anyone out there though: Where are the double-precision SPE cell processors at? Because they sure as hell aren’t in the real system yet. By real system, I mean the one that exists in the data center, and not in press releases. You’ll notice that they’ve not yet passed through the ominous-sounding “Phase 2″. Phase 2 is a technology refresh and assessment of the final system. So by the sounds of it, despite the press releases, they’re still not committed to having Cells in the final system at all.

So despite all the hype, I want to ask the cell: what have you done for me lately? What is really going on, and is anyone really doing anything with this thing. I mean, really. At this point I’d like to invite PS3 fanboys not to comment. Your comments will not appear. No-one wants to hear about Folding @ Home. I’m not about to try to argue on the merits or otherwise of the Cell’s compute architecture, or its performance potential. I want to axe you this: Is it viable commercially?

If you want to use a cell processor for accelerated computing you have two options:

1. Muck around with a PS3, or

2. Buy a proper cell compute board from Mercury/IBM.

1. is fine if all you want to do is evaluate the technology, but it’s not going to do the job for any system that has serious memory capacity, bandwidth, latency needs.

So, if you’re doing serious compute work with Cell, you’ll need to get one of Mercury/IBM products. Mercury have a couple of Bladecenter blades and a PCI Express compute card. IBM have the Qs20 blade. The obvious similarities to the Mercury stuff, this press release and this forum post lead me to the conclusion that all non-PS3 products using the cell on the market are being sold by Mercury (though IBM may be taking a share in some revenues).

Now, I haven’t even bothered to try fishing through IBMs SEC filings to try to find the Cell revenues. I don’t think I’ll find them, big companies don’t work that way. They don’t have to go into that level of detail, so they usually don’t.

Mercury on the other hand have been surprisingly candid in their latest 10-Q. Fishing through it I’ve pulled out Cell revenues and R&D costs for the quarter ended March 31st 2007. Because Cell revenues made up 10% of revenues, I’m assuming that they also made up 10% of revenue cost and operating expenses for which no specific “Cell” references were made in the 10-Q. I think this is fair, if anything it might well put costs and expenses lower than they are in reality. So for the cell, for the quarter ended March 31st 2007, we have:

Revenues: $5.6m

Cost of Revenues: $2.5m (est’d)

Gross Profit: $3.1m

Operating Expenses:

Selling, General & Administrative: $2.1m (est’d)

R&D: $5.3m

Other Costs: $0.2m

Net Profit: -$4.5m

So if my logic holds, all general-purpose cell processor revenues, outside of PS3 sales added up to little more than a measly $5.6m, and those revenues led to a whopping $4.5m loss.

I can’t split the figures for the cell out because they’ve not filed the 10-Q yet, but the bottom-line results for the most recent quarter, show that things haven’t improved at Mercury in the last quarter.

So, if the Cell is all that, then why aren’t Mercury making money from it?

Comments (1)

The Transputer’s Legacy

August 17, 2007 at 5:48 am · Filed under History

This piece is great. It tells a lot more of INMOS’s story than you’ll get from their wikipedia page.

I’m too young to remember the transputer from the first time around, but a lot of the old salts will tend to get all wistful, and stare into the distance when they recall the transputer. It was the great almost was in the accelerated computing days of yore. That was before microprocessors started evolving towards being superscalar and out-of-order, and the transputer was crushed.

I think the thing about the Transputer and INMOS is that it shows how this industry is different from many others. A computing company can go off like a firework, and be a complete financial disaster, but still leave big marks on the landscape. INMOS’s failure sowed the seeds for much more activity, and so from a wider perspective their failure benefited the industry as a whole. A noble failure.

Transputer architect David May’s latest venture is XMOS. It’s tempting to view the proposed technology as an answer to a question nobody asked. Will this company contribute anything to the tech economy, or generate a meme or two like INMOS did? Or will it be another Mathstar?

Leave a Comment

Latency Worries For Market Feeds – Is Hardware the Answer?

August 17, 2007 at 3:49 am · Filed under Acceleration Appliances

I just read a good article by Henry Young, the director of the London and Singapore-based TS Associates.

It’s a good introduction to the issue of low latency in market feeds to financial institutions. Financial houses have automated algorithms that trade in financial instruments and their derivatives (like options). With GHz clock speeds on the microprocessors, the simpler algorithms work in jig-time. However, your trading algorithms are only so good as the data they are working on. You want your algorithm to have access to the latest market data, and the there is a drive to push latency down as far as it can go, and time savings are in the millisecond range. The speed of light is one of your enemies in this game, so financial houses will pay a premium to co-locate ther servers at the source of the market data. I can’t imagine what this game will do for real estate prices in the area, and where it might end! Will Manhattan one day be a glimmering ghost city? Populated only by humming computers and their attendants.

Some companies are looking to lower latency by using hardware acceleration. Exegy are one, and they offer customers an FPGA-based appliance that claims to be able to speed some things up by an impressive degree.

As I understand it, TS Associates‘ approach is different to Exegy’s. Latency is the big problem today, and Exegy are proposing to solve it with a radical new approach. TS Associates’ approach seems more measured: They recognize that latency is the key problem, but also that it is in the critical path. Companies will be reluctant to take the risk of putting an unproven technology (which from their perspective is a bit of a black box) into a key area of their business. TS Associates propose therefore an FPGA-accelerated appliance that monitors latency. This way they can offer customers an entirely new product that is off the critical path. Adoption would be lower risk for customers, and it would be TS Associates’ foot in the door, their chance to become the financial houses’ trusted supplier of FPGA accelerators for latency-related issues.

I’m looking forward to seeing how Exegy and TS Associates fare over the coming months! Does slow and steady win the race, or does fortune favor the bold?

Leave a Comment

Mathstar Wobbling – The End Isn’t Nigh Though

August 16, 2007 at 11:02 am · Filed under Doomed Architectures

Mathstar.

I presume you’ve heard of them, otherwise why would you be on a blog post that discussed them?

They claim to be the FPOA leader. Imagine that, the FPOA leaders! That’s because no-one else is stupid enough to invest their money and careers in such a feeble idea.

These guys have gone to the market now and got tens of millions of dollars more than once with the most water-headed concept I’ve ever seen:

They think they can dislodge the multi-billion dollar FPGA industry by pulling some half-baked architecture out of their asses with no programming model worth speaking of.

The Mathstar people don’t annoy me, they’ve maybe even done a noble thing by putting their careers on the line for an idea.

The investors annoy me a little more, because they seem to have money to throw away at a stupid idea.

What annoys me most though is the number of times people in the industry, who should know better, talk about the Mathstar devices as if they are actually part of the hardware acceleration landscape.

They’re not. They are a joke sideshow that no-one has bought in any serious numbers. It is a zombie architecture, and people are only talking about it because of the millions being spent on little adverts all over the interweb.

Check out their share price and revenues here. The share price is on a constant downer, and the revenues are just noise hovering above zero. The emperor has no clothes! Will this architecture just die already!

The thing is that it won’t though. They’ve got a reasonable amount of cash in the bank. Maybe enough to update their architecture, but then again maybe not. If not, then we can expect them to spend it all on marketing their terrible device, The Arrix. With $30 million dollars this could go on for years.

They’re doomed in the end though. Doomed.

Comments (2)