Product was not bad. Specialized processors need time to gain traction. Look at how long Intel's own SIMD instructions took to gain traction. They got a free ride from the rest of the CPU. Reasons are many, convincing programmer's to learn new tricks, reluctance to try new technologies on the part of management and performance of regular processors catching up.
Hopefully their mesh fabric will be put to good use in Ezchip.
But this an indicator of a major problem in teh chip industry. How do you fund long gestation projects ? Are we stuck only with the existing dominant player's innovating new architectures. Like that is about to happen ! If Intel feel Knight's landing does not have priority over a Broadwell i3, then the world has to wait for 1-2 years. This is simply not acceptable.
It is a pity that it is easier to get funding for a company that can change the encoding format of an emoticon than getting a radical CPU arch. funded !
Product wasn't bad, but it didn't get traction against Broadcom (XLR/XLP/XLP2) or Cavium (Octeon1/2/3).
Some of that might have been that their mesh architecture meant it was substantially more difficult to use in a packet-processing context: Octeon and XLP had no issues with thread/core placement; for TileGX getting decent latency was hamstrung by the I/O sub-system to start, and then you had to consider the routing latency across the mesh when positioning threads.
Their "developer friendliness" was also perhaps an issue. They did the right thing in making the basic Linux and compiler ports open, but their eval systems were expensive and hard to get. The advantages of their Zero-Overhead Linux were whittled away by recent advances in the stock kernel too.
Still, they've haven't had a lot of runway with the Gx chips. With the life-extension from EZchip, they might start to see more success, even if that doesn't benefit the original investors and shareholders. If they follow the herd down the low-power datacenter path, their architecture might scale better to higher core-counts than their competitors, particularly if they start putting some I/O into the middle of the mesh. They might need to switch out the instruction set though ... (sigh)
@gsmd, thanks for the ARM shout out!. Given Tilera's approach of aggregating ALOT of modest level performance CPUs the A53 would probably represent a better approach.
I think this is a great purchase for EZChip, especially at that pricing level. Their use of a proprietary processor architecture forced a path for developing turnkey solutions. In that regard, Tilera does have some design wins at a number very significant customers including communications and embedded. The technology is very credible technically. The combination of Tilera's interconnect EZchip's NPUs throws up some interesting possibilities for next generation products.
Tilera lured Andrew Rava (ex Cavium VP sales) out of retirement, appearing this year on the Tilera management page. The purchase by EZChip may, therefore, represent a very recent change of company strategy.
As a hardware guy, it is disappointing to see an exit like this. It is hard for hardware oriented start-ups to raise VC money and this type of transaction doesn't help change that situation. This industry, particularly in the US, has thrived on hardware innovation initiated in small companies and I worry about the "so whats" in the long term.