AMD Phenom L3 cache bug explained

[Yousuf Khan]

There was a bug detected in the L3 cache of the newly introduced AMD
Phenom and Barcelona quad-core processors. It's apparently what's
causing the delay of the introduction of faster versions of the Phenom &
Opteron(so far it's only available upto 2.3Ghz). Also it's resulted in
AMD having to reintroduce newer, faster versions of the previous
generation Athlon X2 processors to fill the gap.

DailyTech - Understanding AMD's "TLB" Processor Bug
"How does the TLB erratum occur? All AMD quad-core processors utilize a
shared L3 cache. In instances where the software utilizes nested memory
pages, this processor will experience a race condition. "
http://www.dailytech.com/Understanding++AMDs+TLB+Processor+Bug/article9915.htm

 

[Ed]

There was a bug detected in the L3 cache of the newly introduced AMD
Phenom and Barcelona quad-core processors. It's apparently what's
causing the delay of the introduction of faster versions of the Phenom &
Opteron(so far it's only available upto 2.3Ghz). Also it's resulted in
AMD having to reintroduce newer, faster versions of the previous
generation Athlon X2 processors to fill the gap.

DailyTech - Understanding AMD's "TLB" Processor Bug
"How does the TLB erratum occur? All AMD quad-core processors utilize a
shared L3 cache. In instances where the software utilizes nested memory
pages, this processor will experience a race condition. "
http://www.dailytech.com/Understanding++AMDs+TLB+Processor+Bug/article9915.htm

Sad we won't see anything worth buying till March, and even then I would
probably want to wait another 3 months to make sure the boards and bios
have ALL the bugs worked out like I always do before upgrading.

Intel's 45nm chips will be out for 3 months by March and may be the
better bang for the buck though from what I've read their boards won't
have much of a future upgrade path compared to AMD's AM2/AM2+ socket
boards.

Comments?

 

[Yousuf Khan]

Sad we won't see anything worth buying till March, and even then I would
probably want to wait another 3 months to make sure the boards and bios
have ALL the bugs worked out like I always do before upgrading.

Intel's 45nm chips will be out for 3 months by March and may be the
better bang for the buck though from what I've read their boards won't
have much of a future upgrade path compared to AMD's AM2/AM2+ socket
boards.

Comments?

It may sound strange, but I don't think it really matters how late
Phenom comes out. AMD has rediscovered its roots as the cheaper
processor maker. Over the last year, it has been regaining market share
over Intel, despite overwhelming technical superiority from Intel. It's
now clawed back almost all of the market share that it lost to Intel
with its initial misstep from last year. It had 25% market share last
year, it went down to 18% after the disastrous misstep, and now it's
already back upto 23% again. The disastrous misstep was to become a
supplier for Dell and ignore the Channel, but it was not the lack of
performance against Core 2 as most people think; Dell overestimated how
much it could sell, and end up dumping its unsold inventory back to AMD.
I don't think AMD will be so accomodating to Dell's forecasts anymore.

During all of this time it's had some financial issues due to the ATI
acquisition. But even ATI is on the comeback, and it's now contributing
positive cashflow to AMD.

Also I saw a roadmap today from AMD which talks about a "Large L3 cache"
coming in 2008. That may be Z-RAM. Separately, there was even a rumour
that any Z-RAM cache could be upto 100MB!!!

Check page 30 of this PDF, where the roadmap stuff starts:
http://www.cisl.ucar.edu/dir/CAS2K7/Presentations/torricelli.pdf

Next Generation CPUs Could Have 100MB of Cache - Z-Ram technology might
enlarge the cache memory by a factor of 10 - Softpedia
"The company also said that the actual Z-RAM can store "significantly
more charge in the memory bitcell," which translates into lower read and
write cycles. According to Innovative Silicon, Z-RAM can surpass the 1
GHz barrier. The firm said that the memory can store 5 Mb per mm2 when
using 65 nm process and more than 10 Mb per mm2 using 45 nm
manufacturing process. That could easily translate into L2 cache sizes
of about 50 to 100 MB if applied to the die-sizes of actual CPUs.
"
http://news.softpedia.com/news/Next-Generation-CPUs-Could-Have-100MB-of-Cache-41830.shtml

Yousuf Khan

 

[Robert Myers]

It may sound strange, but I don't think it really matters how late
Phenom comes out. AMD has rediscovered its roots as the cheaper
processor maker.

You mean AMD is competing on price. That's not good for anyone.
Commodities are sold on price and commodity markets are not known for
technological innovation.

I'll give AMD credit. I expected AMD to be headed out to sea on a rip-
current by now. It's not completely clear that it's not, but it's
still above water (barely). It's a hole in the ocean into which you
(or people with more money than they know what to do with) pour
capital. Would AMD have that money if Intel weren't so heavily
invested in Israel?

As to AMD's cute little "bug," I suppose that everyone here knows that
AMD published benchmark results with the bux-fix turned off, so the
chip that actually works is significantly slower than advertised.

Anything good to say? Yes. IBM and Power.

Robert.

 

[Jan Panteltje]

On a sunny day (Thu, 6 Dec 2007 13:37:02 -0800 (PST)) it happened Robert Myers
<[email protected]>:

As to AMD's cute little "bug," I suppose that everyone here knows that

AMD published benchmark results with the bux-fix turned off, so the
chip that actually works is significantly slower than advertised.

I have just read there is a Linux kernel with a fix for the bug that
does _not_ slow down.
They emulate the accessed- und dirty-bits bits in software.
https://www.x86-64.org/pipermail/discuss/2007-December/010260.html

 

[Ed]

Also I saw a roadmap today from AMD which talks about a "Large L3 cache"
coming in 2008. That may be Z-RAM. Separately, there was even a rumour
that any Z-RAM cache could be upto 100MB!!!

Check page 30 of this PDF, where the roadmap stuff starts:
http://www.cisl.ucar.edu/dir/CAS2K7/Presentations/torricelli.pdf

At AMD's last analyst day meeting there was no mention of Z-Ram that I
can recall. That pdf is old.

They claimed the first AMD64 K8s were capable of 8MB L2 caches but they
never did more then 1MB per core, probably because of yields, die size,
and the performance gains just weren't noticeable enough, their
architecture just doesn't seem to benefit much from larger caches, even
the new chips with L3 cache don't appear to be making any big impact.

Ed

 

[Yousuf Khan]

I have just read there is a Linux kernel with a fix for the bug that
does _not_ slow down.
They emulate the accessed- und dirty-bits bits in software.
https://www.x86-64.org/pipermail/discuss/2007-December/010260.html

It seems to be something that happens only under virtualization anyways.

Yousuf Khan

 

[Jan Panteltje]

It seems to be something that happens only under virtualization anyways.

Yousuf Khan

Not only under virtualisation, it seems *especially* under virtualisation.
http://www.heise.de/english/newsticker/news/100231
Gernam, according to C'T:
http://www.heise.de/newsticker/meldung/100186

 

[Yousuf Khan]

You mean AMD is competing on price. That's not good for anyone.
Commodities are sold on price and commodity markets are not known for
technological innovation.

It's resulted in a huge expansion of the PC market. It's grown to levels
never seen before. Typically there's 100 million PCs sold in a year,
this year it may hit 200 million. So the effect is exactly what's
needed, PC's are becoming more affordable. I've never seen laptops as
cheap as they are nowadays.

I'll give AMD credit. I expected AMD to be headed out to sea on a rip-
current by now. It's not completely clear that it's not, but it's
still above water (barely). It's a hole in the ocean into which you
(or people with more money than they know what to do with) pour
capital. Would AMD have that money if Intel weren't so heavily
invested in Israel?

If you're talking about Abu Dhabi's Mubadalla investment in AMD, you can
call it an Arab vs. Israeli thing. Both Mubadalla and AMD are sponsors
of the Ferrari F1 team. So I think it has more to do with Formula 1
business connections.

As to AMD's cute little "bug," I suppose that everyone here knows that
AMD published benchmark results with the bux-fix turned off, so the
chip that actually works is significantly slower than advertised.

The bug only shows up under virtualization, no other time.

Yousuf Khan

 

[Yousuf Khan]

At AMD's last analyst day meeting there was no mention of Z-Ram that I
can recall. That pdf is old.

It's likely they aren't going to say too much about it until they get it
out. Something like this will either work or not work, and they don't
want to get too much buzz around it if it doesn't come out. But they had
licensed the technology back in 2006, and if it takes 2 to 3 years to
develop it, then 2008 is just about right time to start seeing it.

They claimed the first AMD64 K8s were capable of 8MB L2 caches but they
never did more then 1MB per core, probably because of yields, die size,
and the performance gains just weren't noticeable enough, their
architecture just doesn't seem to benefit much from larger caches, even
the new chips with L3 cache don't appear to be making any big impact.

You mean like the small impact of 10-20% drop in performance if the L3
is reduced in speed or cut out entirely?

Yousuf Khan

 

[Robert Myers]

It's resulted in a huge expansion of the PC market. It's grown to levels
never seen before. Typically there's 100 million PCs sold in a year,
this year it may hit 200 million. So the effect is exactly what's
needed, PC's are becoming more affordable. I've never seen laptops as
cheap as they are nowadays.

Just like my life is better because I can't go anywhere without
listening to somone else's cell phone conversation. Every idiot's got
one. The conversation is in greater volume and emptier than ever,
only now I have to listen to it.

I can pretty easily make a credible model that will show you that if
you pay a little bit higher prices now, and the higher price goes into
R&D, you will get higher net performance at lower cash outlay over the
long haul. I don't know who thinks the world is made a better place
by selling lots of cheap PC's, but I don't belong to the club. I
don't know why I didn't say that to the low-price AMD boosters long
ago and drop the subject forever. I'd be just as happy to see Intel
and IBM charging higher prices.

Robert.