Saturday, November 30, 2013
RSC Packs a Petaflops of Xeon Phi in a Rack
Like many top-end systems created today, the PetaStream is a hybrid design, and in this case it mixes a modest number of Xeon processors from Intel with a much larger number of Xeon Phi X86 coprocessors to get more than 1 petaflops of aggregate double-precision number-crunching power into a single rack.
___________________________
99% BAD HARDWARE WEEK: Phi SC5120D Coprocessor with 60 cores. 245W TDP
Price at introduction $ 2700+. Cabinet price $ 5M, Overall power 400+ kW
- The co-processor does not support MMX, SSEx and AVX instructions
- The co-processor includes 8 GB of on-board GDDR5 memory with effective speed 5.5 GT/s
- Peak performance is 1,011 GFLOPS for double-precision operations
- The card is passively liquid cooled
- Card with PCI Express 2.0 x 24 card 230pin
COP part number is an OEM/tray co-processor Clock 1053 MHz
SMC means node controller .NVIDIA uses the terms "node"and streaming multiprocessor interchangeably.
Thus Petastream architecture. Note however that this chip remains essentially coprocessor.
Intel's QLogic Infiniband 312G IB EDR for module connections.
Sounds boring. Latest chip version is able in turbo mode 1,33 Tflops at 300W
Entering a new era in computing
___________________________
99% BAD HARDWARE WEEK: The age of efficient Multithreading will peak in year 2026
Wednesday, November 27, 2013
IBM hardware sale sharply down after Snowden's leaks
Hardware sales in China almost halved, taking IBM's China revenue down 22 percent.
___________________________
99% BAD HARDWARE WEEK:
___________________________
99% BAD HARDWARE WEEK:
Tuesday, November 26, 2013
NVIDIA Project Denver-based SoC due in 2015
We don't know how Kayla will be offered as an actual product, but a working example was shown running ray-tracing in real time.
99% BAD HARDWARE WEEK: Alas, Volta below might be in delayed in 2017 and even 2018. Packaging problems and signal integrity. Together, they now seem like gremlins race. In a 5 TFlops , 1TB/sec of stacked memory bandwidth machine. 256 Volta cards will deliver over 1 PFlops.
Dream of Petaflops machine from year 1998 will be delivered in a single cabinet air cooled, some 20 years later !
Saturday, November 23, 2013
Norse God from box, literally !
128 MB L4 eDRAM at Intel
Friday, November 22, 2013
Optimistic interpretation of Moore's Law Death
Intel is trying to convince us that progress i still here. Lower price per transistor ? But not taking in account drive to bigger wafers and better yield ! Than lower node at 14 nm. But something like that doesn't exist at all. Then, there are no advancements in low power and performance. 22 nm node can be somewhat optimized as it was previously, but not both at the same time and at the declared levels. And the last, but not the least lie: Active power is down. But what about leakage, when device is off ? It grows. Exponentially ! For all dark silicon INSIDE !
___________________________
99% BAD HARDWARE WEEK: UV projection sources at 133 nm interfere with SiO2 on wafer bandgap. Thus, industry forced to extend use of 193nm deep UV photolithography
___________________________
99% BAD HARDWARE WEEK: UV projection sources at 133 nm interfere with SiO2 on wafer bandgap. Thus, industry forced to extend use of 193nm deep UV photolithography
Thursday, November 21, 2013
Hackers' arrow into Cupid's heart
DATING WEBSITE OUTFIT Cupid Media has been caught up in a security scandal that suggests that 42 million user credentials have been compromised.
The discovery was made by security researcher Brian Krebs of Krebs on Security. He found the data on the same server that contained data gained from hacking attacks on other well known organisations, including Adobe and PR Newswire.
___________________________
99% BAD HARDWARE WEEK:
Saturday, November 16, 2013
Boeing And Lockheed Martin Team On New Bomber
The two companies originally agreed to work on what was then the Next-Generation Bomber (NGB) program, aimed at delivering a new bomber in 2018. (The deal was announced in early 2008.)
___________________________
99% BAD HARDWARE WEEK: Northorp Grumman's probable proposal
AMD HSA inside Sony's PS4 !
99% BAD HARDWARE WEEK: 8 CPU cores. Something like AMD's Berlin core ? No, obviously pair of Jaguars blended with 18 GCU stream processor. Note all three processing blocks have independent but HSA memory interface.
Thursday, November 14, 2013
IBM makes Watson available via API
IBM has upped the ante in the API game by making its Watson question-answering system available as a service. That’s right, Watson could soon power your smartphone app.
___________________________
99% BAD HARDWARE WEEK:
AMD security processor with ARM trust zone
99% BAD HARDWARE WEEK:
UK spies continue “quantum insert” attack via LinkedIn, Slashdot pages
According to a new report by Der Spiegel, the British signals intelligence spy agency has again employed a “quantum insert” technique as a way to target employees (Google Translate) of two companies that are GRX (Global Roaming Exchange) providers.
___________________________
99% BAD HARDWARE WEEK:
Wednesday, November 13, 2013
How long do disk drives last?
Why does a company that keeps more than 25,000 disk drives spinning all the time not know how long they last? Backblaze has been providing reliable and unlimited online backup for over five years. For the past four years, we’ve had enough drives to provide good statistics, but 74% of the drives we buy are living longer than four years. So while 26% of drives fail in their first four years, and we have detailed information about the failure rates of drives in their first four years.
___________________________
99% BAD HARDWARE WEEK: We are up to 75 petabytes of cloud storage now.
Tuesday, November 12, 2013
AMD's Kaveri at 3,7Ghz 856 GFlops
Kaveri will run at up to 856 gigaflops, between two Steamroller CPU cores at 3.7GHz and eight Radeon GCN compute units, with each of the 512 GPU cores spread amongst the eight Radeon GCN CUs running at 720 MHz.
___________________________
99% BAD HARDWARE WEEK: Server processor Berlin 1P at 0,7 TFlops? Wow.
Friday, November 08, 2013
Microsoft might sell Xbox and drop Bing
Microsoft might sell Xbox and drop Bing if Elop gets CEO job
___________________________
99% BAD HARDWARE WEEK:
___________________________
99% BAD HARDWARE WEEK:
Bit Commitment Based on Quantum Communication and Special Relativity
Bit commitment is a fundamental cryptographic primitive in which Bob
wishes to commit a secret bit to Alice. Perfectly secure bit commitment
between two mistrustful parties is impossible through asynchronous
exchange of quantum information. Perfect security is however possible
when Alice and Bob split into several agents exchanging classical and
quantum information at times and locations suitably chosen to satisfy
specific relativistic constraints. Here we report on an implementation of a bit commitment protocol using quantum communication and special relativity. Our protocol is based on [A. Kent, Phys. Rev. Lett. 109
130501 (2012)] and has the advantage that it is practically feasible
with arbitrary large separations between the agents in order to maximize
the commitment time. By positioning agents in Geneva and Singapore, we
obtain a commitment time of 15 ms.
___________________________
99% BAD HARDWARE WEEK:
___________________________
99% BAD HARDWARE WEEK:
Silicon gets silicone but not silicon ones !
But III-V silicone .
According to Thean, Imecs partners have asked it to further develop the technology for possible deployment at the seven-nanometer node circa 2016-to-2018.
___________________________
99% BAD HARDWARE WEEK: Intel's 22 nm node was in 2011, 14 nm in 2014, 10 nm in 2017, thus probably nothing new before 2020. Intel have to have the solution ready at least 2 year before the first 7nm products. Thus just in time for 2020 frame if delivered in 2018 ! Sorry Intel for indiscretion, but this is BAD Hardware, isn't it ?
Nixon was right !
"The data mining and surveillance of government officials, allowing a constant stream of information on political backroom dealing, could create the most open democracy the world has ever seen."
Nixon had various White House rooms outfitted with telephone line taps and lavalier microphones, all recorded with nine voice-activated Sony TC-800B open-reel tape recorders. Our surveillance technology is far more impressive. Surely we can do better than Nixon's spooks. Neither did he have smartphones, which we know are incredible surveillances sources, and very likely in the hand of every legislator and bureaucrat in this fair land.___________________________
99% BAD HARDWARE WEEK: The problem is that all are now spied. And government's officials too ? Except them, except :) Thus, welcome to Orwellian 1984 :(
Thursday, November 07, 2013
Xbox is worth nothing, only Android taxes protect it from axe !
Canada-based telecom Nortel went bankrupt in 2009 and sold its
biggest asset—a portfolio of more than 6,000 patents covering 4G
wireless innovations and a range of technologies—at an auction in 2011.
Google bid for the patents, but it didn't get them. Instead, the patents went to a group of competitors—Microsoft, Apple, RIM, Ericsson, and Sony—operating under the name "Rockstar Bidco." The companies together bid the shocking sum of $4.5 billion.
Patent insiders knew that the Nortel portfolio was the patent equivalent of a nuclear stockpile: dangerous in the wrong hands, and a bit scary even if held by a "responsible" party.
In 2011 Google placed an initial $900 million bid for Nortel's patents. Google increased its bid several times, ultimately offering as much as $4.4 billion.After losing out to RBX on the Nortel patents, Google went on to acquire Motorola Mobility for $12.5 billion, a deal driven partly by Motorola's library of patents.
___________________________
99% BAD HARDWARE WEEK: Strange coalitions.
Patents will be able to generate $4.5 billion or more in licensing fees is a dangerous assumption ? Eh, only Android royalties bring to Microsoft $2B each year :) But from arch enemy Google. Well, who would expect that coalition to save XBox from Sony's , RBX member too, Playstations.
Google bid for the patents, but it didn't get them. Instead, the patents went to a group of competitors—Microsoft, Apple, RIM, Ericsson, and Sony—operating under the name "Rockstar Bidco." The companies together bid the shocking sum of $4.5 billion.
Patent insiders knew that the Nortel portfolio was the patent equivalent of a nuclear stockpile: dangerous in the wrong hands, and a bit scary even if held by a "responsible" party.
In 2011 Google placed an initial $900 million bid for Nortel's patents. Google increased its bid several times, ultimately offering as much as $4.4 billion.After losing out to RBX on the Nortel patents, Google went on to acquire Motorola Mobility for $12.5 billion, a deal driven partly by Motorola's library of patents.
___________________________
99% BAD HARDWARE WEEK: Strange coalitions.
Patents will be able to generate $4.5 billion or more in licensing fees is a dangerous assumption ? Eh, only Android royalties bring to Microsoft $2B each year :) But from arch enemy Google. Well, who would expect that coalition to save XBox from Sony's , RBX member too, Playstations.
Monday, November 04, 2013
6TB hard disk
99% BAD HARDWARE WEEK:
Broadwell against Kaveri
A lot of optional geometry shaders in Broadwell.
- Right now there are GT1, GT2, and GT3 variants shown for the initial Broadwell processors. The Broadwell GT1 PCI IDs include 0x1602, 0x1606, 0x160A, 0x160B, 0x160D, and 0x160E. Broadwell GT2 PCI IDs are 0x1612, 0x1616, 0x161A, 0x161B, 0x161D, and 0x161E. Lastly, the high-end GT3 graphics for Broadwell are IDed as 0x1622, 0x1626, 0x162A, 0x162B, 0x162D, and 0x162E.
- Over Ivy Bridge / Haswell "Gen7" graphics, Broadwell "Gen8" claims support for a maximum of 280 VS threads over 36 on Gen7 hardware, a maximum of 256 GS threads over 36 on Gen7, and 64 WM threads over 48. The URB size has also quadrupled to 512 and the max VS entries supported are now 1664 over 512 and 640 GS entries over 192 on Gen7 hardware.
Coming back to the details, it has been claimed that Intel Broadwell would pump up the GPU performance of by 40% over Haswell which itself has a powerful GPU side. Intel Broadwell GPU would also bring in new and improved instruction set.
There are three variants of Haswell graphics, GT1, GT2, and GT3. The top end part, GT3, has 40 of the enhanced Ivy shaders. If the chip does not increase GPU clocks at all, has zero performance enhancements, nor anything else, it will be 2.5x as fast as Ivy Bridge. That in turn is 2x as fast as Sandy Bridge, so 5x what you can buy from Intel today. AMD’s Kaveri had better be damn quick to stay in the game. Berlin too ?
___________________________
99% BAD HARDWARE WEEK: But, after latest delay Broadwell is at least 6 months behind Kaveri.
- Right now there are GT1, GT2, and GT3 variants shown for the initial Broadwell processors. The Broadwell GT1 PCI IDs include 0x1602, 0x1606, 0x160A, 0x160B, 0x160D, and 0x160E. Broadwell GT2 PCI IDs are 0x1612, 0x1616, 0x161A, 0x161B, 0x161D, and 0x161E. Lastly, the high-end GT3 graphics for Broadwell are IDed as 0x1622, 0x1626, 0x162A, 0x162B, 0x162D, and 0x162E.
- Over Ivy Bridge / Haswell "Gen7" graphics, Broadwell "Gen8" claims support for a maximum of 280 VS threads over 36 on Gen7 hardware, a maximum of 256 GS threads over 36 on Gen7, and 64 WM threads over 48. The URB size has also quadrupled to 512 and the max VS entries supported are now 1664 over 512 and 640 GS entries over 192 on Gen7 hardware.
Coming back to the details, it has been claimed that Intel Broadwell would pump up the GPU performance of by 40% over Haswell which itself has a powerful GPU side. Intel Broadwell GPU would also bring in new and improved instruction set.
There are three variants of Haswell graphics, GT1, GT2, and GT3. The top end part, GT3, has 40 of the enhanced Ivy shaders. If the chip does not increase GPU clocks at all, has zero performance enhancements, nor anything else, it will be 2.5x as fast as Ivy Bridge. That in turn is 2x as fast as Sandy Bridge, so 5x what you can buy from Intel today. AMD’s Kaveri had better be damn quick to stay in the game. Berlin too ?
99% BAD HARDWARE WEEK: But, after latest delay Broadwell is at least 6 months behind Kaveri.
Friday, November 01, 2013
FINFET: Nobody knows anymore what 16 nm means or what 14 nm means !
It’s actually become a fairly common refrain among industry experts. The practice of attaching measurements to chip generations has “been hijacked by marketers to an enormous extent,” one chip-design expert told me. “A lot of it’s really smoke and mirrors,” says analyst Dan Hutcheson of VLSI Research in Santa Clara, Calif. It’s “spin,” he says, that’s designed to hide widening technological gaps between chip companies.
The switch to FinFETs has made the situation even more complex. Bohr points out, for example, that Intel’s 22-nm chips, the current state of the art, have FinFET transistors with gates that are 35 nm long but fins that are just 8 nm wide.
___________________________
99% BAD HARDWARE WEEK: Well, seems WE knew. Moore's law has been dead long since 2007, but we didn't know :)
The term “350 nm node” actually meant something. But around that same time, the link between performance and node name began to break down.
