ARM, ATIC, TI invest $48M in server chip startup

8/16/2010 2:41 AM EDT

Is this startup a contender or pretender? Thoughts out there?

8/16/2010 12:16 PM EDT

Conventional wisdom in the server world suggests that the most bang for buck and bang for watt comes from faster 64 bit architectures. However, in the early days, Google went directly against that conventional wisdom and build server farms out of masses of cheap motherboards. I don't know if they still do that, but I suspect their reasoning still holds, which would bode well for Smooth-stone.

While much of the world still looks at to absolute numbers for clock speed, power consumption and data thru-put, what's really important are the ratios. If we go back to seemingly-ancient basic economics and apply the concept of the point of diminishing return to the the ratios of performance per watt, performance per dollar and even performance per unit of space, we may very well see that an ARM-based server farm can out perform a top-line x86-based server farm.

8/16/2010 3:59 PM EDT

Of course, this assumes that Intel/AMD don't invade this space. Given Intel's incredible advantage in actually manufacturing stuff at high volumes I would be hesitant to bet against them. Are people forgetting the other start ups some years ago that were going to blow away Intel/AMD in lower power computing?

8/16/2010 4:24 PM EDT

I agree with most of the comments made here. I believe there are many areas of opportunities in datacenter market. The announcement doesn't say much on what their solutions are, so it is a guessing game for most us commentors!

Optimizing server performance at lower power consumption in servers seems obvious as an area of opportunity. How ever, the history of startups is rather checkered in this segment; even funded startups like 3Leaf Systems have folded. So Smooth-Stone must really provide a strong business case, differentiation and a clear segmentation of their solution.

Processors exploiting the digital power architecture of server motherboards for efficient power management would be a good start. Second, server virtualization is here to stay along with storage virtualization, separately or in a common network fabric like FCoE though this latter one is still in a a growth phase. Given this scenario, virtualizaqtion-optimized processors along with API's for communication between hypervisors and network management systems would be something to wish for. This last feature would truly provide the most optimal server scenario for hardware-optimized virtualization.

Third, there are a number of companies tracking energy usage both at power outlets the servers plug into as well as processors at the mother boards. Some of these energy monitoring tools can benefit if there is a move to standardize server motherboard architecture for energy efficiency.

MP Divakar

8/16/2010 4:38 PM EDT

While ARM has a nice architecture for a large set of low power processing requirements I'm skeptical about it use in a general purpose server role.

MIPS processors perform in a slightly higher envelope that ARM processors but deliver substantially more performance.

The same can also be said of PowerPC processors that perform at a marginally higher power envelope than MIPS processors but deliver more performance on more computationally diverse loads.

Also both AMD and Intel have processors that are quite impressive on a performance/watt metric.

I can see entry level low-to-moderate demand web servers as a space for ARM, but what else?

8/16/2010 5:45 PM EDT

Traditionally the relevant metrics for the datacenter market were raw performance, and performance per dollar and per watt. Many low-power offerings (Transmeta etc) failed because Intel delivered relentless improvements, mostly by driving performance up. In recent years, though, that trend failed (we're stuck at 3GHz, to put it crudely); the transistors promised by Moore's law go into multiple cores. Once that trend is established, and we finally work out how to apply the distributed systems in web and database roles, it makes sense to continue past the point of a small number of high-performance processors into compensating with large numbers of cheap and low-power processors of middling-performance. The new datacenter architecture is tricky but doable: massive caches, load balancing, distributed databases in the NoSQL style, etc.

Already today ARM vendors offer low-end ARM microcontrollers that essentially provide the entire System on a Chip for under $1 a piece. The trick for vendors like SmoothStone is to put the savings into creative front-end that will make the whole thing work in spite of the low performance of the CPU.

8/17/2010 1:16 AM EDT

I agree with all of the above comments. To me, Smooth-Stone is either the next Transmeta (a failure) or NexGen (a flop) or Cyrix (a goner) or....a possible star?
Perhaps Smooth-Stone is ARM's alter ego. I assume they are making next-gen server processors, based on ARM's technology.
My question: Is $48 million enough to survive--even in the short term? Look at the design and verification costs alone.
This one will take deep, deep pockets.

8/17/2010 7:43 AM EDT

It is nice to finally see an article on embedded systems. My first attempt at building an embedded sys was early in 1999 although I’m not an expert, i was able to build some nice systems
routers, video boxes, etc, out of ordinary computer hardware.
i think it would be appropriate here to mention that i couldn't have done many of this projects without the help of arm board - http://emblitz.com.

8/17/2010 1:11 PM EDT

It will be curious to see how well SmoothStone does. To address Mark's question, I'd call $48 million a drop in the bucket for what will be needed if this suceds. But it ought to be enough for proof-of-concept to determine if they're on to something worth further investing.

And heaven knows power costs are an issue. Not only do you have the cost of the power consumed by the servers themselves, but you have the ever increasing demands for A/C to handle the heat generated as server geometries shrink and datacenters pack more into the racks. And the faster you push the CPU, the hotter it runs.

The interesting question this poses is how much power a server CPU *needs*. Yep, Intel architecture CPUs provide higher clock speeds and more MIPS. On a desktop, you may need it. Does a server? What tasks is the server performing that require the raw horsepower?

The last I knew, the majority of machines were still I/O bound, not compute bound, and I'd expect servers to be examples. I can see adding more servers in a clustered environment handling the load nicely and not requiring more horsepower per server.

The things I can think of driving the need for more CPU power are all desktop based, like applications for media creations and editing, or 3D gaming environments. Those aren't things we'll run on a server.

We're seeing the first of a flurry of netbook designs that are ARM based and intended for general purpose computing in a web centric world. If won't be a big surprise if servers follow.

8/17/2010 3:49 PM EDT

More proof that the "cloud" is coming? If Smooth Stone is successful than cloud computing becomes more of a reality - and it seems to be getting closer every day. My big issue is security. Some data just has to be offline and in my pocket so to speak - that is the only way to truly protect it.

8/18/2010 1:05 AM EDT

Nice to hear about start up's. But the trick is not in gluing together multiple cores, that any bummer can do, but taking the best out of them, that's the challenge. And add to that the legacy support required, so the game is all about virtualization and software integration.
Well, power reduction is one good thing if it happens in the end!