3 ms·

It is—and I think that, unfortunately, POWER10’s primary niche is to provide the most plausible alternative to x86. Selling to plebs is (1) expensive since the

It is—and I think that, unfortunately, POWER10’s primary niche is to provide the most plausible alternative to x86. Selling to plebs is (1) expensive since the economy of scale isn’t there and (2) not useful to IBM’s business strategy anyway. Like, if you are buying 10,000 x86 processors from Intel, then you tell Intel, “Don’t charge us too much, or we’ll buy POWER systems from IBM instead.” Intel chuckles, since they know you’re not serious, but (1) you buy a few racks of POWER systems and make your engineers build and test everything on POWER to make your negotiation position more credible, and (2) Intel knows they can’t charge you more than IBM does. The other niche for POWER, as far as I can tell, is that it has various certifications, or it’s used in systems with various certifications that x86 doesn’t have (e.g. certain aerospace / defense applications). For example, the RAD5545 is a rad-hard POWER SoC.

wmf 6y ago
By the time POWER10 comes out ARM (e.g. Altra Max) will be a more plausible alternative.
- klodolph 6y ago
  Will it be? What makes you say that? I’m a little skeptical because it looks like they are trying to paper over the weaknesses in ARM CPUs by e.g. throwing more cores on a chip, derating competitor’s benchmarks, etc. Not to say it won’t eventually be successful, but the Xeons don’t have a smaller number of bigger cores by accident.
  - imtringued 6y ago
    My misadventures with multi core programming tell me that there is zero demand for multicore application development. Although I have experienced that the vast, vast majority of problems can be divided to run on multiple cores, assuming you have enough work for the cores. There was a problem where I convinced myself it could not be done without locking and a complicated transaction scheme but I still somehow managed to turn the problem into a data parallel algorithm. The code is extremely counterintuitive, error prone and both hard to write and read. However, writing a purely parallel application is so much work even for an expert that there is a huge productivity drop in many cases and for non experts a lot of the problems are intractable, so it is only done when absolutely necessary. So my general strategy has been to significantly relax maximum possible parallelism in exchange for the ability to pretend that your code is single threaded. The general idea is quite simple. Split your data into smaller segments. If a segment accesses another segment merge the two into a big segment. Then process each segment as if it was single threaded but you can process independent segments in parallel. The big problem with this approach are pathological cases where you get very large segments. One (fictional) example would be a simulation involving long distance transport. Train stations that are connected with rails form a very huge graph of segments consisting of dozens of train stations. However, airports are not connected so each airport can still be simulated in parallel and if you have multiple rail networks you can simulate each on a separate core. It's far faster than staying purely single threaded but it means that you will always be bottlenecked by your slowest core and that means current ARM offerings will be behind unless you are doing something extremely multi core intensive like video encoding which will gobble up as many cores as you have.
  - wmf 6y ago
    The AnandTech and STH benchmarks look much better than I expected, even being close in single-thread. Remember I'm not talking about actual success but just something you can show Intel/AMD salespeople to negotiate lower prices.
    - klodolph 6y ago
      I think specs aren’t enough, you have to have a credible supplier that will land real servers in your DCs in the volumes you need, and I’m not convinced that this has materialized for ARM yet. IBM is exceptionally credible when it comes to actually putting hardware in your DCs. This is why I wasn’t very excited by Altra Max benchmarks. Eventually we will have ARM suppliers that can churn out motherboards and associated components with the kind of magnitude that can put a dent in x86 usage, but in the meantime, the only company that seems to be doing it is Amazon, who uses their own custom ARM server designs. Given enough time, suppliers for ARM servers will rise up, if the processors are competitive, but that won’t happen overnight. Certainly not just because of one new CPU.
      - wmf 6y ago
        I guess we have no way of knowing if Wiwynn is making hundreds or hundreds of thousands.
- imtringued 6y ago
  There is no shortage of decent ARM server designs, there is a massive shortage of ARM server designs that actually were available on the market. The vast majority of them never delivered. It's like having a dozen kickstarter scams about the same idea in a row.
- mhh__ 6y ago
  Who in this situation is going to buy from them when you could just buy IBM and be done with it? The whole point is that POWER actually exists (on 7nm etc.)

wmf 6y ago

By the time POWER10 comes out ARM (e.g. Altra Max) will be a more plausible alternative.

klodolph 6y ago
Will it be? What makes you say that? I’m a little skeptical because it looks like they are trying to paper over the weaknesses in ARM CPUs by e.g. throwing more cores on a chip, derating competitor’s benchmarks, etc. Not to say it won’t eventually be successful, but the Xeons don’t have a smaller number of bigger cores by accident.
- imtringued 6y ago
  My misadventures with multi core programming tell me that there is zero demand for multicore application development. Although I have experienced that the vast, vast majority of problems can be divided to run on multiple cores, assuming you have enough work for the cores. There was a problem where I convinced myself it could not be done without locking and a complicated transaction scheme but I still somehow managed to turn the problem into a data parallel algorithm. The code is extremely counterintuitive, error prone and both hard to write and read. However, writing a purely parallel application is so much work even for an expert that there is a huge productivity drop in many cases and for non experts a lot of the problems are intractable, so it is only done when absolutely necessary. So my general strategy has been to significantly relax maximum possible parallelism in exchange for the ability to pretend that your code is single threaded. The general idea is quite simple. Split your data into smaller segments. If a segment accesses another segment merge the two into a big segment. Then process each segment as if it was single threaded but you can process independent segments in parallel. The big problem with this approach are pathological cases where you get very large segments. One (fictional) example would be a simulation involving long distance transport. Train stations that are connected with rails form a very huge graph of segments consisting of dozens of train stations. However, airports are not connected so each airport can still be simulated in parallel and if you have multiple rail networks you can simulate each on a separate core. It's far faster than staying purely single threaded but it means that you will always be bottlenecked by your slowest core and that means current ARM offerings will be behind unless you are doing something extremely multi core intensive like video encoding which will gobble up as many cores as you have.
- wmf 6y ago
  The AnandTech and STH benchmarks look much better than I expected, even being close in single-thread. Remember I'm not talking about actual success but just something you can show Intel/AMD salespeople to negotiate lower prices.
  - klodolph 6y ago
    I think specs aren’t enough, you have to have a credible supplier that will land real servers in your DCs in the volumes you need, and I’m not convinced that this has materialized for ARM yet. IBM is exceptionally credible when it comes to actually putting hardware in your DCs. This is why I wasn’t very excited by Altra Max benchmarks. Eventually we will have ARM suppliers that can churn out motherboards and associated components with the kind of magnitude that can put a dent in x86 usage, but in the meantime, the only company that seems to be doing it is Amazon, who uses their own custom ARM server designs. Given enough time, suppliers for ARM servers will rise up, if the processors are competitive, but that won’t happen overnight. Certainly not just because of one new CPU.
    - wmf 6y ago
      I guess we have no way of knowing if Wiwynn is making hundreds or hundreds of thousands.
imtringued 6y ago
There is no shortage of decent ARM server designs, there is a massive shortage of ARM server designs that actually were available on the market. The vast majority of them never delivered. It's like having a dozen kickstarter scams about the same idea in a row.
mhh__ 6y ago
Who in this situation is going to buy from them when you could just buy IBM and be done with it? The whole point is that POWER actually exists (on 7nm etc.)