About a year ago, when I was reading about LightPeak(Thunderbolt), Intel claimed it was going to be 10-40gbit and was going to take a few years *after* 22nm became standard.
Just based on Intel claiming LightPeak was meant to come out after 22nm, means it was released early. I am not surprised that the optical version is still some odd years off.
Personally, I think this early release was a mix of Apple and Intel. Apple wanting the fastest and unique, and Intel wanting to make at least some money on their tech as it has been in development for quite a long time. I remember reading about "some optical tech that will scale to 100gbit and become standard", from Intel back near 2005.
You should check out the execution units on the i7, there is a lot of duplicated FP/int units, which allows HT to get some pretty hefty performance benefits. SSE units are shared, so any SSE benchmarks would take a hit with HT.
If you can get more than 1 DP/cycle on your AMD chip with 6 "real" cores, let me know; because Intel is getting 1 DP/cycle on its "fake" cores.
Your highest performance with HT would be one thread doing DP and another doing SSE.
I get a 19ms ping to Chicago, which is somewhere upwards of 500mi via the trace route in another state. Put in some more localized rendering farms, like per State, and you could easily keep latency low enough for the average user.
Jitter could be an issue if the network isn't well designed. It would probably show up as micro-stuttering.
Based on what I've read about Raytracing vs Rasterization, Raytracing *will* win out in the long run. I guess RT scales better than rast, but the overhead is expensive. Once we get to the point where RT is about the same speed as rast, it will only take 1-2 generations before RT is several factors faster.
Whichever company is ready to push out RayTracing, will stomp the market. If you release too early, your product will just be a gimmick, if you release too late, the competition will be several times faster than you.
How does this work if someone buys one of these devices for their kid? I thought ALL contracts with children are void and must be signed via the legal guardian.
If the parent never touches/sees/uses the devices in question, how could they be bound to the EULA?
The service packs fix bugs and restructure a little bit. The OS is considered old because its framework/design is 10 years old.
I can rearrange the furniture and paint the walls, but I can't claim my 2001 house is only 3 years old. ohh.. wait.. car analogy.. I can change the oil and tweak the ECU, but I can't claim my 2001 car is only 3 years old.
Until the competition keeps releasing higher and higher bandwidth switches/router and someone comes up with a new killer apps that makes use of it. Then you're left in the dust, and you lose most of your customers. Network admins don't like to have lots of different switch brands, then there's more to keep track of. Cisco falls behind and maybe HP and/or Juniper fill the void, and in the several years it takes to release a competing product, most of their customers have already jumped ship and gutted their infrastructure of anything Cisco.
Humans hate change. There needs to be enough incentive to change, and even more incentive to change back. It will leave people with a sour taste.
About 1/2 of my in-state tuition was paid for by royalties from patents. My in-state tuition was only supported by about 30% tax payers. The rest was by the student or money made from research. If you were out-of-state, you had to pay everything yourself, unless from a sister-Uni.
My state Uni was almost entirely supported by tuition and Alumni donations. Federal funding only went directly to departments or groups that qualified, not to the Uni.
"my guess is they're using a custom-developed tool to export the data and it's buggy"
Good point. We'd have to see what's common among the customers, but they always eventually fix it so we never dug further than sending back examples problem and telling them to fix it.
We don't do it ourselves, we tell our customers what we need and this is what happens.. Often. We have about 30 different customers who use Oracle, and nearly all of them have had this issues more than once at some point. A select few have never had this issue.
You say "all day", but everything I read says "milli/microseconds". There is many many magnitudes of difference between what you said and what's happening.
What seems to be happening is there is A LOT of competition for "middle men". As more and more people try to squeeze into this "middle man" market. Because of this, they need shorter and shorter times in order to buy/sell before the competition. This brings us to milli/micro second trade times.
At some point, middlemen become a large portion of the market. What happens when a market is "owned" by middlemen and few long term investors relative to middlemen?
We need middlemen, but we don't need a lot of middlemen clawing as fast as they can to artificially inflate a stock just to see who can dump everything before it peaks to become the "winner".
Have you about the issues of HFT causing mini "bubbles" and pop and crash stocks? Sounds great.
If you *know* people and have the money, you can get your computers at the exchange and they measure those latencies in microseconds. They said they may be measured in nanoseconds some day.
Tax starts at $0.01 and doubles every time you do a buy+sell. Counter doesn't reset for 24 hours after your last buy+sell.
When you get these crazy companies doing trades measured in microseconds, this adds up really fast. Think binary. First transaction cost is (2^1-1)*0.01, second is (2^2-1)*0.01, third is (2^3-1)*0.01.. etc.. Those pennies add up. It doesn't stop people from doing short term buy+sells, but it discourages them from doing a bunch of them in a row.
You complain about minor things. I have to import data from Oracle servers and for some reason, Oracle loves to dump out rows with different column counts.
I have to deal with a lot of customers with a lot of different versions of Oracle, and they all seems to have had CSV files that I could not important because the header says 14 columns with their names, but some rows have 5 columns and some rows have 20 columns. Pulled those numbers out of my ass, but that's what happens.. A LOT.. F'n annoying as hell.
Even running select statements against the DB gives me issues like that.. WTF?!
I don't admin Oracle boxes, I just request exports from Oracle DBs from customers and this is what happens quite often when customers try to send us data.
Always fun wasting a week of back-and-forth with a customer just trying to get consistent row schemas.
I expect all hardware I own to run 24/7, not until it dies, but until I no longer have use for it. If my mobo dies when I'm relying on it to work, then it was bad.
"Look, I found this car for $1000 cheaper. The head gasket goes bad every 4 years, but I still save money in the long run."
Go buy a bunch of $50 motherboards and a bunch of $150 motherboards. Now wait 5 years of them running 24/7 and see which ones have the fewest problems. One pays for quality. Go purchase a $25 800watt PSU and load test it, let me know how that goes and tell me quality doesn't matter.
Last time I purchased a $50 mobo, was because I was 13 years old and mowed lawns for money.
Slashdot Mirror
About a year ago, when I was reading about LightPeak(Thunderbolt), Intel claimed it was going to be 10-40gbit and was going to take a few years *after* 22nm became standard.
Just based on Intel claiming LightPeak was meant to come out after 22nm, means it was released early. I am not surprised that the optical version is still some odd years off.
Personally, I think this early release was a mix of Apple and Intel. Apple wanting the fastest and unique, and Intel wanting to make at least some money on their tech as it has been in development for quite a long time. I remember reading about "some optical tech that will scale to 100gbit and become standard", from Intel back near 2005.
When I run a synthetic benchmark on my i7-920(2.66ghz), this is what I get for my Int/DP(64bit float) performance.
5.32bil ints/sec/thread - 8 threads
2.65bil FP/sec/thread - 8 threads - about 21gflops w/o SSE
You should check out the execution units on the i7, there is a lot of duplicated FP/int units, which allows HT to get some pretty hefty performance benefits. SSE units are shared, so any SSE benchmarks would take a hit with HT.
If you can get more than 1 DP/cycle on your AMD chip with 6 "real" cores, let me know; because Intel is getting 1 DP/cycle on its "fake" cores.
Your highest performance with HT would be one thread doing DP and another doing SSE.
I get a 19ms ping to Chicago, which is somewhere upwards of 500mi via the trace route in another state. Put in some more localized rendering farms, like per State, and you could easily keep latency low enough for the average user.
Jitter could be an issue if the network isn't well designed. It would probably show up as micro-stuttering.
Based on what I've read about Raytracing vs Rasterization, Raytracing *will* win out in the long run. I guess RT scales better than rast, but the overhead is expensive. Once we get to the point where RT is about the same speed as rast, it will only take 1-2 generations before RT is several factors faster.
Whichever company is ready to push out RayTracing, will stomp the market. If you release too early, your product will just be a gimmick, if you release too late, the competition will be several times faster than you.
How does this work if someone buys one of these devices for their kid? I thought ALL contracts with children are void and must be signed via the legal guardian.
If the parent never touches/sees/uses the devices in question, how could they be bound to the EULA?
Getting banned for that reason would open them up for another class action suit
It would be like /. putting a ToS that said "You give up your right to reproduce in exchange for using our services"...... ok, bad example.
How would this be legal?
The service packs fix bugs and restructure a little bit. The OS is considered old because its framework/design is 10 years old.
I can rearrange the furniture and paint the walls, but I can't claim my 2001 house is only 3 years old. ohh.. wait.. car analogy.. I can change the oil and tweak the ECU, but I can't claim my 2001 car is only 3 years old.
"Any platform can support H.264 with a small fee"
Wait.. we have to pay for Linux now because of some codec?
We need a completely free codec, not a "cheap" codec.
Until the competition keeps releasing higher and higher bandwidth switches/router and someone comes up with a new killer apps that makes use of it. Then you're left in the dust, and you lose most of your customers. Network admins don't like to have lots of different switch brands, then there's more to keep track of. Cisco falls behind and maybe HP and/or Juniper fill the void, and in the several years it takes to release a competing product, most of their customers have already jumped ship and gutted their infrastructure of anything Cisco.
Humans hate change. There needs to be enough incentive to change, and even more incentive to change back. It will leave people with a sour taste.
Yep, most CEOs like to remove R&D, engineers, programmers. Eventually they're left with a "Dilbert" corp of Managers, Sales, and Lawyers.
About 1/2 of my in-state tuition was paid for by royalties from patents. My in-state tuition was only supported by about 30% tax payers. The rest was by the student or money made from research. If you were out-of-state, you had to pay everything yourself, unless from a sister-Uni.
My state Uni was almost entirely supported by tuition and Alumni donations. Federal funding only went directly to departments or groups that qualified, not to the Uni.
You forget to include the interactions between several HFT computers all cranking away at the same time with-out knowledge of each-other.
How does one HFT tell the difference between true market demand changing and a bunch of other HFT systems tossing money back-and-forth?
copyright != patent
While they could lay claim to the paper, they could never claim ownership to what the paper describes.
"my guess is they're using a custom-developed tool to export the data and it's buggy"
Good point. We'd have to see what's common among the customers, but they always eventually fix it so we never dug further than sending back examples problem and telling them to fix it.
We don't do it ourselves, we tell our customers what we need and this is what happens.. Often. We have about 30 different customers who use Oracle, and nearly all of them have had this issues more than once at some point. A select few have never had this issue.
You say "all day", but everything I read says "milli/microseconds". There is many many magnitudes of difference between what you said and what's happening.
What seems to be happening is there is A LOT of competition for "middle men". As more and more people try to squeeze into this "middle man" market. Because of this, they need shorter and shorter times in order to buy/sell before the competition. This brings us to milli/micro second trade times.
At some point, middlemen become a large portion of the market. What happens when a market is "owned" by middlemen and few long term investors relative to middlemen?
We need middlemen, but we don't need a lot of middlemen clawing as fast as they can to artificially inflate a stock just to see who can dump everything before it peaks to become the "winner".
Have you about the issues of HFT causing mini "bubbles" and pop and crash stocks? Sounds great.
If you *know* people and have the money, you can get your computers at the exchange and they measure those latencies in microseconds. They said they may be measured in nanoseconds some day.
Tax starts at $0.01 and doubles every time you do a buy+sell. Counter doesn't reset for 24 hours after your last buy+sell.
When you get these crazy companies doing trades measured in microseconds, this adds up really fast. Think binary. First transaction cost is (2^1-1)*0.01, second is (2^2-1)*0.01, third is (2^3-1)*0.01.. etc.. Those pennies add up. It doesn't stop people from doing short term buy+sells, but it discourages them from doing a bunch of them in a row.
Or something that scales exponentially.
You complain about minor things. I have to import data from Oracle servers and for some reason, Oracle loves to dump out rows with different column counts.
I have to deal with a lot of customers with a lot of different versions of Oracle, and they all seems to have had CSV files that I could not important because the header says 14 columns with their names, but some rows have 5 columns and some rows have 20 columns. Pulled those numbers out of my ass, but that's what happens.. A LOT.. F'n annoying as hell.
Even running select statements against the DB gives me issues like that.. WTF?!
I don't admin Oracle boxes, I just request exports from Oracle DBs from customers and this is what happens quite often when customers try to send us data.
Always fun wasting a week of back-and-forth with a customer just trying to get consistent row schemas.
It's about as different as a null pointer vs a pointer pointing to an empty string object.
Vista/7 remaps OpenGL calls to DX10/11. Part of the reason why it runs slower.
Why don't they just use CDMA? TDMA/FDMA is quite inferior, although cheaper.
I expect all hardware I own to run 24/7, not until it dies, but until I no longer have use for it. If my mobo dies when I'm relying on it to work, then it was bad.
"Look, I found this car for $1000 cheaper. The head gasket goes bad every 4 years, but I still save money in the long run."
Yeah, if you don't value your time.
Go buy a bunch of $50 motherboards and a bunch of $150 motherboards. Now wait 5 years of them running 24/7 and see which ones have the fewest problems. One pays for quality. Go purchase a $25 800watt PSU and load test it, let me know how that goes and tell me quality doesn't matter.
Last time I purchased a $50 mobo, was because I was 13 years old and mowed lawns for money.