Yes, your an infant. Now shut the hell up./me Had several flights over the atlantic with asshat kids around me. Nothing like 8 hours of screaming to make your arrival all the more pleasant...
No, annoying is stupid fuckin asians (Li Yu, Yung Bi Fong, etc) from Toronto, calling my VoIP number [also in Toronto] at 2am thinking it's a calling card #.
Look you stupid "gooks", put the calling card # in your phone's memory. Then memory dial the damn thing. God damn...
How you tell a FOB from a native? Natives will apologize for the wrong #, fobs just hang up on you....
Can you do a 32x32 multiply-add => 96-bit accumulator in less than four opcodes? [16 bytes with ARM]. Even the x86 can do that in 5 opcodes [load, mul, add, adc, adc] and takes about the same space. Sure you can accomplish the operation, just is it efficient or not?
In many cases where MIPS are being used the hard work is offloaded. Which isn't a bad thing. It just doesn't speak volumes about the core as a highly efficient design [in general].
If your idea of running a solid board is just hosting the server and letting anything fly you're really no better than a pin board hanging at a bus stop or something. Expect to get spammed.
It takes a split second to tell if a post is spam or not. Unless your forum is getting 1000s of posts a day [in which case you could also delegate the work out], you can easily sift through a pile of posts in a few minutes.... of course that requires thought, and as we all know, thinking is a damn right dangerous proposition for most.
Why? It encodes a single pass at ~60fps, it's free software, it does a very good job with libavcodec of making mpeg4:ac3 movies, etc.
I was encoding three movies at a time for a while. All hitting ~50fps per pass [25fps overall]. Good enough for me. Oh, yeah I guess I could pay for [or steal] a commercial encoder. But since I only have to encode my DVD once, it doesn't really matter.
Um? are you kiddin? I have a 2P 285 box [dual core 2.6Ghz] and in certain applications [e.g. video encoding] only a single core is used [damn you mencoder!!!]. Why would I clock up the other cores? Now cpufreqd does clock down the other PROCESSOR but the other core on the same processor as the mencoder process is clocked up.
You are right that in certain high load applications you may not need it. But remember for every live server in the world there are dozens of test boxes which take power just the same. Those test boxes RARELY have anywhere close to 100% load and could benefit greatly.
Instead of making two processors, one which can idle and one which can't, you might as well just make one that can idle.
The cell SPEs are also very large. 256KB of full speed SRAM, DMA engine and a SIMD engine. That's a lot of transistors.
Also the paper you linked to is just about having spare cache areas [quadrants] which are permentantly routed to [via a flash bit] when a defect is found.
So your 4MB cache may actually be [say] 5MB of cells. You can tolerate a loss of upto 1MB [at most, depending on the spread]. So you're still using more die space and you have mitigated the defect rate so you "lose less".
By making the die larger it means fewer fit on the wafer which raises the cost, but you lose fewer dies to defects, reduces the cost. On the other hand, if you just make the damn thing smaller to start with you put more on the wafer and yield more with the same defect rate.
It really depends on how much the redundancy costs. If it's only say 5% of the die then it's probably a better idea. If it's like 30% it may not be.
Still it has to cost something to expect a 4MB cache. It's not free.
1. Comes on the same SIZE MEDIA as before [check]
- no media advantage like VHS vs DVD... hmm 2. Comes with new added value DRM that makes hardware incompatible with one another? [check] 3. Are nowhere close to being compatible with one another [check] 4. Are only different so they can vendor lockin consumers [check] 5. Are fundamentally no better than before [check] 6. Still hold the same shitty two-bit plot movies [check] 7. Players are expensive [check]
- DVD players cost like 50$ now. And if you run Linux they're free 8. The terminology is confusing for most and awkward [check]
- What the hell is progressive scan? interlaced? HDCP? is blu-ray HD? What the hell is HD-DVD then? What is HDMI? etc
Tips
1. Merge the damn standards or deprecate one of them 2. Drop the DRM crap 3. Pay your actors less and your writers more 4. Host a website like "hdfacts.org" where people can learn about the tech in a vendor neutral environment 5. Give up the proprietary crap already
Ahem EM64T vs. x86_64? Pacifica vs. VT, 3DNOW vs SSE?
Intel isn't exactly known for making exact equivalent implementations of anything AMD comes out with. Yeah it would be good if CSI was just HT with their own cache coherency protocol (hint: AMD keeps their cache control specs out of the HT consortium), but who knows what they'll actually do.
In all reality, [this is my personal opinion] it would be good if BOTH AMD and Intel worked together on socket specifications. That way I can upgrade my AMD box with Intel cores and vice versa. Instead of having to scrap the entire box just to buy a new processor.
Imagine if WD, Samsung, etc all had their own SATA specs? That'd be a hell of a nightmare!
That makes no sense at all. Cache transistors are... um... transistors. They may be specially formed [or laid out] because of the nature of cache memory but they're transistors none the less.
If your cache is sporadically functioning you have to scrap the die. You can't sell a 4MB cache where a dozen cache lines aren't functioning. That'd lead to a really unstable design and it's technically fraud.
Yes, you can lower the cache size by cutting off ways [or banks]. E.g. 4MB => 2MB => 1MB etc... but that still counts as a lower yield.
If you wanted to make 350 4MB cores and you ended up with 275 4MB cores, 25 2MB cores and 50 duds, you still have a yield of only 275 since your cash cow is those processors. You sell off the 2MB cores to STEM your loses, not turn a profit.
In this regard I think it's more incompetence than anything else. I know at my ISP they're a bunch of MCSE reject windows toting retards. They removed usenet support because "nobody was using it" and they were talking about packet shaping [specially bt and voip] but haven't yet.
Chances are the whole idea of local mirrors for popular content just hasn't sunk in yet.
1. Will CSI be HT compatible? Where is the spec? Will they take part in a consortium like HT?
2. The revision of the 754-pin and 939-pin processors were COMPATIBLE. None of this "your i845 northbridge is no good, get a 915, oh your 915 is no good, get a 945, oops, I meant 955, no 965 no 975!!!!". I went through several revisions of AMD64s in my Asus K8V motherboard including the dual-core 4200 and 4800 series.
3. Yes, it's nice that it's smaller but transistors still cost money. You have to get all 201,326,592 transistors in the cache correctly masked, etched and formed.
A smaller cpu means you get more per waffer, cost--
While I appreciate the sentiment of buying = voting, you also should be buying what you actually need. If that happens to be an Intel or Samsung or whatever processor [Samsung makes ARM processors...] so be it.
I agree that AMD is a smart choice both on technological and political fronts. But for the general public with a 1P configuration the benefits of AMD can be lost on them.... I also happen to think more people need 2P setups:-) nothing like encoding two movies at high quality, two pass mode at 30fps each simulaneously to show off the power of the Opteron.
No, they "should" use that. Doesn't mean they will.
Chances are they'll do one of two things
1. Roll out their own CPU interconnect with a bridge to HTX 2. Roll out their own HT but call it ZippyDataTransit or say Lightning Data Transport [hehehe chuckle, *] and claim it's 1.97x faster than AMDs
Tom
[*] Bonus points for anyone who can tell me why the latter name suggestion is funny.
You do realize that many of these "leaks" are on purpose to get kids and stupid people foaming at the mouth for the "finished product" right?
Yeah, it's an accident that a 100 million dollar studio keeps losing track of their works in progress... right, cuz a firewall and SSL server would be so hard.
1. Get with direct link interconnects, FSB is teh stupid 2. Stop making a new core every other Tuesday, m'kay? 3. 4MB of cache is nice, but it has to be hella expensive right? [*] 4. Merge with Nvidia, totally mess up the PC scene, it'll be fun:-)
[*] Don't look at the retail cost for the true margins they make [if any] on the cores. Selling at a loss or near loss is not a new tactic.
I'd hardly praise C# as a language that should be included with an OS.
I mean a real compiler where I can do real work in. All these new langs are a fad. Look at what went on with Java, all of a suddent PHP comes out and it's like "they don't solve the same problems but I simply cannot resist the urge to write everything in PHP now!" then ASP shit then.NET same stuff over and over.
All the while good ol' C and C++ are still the driving force of this software and other stuff anyways.
I'd rather have a good C compiler bundled with my OS than any other compiler/language combo. At least then I could use it to develop applications which are portable, hence broaden my possible customer base and make more money like a good capitalist.
Slashdot Mirror
I could do with less whiny little brat kids.
/me Had several flights over the atlantic with asshat kids around me. Nothing like 8 hours of screaming to make your arrival all the more pleasant...
Yes, your an infant. Now shut the hell up.
Tom
Where's the Linux lab? :-)
Nothing like enforcing your monopoly like buying out the schools... or at least making them think you bought them out...
Tom
No, annoying is stupid fuckin asians (Li Yu, Yung Bi Fong, etc) from Toronto, calling my VoIP number [also in Toronto] at 2am thinking it's a calling card #.
Look you stupid "gooks", put the calling card # in your phone's memory. Then memory dial the damn thing. God damn...
How you tell a FOB from a native? Natives will apologize for the wrong #, fobs just hang up on you....
Tom
Yeah which means bloat?
Can you do a 32x32 multiply-add => 96-bit accumulator in less than four opcodes? [16 bytes with ARM]. Even the x86 can do that in 5 opcodes [load, mul, add, adc, adc] and takes about the same space. Sure you can accomplish the operation, just is it efficient or not?
In many cases where MIPS are being used the hard work is offloaded. Which isn't a bad thing. It just doesn't speak volumes about the core as a highly efficient design [in general].
Tom
Um what? MIPS lacks a add with carry instruction. That makes it suck. :-)
Tom
Why not just moderate the board to start with?
... of course that requires thought, and as we all know, thinking is a damn right dangerous proposition for most.
If your idea of running a solid board is just hosting the server and letting anything fly you're really no better than a pin board hanging at a bus stop or something. Expect to get spammed.
It takes a split second to tell if a post is spam or not. Unless your forum is getting 1000s of posts a day [in which case you could also delegate the work out], you can easily sift through a pile of posts in a few minutes.
Tom
Why? It encodes a single pass at ~60fps, it's free software, it does a very good job with libavcodec of making mpeg4:ac3 movies, etc.
I was encoding three movies at a time for a while. All hitting ~50fps per pass [25fps overall]. Good enough for me. Oh, yeah I guess I could pay for [or steal] a commercial encoder. But since I only have to encode my DVD once, it doesn't really matter.
Tom
Um? are you kiddin? I have a 2P 285 box [dual core 2.6Ghz] and in certain applications [e.g. video encoding] only a single core is used [damn you mencoder!!!]. Why would I clock up the other cores? Now cpufreqd does clock down the other PROCESSOR but the other core on the same processor as the mencoder process is clocked up.
You are right that in certain high load applications you may not need it. But remember for every live server in the world there are dozens of test boxes which take power just the same. Those test boxes RARELY have anywhere close to 100% load and could benefit greatly.
Instead of making two processors, one which can idle and one which can't, you might as well just make one that can idle.
Tom
64K code+data.
Not 32/32.
Tom [not official...]
On die memory controller, HT links and yeah it's 90nm.
At 45nm they're going quad-core. I can't tell you the config exactly but they're not leaning in Intels direction exactly.
Tom
The cell SPEs are also very large. 256KB of full speed SRAM, DMA engine and a SIMD engine. That's a lot of transistors.
Also the paper you linked to is just about having spare cache areas [quadrants] which are permentantly routed to [via a flash bit] when a defect is found.
So your 4MB cache may actually be [say] 5MB of cells. You can tolerate a loss of upto 1MB [at most, depending on the spread]. So you're still using more die space and you have mitigated the defect rate so you "lose less".
By making the die larger it means fewer fit on the wafer which raises the cost, but you lose fewer dies to defects, reduces the cost. On the other hand, if you just make the damn thing smaller to start with you put more on the wafer and yield more with the same defect rate.
It really depends on how much the redundancy costs. If it's only say 5% of the die then it's probably a better idea. If it's like 30% it may not be.
Still it has to cost something to expect a 4MB cache. It's not free.
Tom
1. Comes on the same SIZE MEDIA as before [check] ... hmm
- no media advantage like VHS vs DVD
2. Comes with new added value DRM that makes hardware incompatible with one another? [check]
3. Are nowhere close to being compatible with one another [check]
4. Are only different so they can vendor lockin consumers [check]
5. Are fundamentally no better than before [check]
6. Still hold the same shitty two-bit plot movies [check]
7. Players are expensive [check]
- DVD players cost like 50$ now. And if you run Linux they're free
8. The terminology is confusing for most and awkward [check]
- What the hell is progressive scan? interlaced? HDCP? is blu-ray HD? What the hell is HD-DVD then? What is HDMI? etc
Tips
1. Merge the damn standards or deprecate one of them
2. Drop the DRM crap
3. Pay your actors less and your writers more
4. Host a website like "hdfacts.org" where people can learn about the tech in a vendor neutral environment
5. Give up the proprietary crap already
Tom
Ahem EM64T vs. x86_64? Pacifica vs. VT, 3DNOW vs SSE?
Intel isn't exactly known for making exact equivalent implementations of anything AMD comes out with. Yeah it would be good if CSI was just HT with their own cache coherency protocol (hint: AMD keeps their cache control specs out of the HT consortium), but who knows what they'll actually do.
In all reality, [this is my personal opinion] it would be good if BOTH AMD and Intel worked together on socket specifications. That way I can upgrade my AMD box with Intel cores and vice versa. Instead of having to scrap the entire box just to buy a new processor.
Imagine if WD, Samsung, etc all had their own SATA specs? That'd be a hell of a nightmare!
Tom
That makes no sense at all. Cache transistors are ... um ... transistors. They may be specially formed [or laid out] because of the nature of cache memory but they're transistors none the less.
If your cache is sporadically functioning you have to scrap the die. You can't sell a 4MB cache where a dozen cache lines aren't functioning. That'd lead to a really unstable design and it's technically fraud.
Yes, you can lower the cache size by cutting off ways [or banks]. E.g. 4MB => 2MB => 1MB etc... but that still counts as a lower yield.
If you wanted to make 350 4MB cores and you ended up with 275 4MB cores, 25 2MB cores and 50 duds, you still have a yield of only 275 since your cash cow is those processors. You sell off the 2MB cores to STEM your loses, not turn a profit.
Tom
In this regard I think it's more incompetence than anything else. I know at my ISP they're a bunch of MCSE reject windows toting retards. They removed usenet support because "nobody was using it" and they were talking about packet shaping [specially bt and voip] but haven't yet.
Chances are the whole idea of local mirrors for popular content just hasn't sunk in yet.
Tom
Why don't ISPs that worry about their net usage outside their network just mirror shit?
Would it be really hard to throw together a 1TB file store with the latest patches, demos, ISOs and the like?
That way the customers can get stuff inside the network and the ISP doesn't have to worry about upstream net usage.
OMG it's like I'm smart and all.
Tom
1. Will CSI be HT compatible? Where is the spec? Will they take part in a consortium like HT?
2. The revision of the 754-pin and 939-pin processors were COMPATIBLE. None of this "your i845 northbridge is no good, get a 915, oh your 915 is no good, get a 945, oops, I meant 955, no 965 no 975!!!!". I went through several revisions of AMD64s in my Asus K8V motherboard including the dual-core 4200 and 4800 series.
3. Yes, it's nice that it's smaller but transistors still cost money. You have to get all 201,326,592 transistors in the cache correctly masked, etched and formed.
A smaller cpu means you get more per waffer, cost--
More transitors lowers your yield though, cost++
Tom
Non-compete clauses don't last forever.
And aren't enforceable everywhere.
Tom
You win bonus points. To be spent as you see fit at any bonus store...
Um...
Righto
While I appreciate the sentiment of buying = voting, you also should be buying what you actually need. If that happens to be an Intel or Samsung or whatever processor [Samsung makes ARM processors...] so be it.
... I also happen to think more people need 2P setups :-) nothing like encoding two movies at high quality, two pass mode at 30fps each simulaneously to show off the power of the Opteron.
I agree that AMD is a smart choice both on technological and political fronts. But for the general public with a 1P configuration the benefits of AMD can be lost on them.
Tom
No, they "should" use that. Doesn't mean they will.
Chances are they'll do one of two things
1. Roll out their own CPU interconnect with a bridge to HTX
2. Roll out their own HT but call it ZippyDataTransit or say Lightning Data Transport [hehehe chuckle, *] and claim it's 1.97x faster than AMDs
Tom
[*] Bonus points for anyone who can tell me why the latter name suggestion is funny.
Yeah that's why it would be fun. I think we need more vendor-lockin nowadays. If we can't beat them join em.
/me looks to employer, how about we support OSS? /me takes that back, wants to keep job.
And yeah, it is nice that Intel is more pro-OSS...
Tom
You do realize that many of these "leaks" are on purpose to get kids and stupid people foaming at the mouth for the "finished product" right?
Yeah, it's an accident that a 100 million dollar studio keeps losing track of their works in progress... right, cuz a firewall and SSL server would be so hard.
Tom
1. Get with direct link interconnects, FSB is teh stupid :-)
2. Stop making a new core every other Tuesday, m'kay?
3. 4MB of cache is nice, but it has to be hella expensive right? [*]
4. Merge with Nvidia, totally mess up the PC scene, it'll be fun
[*] Don't look at the retail cost for the true margins they make [if any] on the cores. Selling at a loss or near loss is not a new tactic.
Tom
I'd hardly praise C# as a language that should be included with an OS.
.NET same stuff over and over.
I mean a real compiler where I can do real work in. All these new langs are a fad. Look at what went on with Java, all of a suddent PHP comes out and it's like "they don't solve the same problems but I simply cannot resist the urge to write everything in PHP now!" then ASP shit then
All the while good ol' C and C++ are still the driving force of this software and other stuff anyways.
I'd rather have a good C compiler bundled with my OS than any other compiler/language combo. At least then I could use it to develop applications which are portable, hence broaden my possible customer base and make more money like a good capitalist.
Tom