And one where an Atom Z3735F gets the same score as a Core i5 3317U...
Look closer: it takes the Atom 4 cores to do what the Core i5 3317U does using 1 thread.
And one where an Atom Z3735F gets the same score as a Core i5 3317U...
I'm not going to detabe with you anymore on this. The numbers speak for themselves.
Max load on an ipad air 2 is 11W vs 29W on the macbook.
Are you seriously saying that a faster SSD, more RAM and marginally bigger screen uses 18W? That is ridiculous.
You only have to look at the idle and max to see isolate the CPU/GPU components as the max is measured when loading the CPU/GPU.
The delta on the ipad air 2 is 6W and macbook is 23W.
Seriously people need to use their brain a bit. Notebook check is measuring power from the wall not the SoC, they are essentially just measuring the size of the power adapter the device comes with. There is no real easy way of measuring the SoC power. If the iPad 2 Air came with a 5W adapter you'll get a max reading of 5W (of course your battery would probably not charge when device is on). Drawing more power from the wall doesn't mean the CPU is actually drawing all that power as the extra power draw is used to charge the battery at a quicker pace.
Like you said the Max load on an iPad 2 Air (as measured by the wall) is 11W and the macbook 29W... I'll give you two guesses on what the wattage on the power adapter that comes with these device are. 10W and 29W... Coincidence? I think not. Once you factor in adapter efficiency you get the 11W and 29W numbers.
The more I learn about software optimization, the more sceptical I am about cpu benchmarks....
I don't think it is a problem if a benchmark is not vectorized, if it is application logic that the average programer would write. But the algorithms geekbench uses don't belong into this category. The algorithms geekbench uses are normally heavenly vectorized and optimized.
...
The more I learn about software optimization, the more sceptical I am about cpu benchmarks.
Usually, you are testing a particular software implementation, a compiler and some piece of hardware jointly. Trying to compare two pieces of hardware this way is hard.
My experience is that software implementations (and compilers) tend to be quite suboptimal on a pure performance basis, but the degree of "optimalness" is highly non-linear and hard to predict. Thus, small changes in code (or compiler) can result in large changes in performance.
It is often hard to know if a piece of software is close to "optimal". So (often) you don't even know robustly that if you are close to hw limits. In simple cases, you might be able to determine that e.g. memory limits your application and that this memory traffic is unavoidable. But often the problem is so complex (and the hardware is so hard to fathom) that such estimates are crude.
Testing the "sw ecosystem" might be equally relevant. I.e. how many man-hours does it take to implement operation X at a running speed of N seconds? How many dollars worth of tools/licenses? How many prospective customers will the platform offer the developer to distribute costs on? If the answer is that "this platform allows the dev to run some given matrix multiplications at one million/second by using a free library", then that may (or may not) be more relevant than "this platform allows the dev to run those same matrix mults at four million/second by investing 6 months of development time writing assembly and understanding the quirks of cache implementations".
One might expect that any platform/task combination will have some unique performance vs "cost" curve. Measuring absolute hardware limits only tells you the (expected) asymptote of that curve for those willing to put endless effort into the project, while details of that curve tells you more about what you can get for a more moderate effort.
That said, if your box is going to be devoted to one or a few tasks (calculate FFTs or run Quake or whatever), then measuring performance for the application of interest is going to tell you how fast that application is going to run on two or more hw platforms. Until the next recompile at least.
-k
I predict that
A9 duel core with 1.8Ghz
A9X Tri core with 2Ghz, GT6850 GPU
From chinese information.
We will have to wait for benchmarks but it does look like Apple quoted performance increases are for singled threaded tasks.
I.e. In apple's tech specs for iPad it notes that:
- a8 in the mini 4 is 1.3x faster than a7
- a8x in the air 2 is 1.4x faster than a7
- a9x in the pro is 2.5x faster than a7
A8x and a8 difference (even when both are running at 1.5ghz) in single thread can be explained by the larger l2 cache 2mb vs 1mb and faster memory interface.
If they are quoting multithreaded perf they would have put a8x much higher than a8 because of core counts.
If the information quoted here based on Chinese ministry of industry A9 is 1.8 GHz with 2 cores, we are talking of about ~30% better IPC, which would definitely be impressive. I'll wait for benchmarks...Yeah, if Apple's performance increase claims are at iso core counts, then they're damn impressive.
A real-life example for myself is TMPGEnc. Some say it is well optimized for Intel hardware and thus will always turn out faster encode times on an Intel CPU than a comparably priced AMD CPU. For me this distinction is irrelevant, all I am interested in is price/performance for the software as it comes to me from the distributor as I cannot acquire a hypothetically better optimized version of it for an AMD processor, so allegations of compiler optimization bias are irrelevant in this situation.
A positive side effect of nearly owning the market. Same is to a lesser extent true for Nvidia.- Excatly this. At the end of the day, the only thing that matters.
Yeah, if Apple's performance increase claims are at iso core counts, then they're damn impressive.
A9X is 80% faster than A9 according to Apple. Assuming thats a pure multithreaded statement we still get only 33% improvement from a 4th core. The rest of 30% or more improvement will have to come from increased clocks and higher IPC. There is no way IPC can be improved 30% or more on an already impressive high performance CPU core. A9X is definitely quad core and A9 is tri-core. You only need to see Cyclone to Enhanced Cyclone or Broadwell to Skylake to see that a 10% IPC improvement itself is at the upper end of the improvements we can expect on a existing high performance CPU core. The only other way to drastically improve IPC is through new CPU instructions and extending the ISA.
Too bad we don't have the Lua and Dijkstra scores.
IPC about 10% better on integer and 20% on FP. Even though that's less than the 30% that would have come from the Apple claim of 70% faster, that's still very good.