Intel CPUs afflicted with simple spec-exec vulnerability

[thorhammerz]

5 minutes ago, Arika S said:

For the people who complain about Intel's monopoly on this forum, they sure are happy for Intel to fail and go under giving the monopoly to AMD. 

Nearsightedness afflicts vision in more ways than one .

[dalekphalm]

7 minutes ago, Arika S said:

For the people who complain about Intel's monopoly on this forum, they sure are happy for Intel to fail and go under giving the monopoly to AMD. 

Interesting, isn't it.

 

I'm definitely rooting for AMD, but I don't want Intel to disappear - that would be just as bad as if AMD went under. We need healthy, somewhat equal competition.

[linustechtips fan]

when will intel come out with a les nono meater cpu ??????????????????

[GoldenLag]

20 minutes ago, linustechtips fan said:

when will intel come out with a les nono meater cpu ??????????????????

give it 10-15 years. 

 

while possible, physics and quantum mechanics start making it hard. 

[CarlBar]

2 hours ago, Plutosaurus said:

There's a clear difference in doing everything right, and still failing.

 

 

Sure but there comes a point where if you've made enough mistakes over a long enough time period that you have to question how hard they were looking in the first place. Personally i don't think where there yet, but at the same time this many super long term bugs does raise a lot of eyebrows. Mostly for me it actually comes back to how widespread specter and meltdown where, it's obvious there where industry wide bad assumptions being made, but it's so widespread that saying those where not happening in good faith is a fairly improbable thing. On that basis i think it's reasonable that good faith assumptions that where wrong may have lead to other bad practices also done in good faith.

[Mister Woof]

1 hour ago, CarlBar said:

 

Sure but there comes a point where if you've made enough mistakes over a long enough time period that you have to question how hard they were looking in the first place. Personally i don't think where there yet, but at the same time this many super long term bugs does raise a lot of eyebrows. Mostly for me it actually comes back to how widespread specter and meltdown where, it's obvious there where industry wide bad assumptions being made, but it's so widespread that saying those where not happening in good faith is a fairly improbable thing. On that basis i think it's reasonable that good faith assumptions that where wrong may have lead to other bad practices also done in good faith.

People seem to think that these CPU designs go through R&D in like 6 months.

 

Intel core architecture (the fundamental design that is the target of these attacks) has been in development for over a decade, and it takes a really long time to push out new tech.

 

They just learned about a vulnerability LAST YEAR that affects products as old as ten years, on a fundamental technology design that's over 13 years old. A hardware fix is not just flipping a switch. The issue is an exploit in the fundamental design that needs to be fixed.

 

It takes years for r&d for new designs, years for fabs to accommodate new productions, and years to push it out to the public.

 

People act like it's been one year of Intel knowing, and they have done nothing, when the reality is they've been pushing out microcode fixes and working with Microsoft to push mitigations within days and weeks (or even before it became public) of the issues being known.

 

It will take time for hardware fixes. They've responded very well to the exploits and the expectations of miracle hardware fixes overnight are just ridiculous.

 

Not even counting Intel engineering a solution that doesn't castrate performance, you have to look at fab planning, which takes an eternity, and that's not even counting actual production.

 

Even if they had an engineering fix for meltdown on Jan 1, 2018, it STILL wouldn't be in production today.

 

And then they'd have this new one to worry about.

[leadeater]

9 hours ago, dalekphalm said:

Mistakes? Well, the bugs we've discovered multiple times now. Some of those were not just Intel, and affect all modern CPU design (Example: Spectre + Meltdown situation), but in addition to that, there's the terrible naming conventions that they occasionally decide to use. There's also the whole HEDT fiasco that Linus lambasted them about (They released essentially mainstream grade CPU's with a huge markup on the HEDT platform to "combat" Ryzen).

 

Then, on top of that, there's been multiple lawsuits and patent infringement - though to be perfectly fair, that's an industry wide problem.

 

Intel makes some great CPU's. But they're not even remotely close to a perfect company, and by the time Ryzen was released, they were getting so stagnant as to become anti-consumer. Fortunately AMD kicked them straight in the ass, and they're at least marginally trying now.

Other than the security flaws the big issues over at Intel currently is the 10nm fab processing. Intel is (now was?) known as the industry benchmark for fab technology excellence by a significant amount but recently they've been having big issues with that. Some of that is setting too higher goals but those goals, and meeting them, is why Intel is known for what they are known for.

 

Intel without a fab process lead is just another CPU company.

[Curufinwe_wins]

8 hours ago, dalekphalm said:

snip

Just in case anyone didn't clarify...

 

Qualcomm and MS together managed to emulate x86-32, they have not in any way, shape, or form done the same for x86-64 on arm. While yes, many programs can run on both, and most probably still don't 'need' 64 bit x86 support, it is a speed up these days with the advanced instruction sets, and there are quite a number of programs still out in the cold that won't run if you are relying on a purely x86-32 layer.

 

 

https://arstechnica.com/gadgets/2018/02/windows-10-on-arm-limits-briefly-confirmed-no-virtualization-no-opengl/

[Mira Yurizaki]

9 minutes ago, leadeater said:

Intel without a fab process lead is just another CPU company.

I'd argue Intel having manufacturing capability period especially in their capacity and regardless of whether or not they're in the lead sets them apart.

 

Intel's "direct" competitors are all fabless.

[TopHatProductions115]

Wake me up when it's over...

[leadeater]

8 hours ago, dalekphalm said:

Disagree strong. Bulldozer had 8 execution units that happened to share some resources (Int units, if I recall correctly). You can't use that to say they weren't cores though - because old school CPU's didn't even have Int units. Yes, they were garbage cores, and performed similarly to Intel cores + HT, but that doesn't make it some form of hyperthreading (SMT). They were just shit cores.

They shared float units which was the true big mistake, at the time hardly anyone was making any proper use of vectored code or SSE/AVX so AMD went with an INT strong design, boy that was a mistake. Originally they also shared Fetch and Decode but later dedicated Decode per core was used. Hindsight is a wonderful thing.

 

Bulldozer is 8 cores, it just is and meets the definition of what a CPU core is. SMT shares execution units and caches, CMT has dedicated execution units and caches with shared FLOAT units.

 

Lets play a game of spot the difference 

 

CMT vs Zen core:

 

Zen CCX:

 

Once you group 4 Zen cores in to a CCX it's strikingly similar to Bulldozer/CMT, without shared FLOAT and dedicated Fetch per core. Ignoring all the other vast array of improvements, just a logical view comparison.

[leadeater]

25 minutes ago, Mira Yurizaki said:

I'd argue Intel having manufacturing capability period especially in their capacity and regardless of whether or not they're in the lead sets them apart.

 

Intel's "direct" competitors are all fabless.

True good point, was mostly thinking performance wise or why you would buy their product over another. Intel and AMD, and the industry in general, have good engineers so if you remove the physical aspect of CPU manufacturing it's just a competition of who can design the best architecture relative to desired cost of product. In my view in a market like that all we'd get is something like the differences between ARM based phones, they're all generally the same. Competition and improvements wouldn't slow down but things would end up very similar, with some flip flopping as to which is the current best as software changes etc.

 

Intel's fabing and design customization to that is just so good architecture differences are very much negated, you can clock Intel CPUs so high it makes near as much no difference if Zen/Zen+ is 20%+ better in some aspect when you can just raw out muscle it on the Intel hardware.

[aezakmi]

Here it comes another update that cripples 10% of the singlecore performance lol

[justpoet]

After a brief skim through this thread.  A couple points.

 

1) Nobody with any logical sense is arguing that intel will simply die and cease to exist because of this, financially.  Heck, they're about to add an additional source of revenue with graphics cards.  It does, however, hurt both mindshare and marketshare, as well as their security cred even further than it already had been.

2) "They need to already have taken over your system" is simply not true.

Quote

The issue is separate from the Spectre vulnerabilities, and is not addressed by existing mitigations. It can be exploited from user space without elevated privileges.

So, any sort of guest account or a non-priveledged sandbox such as a web browser with java script (which everything has to have these days with how HTML5/CSS work).  

Quote

[The report] describes a JavaScript-based cache prime+probe technique that can be triggered with a click to leak private data and cryptographic keys...

3) "This is just another thing to affect everybody, like the whole Spectre thing, they're just picking on intel because they're the big name"…nope.

Quote

The researchers also examined Arm and AMD processor cores, but found they did not exhibit similar behavior.

 

In short.  This is yet another big deal.  It isn't the death knell of intel CPUs by any means, and even less so the company as a whole, but it isn't trivial in the scheme of things either, especially since JavaScript can now be used well enough to do this.  I'm sure browsers will add mitigations, and I believe DDR5 has some RowHammer mitigations built into it as well, which will also help make this CPU security vulnerability matter much much less.  But that's not today, and this isn't just a conceptual exercise to be ignored.

[leadeater]

1 minute ago, justpoet said:

-snip-

Agreed, well said.

[Stefan Payne]

46 minutes ago, leadeater said:

They shared float units which was the true big mistake, at the time hardly anyone was making any proper use of vectored code or SSE/AVX so AMD went with an INT strong design, boy that was a mistake.

I disagree, a bit.

It was a big mistake but not the "true big mistake", that was to use the darn slow K10 Northbridge and the darn slow L3 Cache.

When 5% NB OC gives you the same or better performance improvement as CPU Core OC, something is wrong here...

 

The shared Scheduler/decoder doesn't help either...

 

To make matters worse, AMD was able to increase the "IPC" by as much as, well, look at this...
https://www.planet3dnow.de/cms/18564-amd-piledriver-vs-steamroller-vs-excavator-leistungsvergleich-der-architekturen/

A Desktop Version of Excavator wouldn't have been too bad...

Especially with decent I/O Part..

46 minutes ago, leadeater said:

Originally they also shared Fetch and Decode but later dedicated Decode per core was used. Hindsight is a wonderful thing.

Yeah, the AM3+ Version always had the shared decoder the independant Decoder was in Kaveri for example and Carrizo.

 

46 minutes ago, leadeater said:

Once you group 4 Zen cores in to a CCX it's strikingly similar to Bulldozer/CMT, without shared FLOAT and dedicated Fetch per core. Ignoring all the other vast array of improvements, just a logical view comparison.

Yeah in that diagramm.

The Pipeline is a completely different beast.

 

And AMD seems to finally have fixed their weak Northbridge/L3 Design as well...

[Stefan Payne]

1 hour ago, justpoet said:

1) Nobody with any logical sense is arguing that intel will simply die and cease to exist because of this, financially.  

No, but the Shareholders getting pissed and running amock because the revenue tanks.

Possibly some lawsuits because they don't get any revenue...

Stock Price might/will probably also tank and, at worst, somewhere around the $20 Mark...

And at worst even a couple of layoffs...

Quote

It isn't the death knell of intel CPUs by any means, and even less so the company as a whole

No, but a big dent in the Image of the company, so that people see them less as infallible like some people do now.

Thus more people looking at the other side what they might not have done when it didn't happen.

1 hour ago, justpoet said:

3) "This is just another thing to affect everybody, like the whole Spectre thing, they're just picking on intel because they're the big name"…nope.

Yeah, that's what people claimed last time because there was little to no testing done on "alternative Hardware", this time around that is not the case and they also did test other possibilitys.

[leadeater]

1 hour ago, Stefan Payne said:

I disagree, a bit.

It was a big mistake but not the "true big mistake", that was to use the darn slow K10 Northbridge and the darn slow L3 Cache.

When 5% NB OC gives you the same or better performance improvement as CPU Core OC, something is wrong here...

None of that matters compared to the share float, almost all high demand software uses SSE/AVX and all the benchmarks in reviews use it to. You're pitting half the number of resources at the task on an architecture that is weak and relies on scaling to get good performance, with the scaling removed.

 

Would have been very interesting to see CMT with shared int dual float, who knows and meaningless now.

[leadeater]

1 hour ago, Stefan Payne said:

And AMD seems to finally have fixed their weak Northbridge/

Northbridge? Those don't exist anymore, you mean the on die IMC? Northbridge (MCH) and Southbridge (ICH) died 10 years ago. We just have PCH now, or in the case of EPYC none with full SoC.

[Stefan Payne]

8 minutes ago, leadeater said:

None of that matters compared to the share float, almost all high demand software uses SSE/AVX and all the benchmarks in reviews use it to. You're pitting half the number of resources at the task on an architecture that is weak and relies on scaling to get good performance, with the scaling removed.

 

Would have been very interesting to see CMT with shared int dual double float, who knows and meaningless now.

Well, the rumor mill has it that it wasn't planned with that and the Original plan was go with a so called "Technical Float Point Unit"...

 

Anyway, the FPU was improved a bit as well, as you can see in the Link I provided in the AIDA FPU Test, wich is somewhere between +50 up to 100% more on Excavator (= Carrizo + Bristol Ridge).

That Architecture sadly got never backported to the Desktop market...

 

TL; DR:

The Bulldozer Architecture was not bad, it was "just" released too soon, with the shitty, slow K10 Northbridge...

Sadly Bulldozer was already late so another delay might not have been possible but was needed...

1 minute ago, leadeater said:

Northbridge? Those don't exist anymore, you mean the on die IMC? Northbridge (MCH) and Southbridge (ICH) died 10 years ago. We just have PCH now, or in the case of EPYC none with full SoC.

Yeah, that's what AMD called Northbridge at the time. The Intel Name for that was Uncore, wich quickly died.

 

If you happen to get access to a higher end K10 or Bulldozer System and can look into the BIOS; you see some Options that are titled "CPU/NB".

That Part contains Memory Controller and sadly also L3 Cache. Wich wasn't really a Problem for the Server Chip as the difference between CPU and NB Clock was rather small but still it might have improved performance there as well.

[leadeater]

21 minutes ago, Stefan Payne said:

Yeah, that's what AMD called Northbridge at the time. The Intel Name for that was Uncore, wich quickly died.

Very rarely does AMD refer to it as Northbridge, I have seen them do it but that's extremely rare. AMD coined the term Integrated Memory Controller and were the first to do it, referring to that as a Northbridge is just a hang over from the old technology and was done so people could understand what it was. It was also common to mistake the PCH as the Northbridge back then too, hence another reason why people referred to the IMC as Northbridge to help prevent that mistake.

 

Here's an actual AMD slide vs a tech publication labeled one:

 

 

It also didn't help that AMD didn't update their internal driver naming so you'd see things like Northbridge driver in Event Viewer or in driver descriptions, common failing with reusing/iterating on existing things.

[Stefan Payne]

10 minutes ago, leadeater said:

Very rarely does AMD refer to it as Northbridge, I have seen them do it but that's extremely rare. AMD coined the term Integrated Memory Controller and were the first to do it, referring to that as a Northbridge is just a hang over from the old technology and was done so people could understand what it was. It was also common to mistake the PCH as the Northbridge back then too, hence another reason why people referred to the IMC as Northbridge to help prevent that mistake.

That is mostly correct.

For K8 the Term Northbridge was not used and I don't remember that term ever beeing used for a K8 based processor (Socket 754, 939, 940 and also AM2).

 

 

HOWEVER with K10 AMD introduced a split voltage for the Core and the other stuff, they also introduced a new frequency domain for it as well, it was at that time, that the Term "Northbridge" came back. Not in Marketing Documents but internally, proof, for example:

 

41322_10h_Rev_Gd.pdf

Revision Guide for AMD Family 10h Processors

 

Or:

AMD_FX_Performance_Tuning_Guide.pdf

AMD FX Processors Unleashed | a Guide to Performance Tuning with AMD OverDrive and the new AMD FX Processors

 

Quote

CPU NB: CPU NorthBridge (should not be confused with NorthBridge chipset, such as the AMD 990FX chipset) – part of the CPU that has its own clock domain and voltage plane. CPU NB clock frequency determines the Memory controller and L3 cache speed. CPU NB has a notable impact on overall system performance.

 

SO AMD indeed called the thing that Intel called Uncore with the Bloomfield processors "CPU Northbridge".

Those documents are both from AMD and hosted on AMD.com.

[leadeater]

2 minutes ago, Stefan Payne said:

-snip-

You'll need to fix those hyperlinks fyi

[Stefan Payne]

1 minute ago, leadeater said:

You'll need to fix those hyperlinks fyi

I didn't link it for copyright reasons so I don't directly link on it...

Yeah, that stuff is totally broken in Europe...

[leadeater]

14 minutes ago, Stefan Payne said:

I didn't link it for copyright reasons so I don't directly link on it...

Providing the link, or a link itself isn't breaking copyright. If you're worried about copyright it's safer to link than to quote anything at all. 

 

That Barcelona diagram is K10 family Opteron btw, I guess they were less keen on using Northbridge references on the server products. Seems totally weird to bring back that name though.