3 ms·

May not be a winnable argument on HN (where the consensus seems to be that this is nearly-100% Intel's fault), but I think a significant fraction of the blame p

by rphlx 9y ago

May not be a winnable argument on HN (where the consensus seems to be that this is nearly-100% Intel's fault), but I think a significant fraction of the blame properly belongs on their customers, particularly their multi-tenant public cloud customers, for assuming a level of HW thread and guest isolation that has never actually existed. These huge-scale customers did not do proper security due dil on the silicon they were buying and made assumptions that were never, strictly speaking, guaranteed by Intel, and in some cases were explicitly documented as not guaranteed within the Intel optimization manuals.

bsder 9y ago
> These huge-scale customers did not do proper security due dil on the silicon they were buying and made assumptions that were never, strictly speaking, guaranteed by Intel, and in some cases were explicitly documented as not guaranteed within the Intel optimization manuals. But the commodity hardware is SOOOOOO much cheaper, you know? Everything about Intel was always about "good enough" from since probably ... 1982? And "good enough" security in hardware was always "nobody cares". IBM, certainly, was screaming about the level of insecurity in commodity hardware and software forever. DEC similarly. But commodity hardware is SOOOO much cheaper. No one. And I mean NO ONE was every going to give up even 10% on performance or cost in order to be even slightly more secure. At any level of the stack. Intel, Microsoft, Google ... all are guilty of this up until probably this year. Anybody who suggested that would have gotten laughed at and/or fired. The market spoke--and security became an afterthought. Sadly, this is STILL true. While there is much gnashing of teeth about Intel, everybody's implementation of security is far worse. The thing that is biting Intel is that the monoculture means that it is a universal and scalable tool as opposed to at the software level where each individual company has to be compromised in a slightly different way. It's only been since everybody is putting everything in the cloud that people now care about actual absolute values of security. The problem is that all of the hardware solutions to these bugs cost RAM somewhere. And RAM is now the gating factor of cost and performance on most chips. RAM fell off the Moore's Law curve back about 32-22nm and isn't coming back. So, the performance hit to mitigate this is real. No one was going to be the first to fix this EVEN IF THEY KNEW AND CARED. Everybody learned from painful business experience that the first guy gets all the arrows and the second guy gets all the profits. So, everybody was going to wait until they were the second guy--which was only going to happen when something bit everybody.
- rphlx 9y ago
  I would have agreed with you about very poor SRAM scaling until very recently; it looks like 10nm/"7nm" made some decent progress there at last, though nothing like The Good Old Days when it actually scaled close to what the marketing number implied. In any case, it is not clear to me that you need that much SRAM to greatly minimize the various side channels that Spectre uses; it's on the order of a portion of the L1D size, which is pretty minimal compared to how much SRAM there is on-chip already for the LLC.

by rphlx 9y ago

bsder 9y ago

> These huge-scale customers did not do proper security due dil on the silicon they were buying and made assumptions that were never, strictly speaking, guaranteed by Intel, and in some cases were explicitly documented as not guaranteed within the Intel optimization manuals. But the commodity hardware is SOOOOOO much cheaper, you know? Everything about Intel was always about "good enough" from since probably ... 1982? And "good enough" security in hardware was always "nobody cares". IBM, certainly, was screaming about the level of insecurity in commodity hardware and software forever. DEC similarly. But commodity hardware is SOOOO much cheaper. No one. And I mean NO ONE was every going to give up even 10% on performance or cost in order to be even slightly more secure. At any level of the stack. Intel, Microsoft, Google ... all are guilty of this up until probably this year. Anybody who suggested that would have gotten laughed at and/or fired. The market spoke--and security became an afterthought. Sadly, this is STILL true. While there is much gnashing of teeth about Intel, everybody's implementation of security is far worse. The thing that is biting Intel is that the monoculture means that it is a universal and scalable tool as opposed to at the software level where each individual company has to be compromised in a slightly different way. It's only been since everybody is putting everything in the cloud that people now care about actual absolute values of security. The problem is that all of the hardware solutions to these bugs cost RAM somewhere. And RAM is now the gating factor of cost and performance on most chips. RAM fell off the Moore's Law curve back about 32-22nm and isn't coming back. So, the performance hit to mitigate this is real. No one was going to be the first to fix this EVEN IF THEY KNEW AND CARED. Everybody learned from painful business experience that the first guy gets all the arrows and the second guy gets all the profits. So, everybody was going to wait until they were the second guy--which was only going to happen when something bit everybody.

rphlx 9y ago
I would have agreed with you about very poor SRAM scaling until very recently; it looks like 10nm/"7nm" made some decent progress there at last, though nothing like The Good Old Days when it actually scaled close to what the marketing number implied. In any case, it is not clear to me that you need that much SRAM to greatly minimize the various side channels that Spectre uses; it's on the order of a portion of the L1D size, which is pretty minimal compared to how much SRAM there is on-chip already for the LLC.