Slashdot Mirror
Next-Gen Intel Chip Brings Big Gains For Floating-Point Apps
An anonymous reader writes "Tom's Hardware has published a lengthy article and a set of benchmarks on the new "Haswell" CPUs from Intel. It's just a performance preview, but it isn't just more of the same. While it's got the expected 10-15% faster for the same clock speed for integer applications, floating point applications are almost twice as a fast which might be important for digital imaging applications and scientific computing."
The serious performance increase has a few caveats: you have to use either AVX2 or FMA3, and then only in code that takes advantage of vectorization. Floating point operations using AVX or plain old SSE3 see more modest increases in performance (in line with integer performance increases).
" Next-Gen Intel Chip Brings Big Gains For Floating-Point Apps "
How much of a gain? More or less than 0.00013572067699?
I hope there's really a new Mac Pro coming and that it has these chips in it! I do a heck of a lot of PDE solving, statistics and simulations, and would love to have a screamin' machine again.
For problems where you need floating point AND is not multithread friendly AND need large computing power AND is specially coded, then this will be of great use. However, most massive computing problems like this are multi-thread friendly and this will still be roughly an order of magnitude from the speeds you can get by using a GPU.
Slightly, but you haven't been keeping up on the latest hardware? My pair of Sapphire 5830's graphics cards would top off at about 435MH/s at a total system wattage of around 520W. The new Jalapeno chips from butterfly labs will do 4500 MH/s using 2 watts total system power. For comparison, my i5-2400 performed 14MH/s at 95W or so. So the Jalapeno is about 321x faster and about 47x more power efficient so combined, I believe that's 15,267.864x more efficient.
While it's got the expected 10-15% faster for the same clock speed for integer applications, floating point applications are almost twice as a fast HTH
Integer and floating point are separately implemented in the hardware, so an improvement to one often doesn't apply to the other. You can add integers by counting on your fingers. To do that with floating point, you have to cut your fingers into fractions of fingers - a very different process.
See: http://en.wikipedia.org/wiki/FMA3
It's common to have an accumulator like this:
X = X + (Y * Z)
To compute that in floating points, the processor normally does:
A= ROUND(Y*Z) X=ROUND(X+A)
Each ROUND() is necessary because the processor only has 64 bits in which to store the endless digits after the decimal point. FMA can fuse the multiply and the add, getting rid of one rounding step, and the intermediate variable:
X= ROUND( X + (Y*Z) )
That makes it faster. Since integers don't get rounded to the available precision, the optimization doesn't apply to integers. The above processor would do Y*Z, then +X, then round, then X=. A CPU designer can make that faster by including either a "add and multiply" circuit or a "add and round" circuit or a "round and assign' circuit. Any set of operations can be done in two clock cycles, if the maker decides to include a hardware circuit for it.
Would that improve hashing speeds in, say, Bitcoin?
Bitcoin is based on SHA256 hashing, which has zero floating point operations. So no, this will not impact Bitcoin mining at all.
While speed for single and double floats is all well and good, I wonder - when will there finally be hardware support for 128 bit (quadruple precission) floats?
It was there on PowerPC for many years, and with Haswell it will be there for x86 as well. FMA is all you need for efficient 128 bit arithmetic.
Pah. AMD had FMA4 since 2011
It would prevent the need to some extra math for extra high numbers (not just those that end on a high numbers, but where the intermediate calculation may be high (e.g.: factorial math to find out the probability of something if I recall). Plus, 96 bits is more than enough for the fraction if you ask me - very greedy in fact to take that to 112 at the cost of 16 bits the exponent could well do with.
Why OpalCalc is the best Windows calc
AMD has lost the CPU race a long time ago, but still beats Intel with integrated graphics. Now, It looks like Haswell could win that battle too.
The article shows GT2 to be 15% - 50% faster than the old HD4000. That's still a bit slower than Trinity, but GT3 has double the execution units than GT2, potentially blowing anything away that AMD could offer.
AMD already has FMA3. They also published great results. Of course nobody read it, at least I have seen mentioned it in the usual generic benchmark articles people like to refer (which does not use FMA3).
In the early 2000's we had some, every week one of them would crash. All the other servers w/ECC, no crash. Hardly a marketing gimmick.
You can sell them on the exchange quickly and easily for USD (or 5 other major currencies)
here's an old paper describing octuple precision on the PowerPC G4
Many problems in number theory and the computational and physical sciences, espe- cially in recent times, require more floating point precision than is commonly available in fundamental computer hardware. For example, the new science of “experimental mathematics,” whereby algebraic truths are foreshadowed, even discovered numerically, requires much more than single (32-bit) or double (64-bit) precision.
That paper references Bailey's 2000 paper on Quad double algorithms, which alludes to "pure mathematics, study of mathematical constants, cryptography, and computational geometry