- Ivan GodardKeymasterJuly 8, 2015 at 10:29 pmPost count: 627
Rob Abrams asked these questions in email, reposted here (with answers) by permission.
It looks like you’re not abandoning Von Neumann but how does the MILL cpu arch compare with IBM’s TrueNorth or HP’s memristor technology?
We are not trying to create new markets; we address existing markets with a better product. Both TrueNorth and memristor are fascinating, but they are educational sales, which wise startups avoid :-).
If memristor tech (or other approaches to cutting memory latency and volatility; there are several) ever pan out, the Mill could use the same tech in our CPUs too, just as other CPU vendors could. The tech would obviate architecture that is used today to hide the latency of DRAM latency, of which the most important is out-of-order (OOO) execution, used in high end processors intended for servers and supercomputers from many vendors. OOO is extremely area- and power-costly, but does give the highest performance of non-Mill CPUs.
We have never used OOO (that’s a large part of our gain in power/performance at the high end) so if if memristors obviate OOO we will not be directly impacted. However, the change would remove some of the wastefullness of competitive high-end chips, so it might drop our competitive advantage on the high end toward the level of advantage we now have on the low end (for which other vendors do not use OOO).
Low-latency memory would obviate only one aspect In the Mill architecture itself: the “deferred load” (see millcomputing.com/docs/memory for an explanation of how that works). Deferral would become unnecessary. However, deferral is only a minor feature of the Mill, and its use in the code is transparent to the application, so there would be no loss in customer experience nor need to adapt the applications to the change.
TrueNorth is a different story – it is a completely non-von Neuman technology and doesn’t run conventional computer applications at all. Consequently we don’t consider it a competitor, nor would any vendor of conventional CPUs. If it works then it will be able to do things that no regular CPU (including the Mill) can do, but vice versa as well: you can’t use a TrueNorth to run your payroll.
There seems to be quite a bit of movement toward redefining the cpu arch as we enter peak computer power. Can nanophotonics extend the scaling of the conventional cpu arch enough to be a competitive issue for the MILL?
We do not expect to be in the fab business (physically manufacturing chips) for a long time if ever. Consequently we will use the fabrication processes of the merchant fabs such as TSMC. The Mill is quite agnostic when it comes to fab processes and technology, so if photonics becomes commercially viable then the merchant fabs will offer it and we will use it. The first mover in a new fab technology such as photonics always has an advantage, typically a year or two, over the merchant fabs. They also have the first-mover costs 🙂
What early target market can the MILL dominate?
It’s less of an issue of which we dominate, and more which can we afford to initially address; building sales channels is a time consuming and expensive process. The Mill runs conventional applications in conventional ways, but faster, cheaper, and using less power. Those existing applications that have built-in dependencies on legacy architectures will be reluctant to switch to a new chip despite the advantages, but those who are just interested in the most bang for the buck will buy Mills.
Certain markets are more willing to take on new things than others – for example, the high-performance computing market is willing to switch CPUs even for gains much smaller than the Mill gives. When we get closer to having salable products we will select our target entry sub-markets based on the industry as it is then. However, those will just be the entry points; as we grow we fully expect to take on the general-purpose CPU market as a whole, because our price-performance advantage is pervasive throughout.
However, we must grow cautiously: it is easy for a new entrant to a huge multifaceted market to try to enter too many sub-markets at once, thereby incurring all the costs of each sub while not achieving all the revenues of any. We will pick verticals that are small enough to handle (at whatever is our then size), gobble them, and move on to more. In parallel, we will offer product to non-targeted subs via low-cost channels (web, dealers) to pick up added good-margin revenues from early adapters in those subs.
How many quarters do you have to scale sales before your price competitiveness becomes synergistic?
Good question. As we get closer to product we will need to do some hard thinking about pricing strategy. Basically, we can spend our advantage for margin or for market share. The classic approach for a disruptor is to enter at the bottom, go for share, and work up. However, for at least initial sub-markets it might be better for us to enter at the top, go for margin, sell to price-insensitive customers (such as the high-performance markets), and work down. Or both – the Mill is a family, like IBM’s S360 was and is, and we might segment using different family members at different price points. At this point we simply don’t know.
Is there a defined TOC of ownership roi that propels you early on?
Customers will buy Mills for different reasons: some will care that they get more execution than any other CPU will give them regardless of price or power; some will care that they get more execution within a given power/heat budget (large data centers are usually like that); some will care that the performance and power is acceptable for the application and the per-unit cost fits in the retail pricing of the containing device. We can address any of those, and eventually all of them and more, but it’s too early to decide which to focus on.
Can the MILL be applied cost effectively to mobile?
Yes technically, no problem, although probably not commercially at the beginning; mobile is just to big a market for a new entrant to handle. It will be a while before we can afford large marketing, sales, and support staffs in China and Korea.
Sorry, my brain went into orbit. 🙂
You have company; the Mill tends to do that 🙂
You must be logged in to reply to this topic.