Part 1 .... Part 21
(Dr Dobbs series... DDJ is one site thats worth registering .. they dont block safetymail.info...)
DDJ
NVidia Graphics cards , massively parallel, speed up calculations 10x-100x
Since 2007 NVIDIAs CUDA has brought super-computing to anyone with a game card...
"GeForce GT 220 packs 48 processing cores into a compact power efficient.."(< $80) 2009
this year: NVIDIA® GeForce® GTX 550 Ti graphics processing unit (GPU) $150 "GPU called GF116,"
... Ti is a Fermi?? newegg
:The refreshed Fermi chip is large: it includes 512 stream processors, grouped in 16 stream multiprocessors clusters (each with 32 CUDA cores)," ??
The GF116 has a single Graphics Processing Cluster (GPC), with four Streaming Multiprocessors (SMs). Each SM contains 48 shader cores, four dispatch units, and eight texture units. All told, GF116 employs 192 shader cores, four Polymorph engines (one per SM), and 32 texture units.??
... and all three ROP partitions are fully functional.. similar to the uncut GF106 GPU in Nvidia's GeForce GTX 460M mobile graphics module. With each of the three ROP partitions capable of eight 32-bit integer pixels per clock, we have 24 ROPs and a cumulative 192-bit memory interface..
(24 * 32bits per clock doesnt sound all that grand .. maybe a "shader unit" can do a multiply?)
..............................................
ATI (AMD) Radeon HD 5770 features the exact core configuration of the Radeon HD 4870 and 4890: 800 SPUs, 40 TAUs (Texture Address Units) and 16 ROPs (Rasterization Operator Units).
might be better than NVIDIA, but as NVIDIA write CUDA , they might be more copasetic
"The Radeon HD 5770 features a remarkably low 18 watt idle consumption level, making it one of the most efficient graphics cards available today. When pushing the card to the extreme it will still suck up to 108 watts, but even with the increased thermal stress, noise levels were comparable to those of the Radeon HD 4770 or GeForce GTS 250 graphics cards."
...........................................
Jargon ROP Raster Operations Partition? ROP partition = ROPs + memory controllers + L2 cache .
...........................................
GeForce GTX 295 Each GPU features the full array of 240 processing cores and 80 texture filtering units. The processor cores and filtering units operate at 1242 MHz and 576 MHz respectively ($500? fits in one slot?)
There are GPU cards up to $700.
Either pay $$ for Mathematica 8 ..'Native support for Compute Unified Device Architecture (CUDA) and OpenCL GPU'
or write yr own.
Me, I'd like to set a couple of cards making Carmicheal numbers and seeing how many Miller-Rabin bases I can fox.
At 2048 bits, any more than 2 bases would be fun!
I'm assuming that GeForce work well on Integers. Surely better than AMD 8-core Bulldozers ($205-4245)
cnet
Since I dont do games, or huge monitors, most of the fun would be lost if I bought a DEsktop and a couple of GPU cards
...........................................................................................
Java
jcuda
JCublas is a library that makes it it possible to use CUBLAS, the NVIDIA CUDA implementation of the Basic Linear Algebra Subprograms, in Java applications.
JCublas provides methods for
Vector operations (Level 1 BLAS)
Matrix-Vector operations (Level 2 BLAS)
Matrix-Matrix operations (Level 3 BLAS)
- so no Integers??
I suppose that means dusting off JNI
nvidia
................................
Some Integer work in C: Jeet Chauhan Summer 2010 - a quick eyeball seems to show that Chauhan is NOT using BigInt, ie integers < 2^32 - no fun!!! calstate
__________________________________________________________________
Jolt
More DDJ stuff...
winners of the 2011 Jolt Awards for books, the winner of the top prize,
Continuous Delivery by Jez Humble and David Farley, was a run-away #1 choice
Continuous Delivery: Reliable Software Releases through Build, Test, and Deployment Automation (Addison-Wesley Signature Series (Fowler)) [Hardcover]Jez Humble (Author), David Farley (Author)
DDJ
Reliable Releases - this seems to be a lasting fashion, cant be bad..
(Jolt was (is?) a cola with "all of the sugar and twice the caffeine?)
realworldtech
ReplyDeleteAlmost all 256-bit AVX instructions are decoded into and execute as a single uop – in contrast to AMD’s more cautious embrace of AVX, which will crack 256-bit instructions into two 128-bit operations on Bulldozer.
For example, a 256-bit multiply can issue to port 0 and simultaneously use the 128-bit SIMD data path for the low half and the 128-bit FP data path for the high half
if SandyBridge really does 128bit * 128 bit integer in one op thats 16 times faster maths than 32 bit ops
so the NVIDIA® GeForce® GTX 550 Ti is only 1.5 times faster that a SandyBridge?
That doesnt sound right...
actually per second GTX550:SandyBridge 1 : 3.6
ReplyDeleteper $1K dollar 6 : 3.6
But an Extra Sandybridge Desktop = $2K+
and an extra 550 only $150
so the benefit is > 3:1
but not "10-100x"
and the room is cooler and less crowded..