Friday, August 25, 2006

K8L in sample production since June

We now know the next generation quad-core (K8L) described in the June analyst meeting is Rev G. There is a confusion about the names, but Rev G is the one with native quad-core, HT 3.0 and L3 cache (check the link to a report at IDF)*. The chip is code named deerhound, and we were calling it K8L, and we thought it's Rev H, but it is not. Rev G core will be 60% faster, Rev H (2008) will be 150% faster. Hector Ruiz said 2008 will see a real killer, he was talking about Rev H. If your memory doesn't fail you, you should recall that Rev G had been sampled and production ready since June. See Daryl Ostrander's presensation.

Someone analysed the bugs list of Intel's Core 2 Duo (Woodcrest, Conroe). One word, unreliable.

DELL Korea decided to stop selling Intel Xeon servers and switch to AMD Opteron.

* See http://www.itjungle.com/breaking/bn031506-story01.html

"To get to quad-core processors, AMD will be moving to a 65 nanometer process in 2007, which will include a totally revamped Operton core, code-named "Deerhound" apparently and presumably also known as Rev G. These Rev G chips make jump from HyperTransport 1.0 to the HyperTransport 3.0 interconnect (hey, what happened to HyperTransport 2.0?), and a new architecture that incorporates L3 cache. "

24 Comments:

Anonymous Anonymous said...

http://www.theinquirer.net/default.aspx?article=33916

serius bugs found in core2duo.

8:36 AM, August 25, 2006  
Anonymous Anonymous said...

Shakirou said...

...See Daryl Ostrander's presensation.


Very interesting! Can somebody tell me what "mature yields" mean in numbers?

10:23 AM, August 25, 2006  
Anonymous enumae said...

Anonymous said...

"serius bugs found in core2duo."

Do you know how to read?

10:30 AM, August 25, 2006  
Blogger Sharikou, Ph. D said...

By g1ugglug at yahoo message boards:

Conroe to be recalled (17 Ratings) 24-Aug-06 11:35 pm
Check out the Conroe errata document.

http://download.intel.com/design/process...

Some real show-stoppers in there:

AI3 - Address Reported By Machine Check Architecture on Single bit L2 ECC errors may be incorrect:

Translation: The new cores are not suitable for an enterprise server environment because they don't really have ECC.

AI20 - FP Inexact-Result Exception Flag May Not Be Set
Translation: The new Core architecture is unsuitable for scientific applications. Software which tries to get more than long double precision, like SuperPI, will not detect that it needs to increase the precision of the numbers being worked with when rounding occurs.

AI22 - Sequential Code Fetch To Non-Canonical Address May Have Non Deterministic Results
Translation: The New Core Architecture will crash frequently and have other unpredictable (or perhaps exploitable) results when a process is using more than 2GB and dynamically loads a library. A frequent occurence on database servers.

AI28 - EIP May Be Incorrect After Shutdown In IA32e mode
Translation: The new cores have problems waking from sleep in any 64-bit OS.

AI32 - Upper 32 bits of 'From' Address Reported through BTMs or BTSs may be Incorrect
Translation: thunking between 32 and 64 bit libraries within a process requires an ugly OS hack due to this bug. Maybe that's why WinXP 64 didn't support it (but Vista does, wonder how it runs on Core....)

AI36 - Split Locked Stores May not Trigger the Monitoring Hardware

Translation: Software that does frequent Interlockedxxx API calls without #pragma pack(4) or #pragma pack(8) (which are often undesirable for memory or network bandwidth conservation reasons) will slow to a crawl. Affects a few custom server applications that I've worked with.

AI42 - Upper 32 Bits of the FPU Data (Operand) Pointer in the FXSAVE Memory Image may be Unexpectedly All 1's after FXSAVE

Translation: Intel didn't learn from the Pentium FDIV bug. Instead they made a similar bug with implications that totally dwarf it.

AI43 - Concurrent Multi Processor Writes To Non-dirty Page May Result in Unpredictable Behavior

Translation: Multi threaded processes will experience data corruption.

Plenty of other big ones in the doc.

11:49 AM, August 25, 2006  
Anonymous Anonymous said...

AMD Errata Document - http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/25759.pdf#search=%22amd%20processor%20errata%22

Translation - All CPUs have errors.

Love number 52- Voltage Issue
Translation - Your AMD desktop will heat up and explode at will. :)

p.s. - K8L is 2008, not running yet.

1:03 PM, August 25, 2006  
Anonymous Anonymous said...

We now know the next generation quad-core (K8L) described in the June analyst meeting is Rev G

That makes no sense. It's well known that K8L is Rev H. Rev G is still K8 based and is basically a shrink. If a quad core is Rev G based then it is therefore K8 based and not K8L based.

1:22 PM, August 25, 2006  
Anonymous Anonymous said...

The Inquirer responded to their own errate story.

http://www.theinquirer.net/default.aspx?article=33942

First off, this is not a reason to panic or not buy the chips, they are fine as is, the errata lists always get overblown by the fanboi set, and this one is no exception.

Anyways, they say Intel is doing a new stepping to correct some of these errors. Still, they say there is no reason not to buy the chips right now since:

End users will probably see all of no difference.

They seem to suggest that Intel is doing a completely new stepping to specifically correct these problems. I don't think that's the case. They are going to have to do new steppings anyways to release Kentsfield and Cloverton on. It's just a matter of taking advantage of that to correct any critical problems. Errata lists always get blown out of proportion just like Yonah's when it was first released. It's been 8 months since Yonah's release and I haven't heard people suing Intel for selling them defective processors just like people aren't doing so for AMD's Errate list.

1:59 PM, August 25, 2006  
Blogger Kalle said...

"We now know the next generation quad-core (K8L) described in the June analyst meeting is Rev G."

http://en.wikipedia.org/wiki/AMD_K8L
"K8L comes after the Revision G of the AMD64 microprocessors"

I wonder who is wrong ...

2:44 PM, August 25, 2006  
Blogger Sharikou, Ph. D said...

That makes no sense. It's well known that K8L is Rev H. Rev G is still K8 based and is basically a shrink. If a quad core is Rev G based then it is therefore K8 based and not K8L based.

There is a name mismatch there. In any case, Rev G is the quad-core chip described by Phil Hester during the analyst meeting -- which we thought was K8L. If you look at AMD's roadmap, 2007 will see 60% integer performance/core and 3x FP performance, that's Rev G, and it will give AMD a 40% lead in integer and a 200% lead in FP performance. 2008 will see 150% integer performance increase, Hector Ruiz said it will be the real killer.

2:51 PM, August 25, 2006  
Anonymous Anonymous said...

"2008 will see 150% integer performance increase, Hector Ruiz said it will be the real killer."

2008 is even more interesting than this year. By the time K8L comes out, Intel will have Nehalem/Penryn ready.

3:03 PM, August 25, 2006  
Blogger Kalle said...

"In any case, Rev G is the quad-core chip described by Phil Hester during the analyst meeting -- which we thought was K8L"

From Wikipedia K8L page:
"On July 21, 2006, AMD President and COO Dirk Meyer and Senior VP Marty Seyer confirmed that the launch date of K8L (Revision H) microprocessors is slated for mid-2007;"
Could you please give some sources where they say that K8L is rev G?

Some more info on your "reliable" source of information:
http://www.theinquirer.net/default.aspx?article=30042

Concluding from that arcticle I can say that rev G is not K8L.

"200% lead in FP performance"

That would need a situation where you need to make two 128bit loads per cycle and can actually compute fast enough to need those two loads. Not a very common case in my experience. If you know algorithms that have such behaviour feel free to describe them.

There are some speculations that K8L can do only one 128bit store per two cycles. At least that is what AMD data says but the data might be wrong.

Of cource if you meant x87 vs SIMD FP power then that 200% number is probably not big enough :)


Some more info here:
http://www.xbitlabs.com/articles/cpu/print/amd-k8l.html

3:27 PM, August 25, 2006  
Anonymous Edward said...

"Love number 52- Voltage Issue
Translation - Your AMD desktop will heat up and explode at will. :)
"

You didn't read the erreta, did you? The processor "stays in the elevated voltage" longer than expected, until the next software state transition kicks in. That doesn't imply heat up as you suggested - only that Cool&Quiet won't be as effective as intended (well, it's already better than Speedstep).

Besides, Athlon 64 do not explode, even under high load and extremely poor cooling. We tested it ourselves. We stopped the fan and ran games on an Athlon 64 X2 4200+, whose temperature went higher than 80C, then we kept it on for 3 minutes or so until the processor shut itself off.

The best part is, after that we put back on proper cooling and the 2.2Ghz X2 4200+ runs flawless since then. It even still overclocks to 2.6Ghz with air cooling. The quality of that AMD CPU is just one word: perfect.

Sorry for this divergence to the topic, but I have to dispell your FUD that Athlon64 could explode. That's plain false.

3:44 PM, August 25, 2006  
Blogger Kalle said...

More fun stuff!

From modified original post:
"Rev G is the one with native quad-core, HT 3.0 and L3 cache (check the link to a report at IDF)"
Published: March 15, 2006

From wikipedia:
"On July 21, 2006, AMD President and COO Dirk Meyer and Senior VP Marty Seyer confirmed that the launch date of K8L (Revision H) microprocessors is slated for mid-2007"
http://en.wikipedia.org/wiki/AMD_K8L

For the third time I wonder what to believe. This time there are two sources. The one supporting OP oppinion is about three months older.

4:00 PM, August 25, 2006  
Anonymous Anonymous said...

" enumae said...

Anonymous said...

"serius bugs found in core2duo."

Do you know how to read?

10:30 AM, August 25, 2006 "
I'm afraid I dont, Im a down syndrome kid who lives in afghanistan after being bombed the hell out of your country forces :D

anyway try reading by yourself.
cause you didnt seem to do so

5:22 PM, August 25, 2006  
Anonymous Anonymous said...

These Rev G chips make jump from HyperTransport 1.0 to the HyperTransport 3.0 interconnect (hey, what happened to HyperTransport 2.0?), and a new architecture that incorporates L3 cache. "

All current AMD processors use HT2.0. HT1.0 only defines speeds up to 800MHz which was used in the first K8 processors. HT2.0 defines speeds up to 1.4GHz. There is no way current AMD processors could use HT1.0 unless they are shipping out of spec and overclocked, which motherboard makers would never agree to. AMD just doesn't use all of HT2.0's capabilties.

This is also why I've said in the past that it's doubtful that AMD will use the full 2.8GHz speed of HT3.0. Running at the top of spec doesn't leave them with much wiggle room with every motherboard having to be built very well to withstand that. By being below spec they have flexibility and allow overclock room to enthousiasts.

6:49 PM, August 25, 2006  
Anonymous Anonymous said...

please this is clearly misinformation. Rev G is a die shrink and its coming out q4 06 as was clearly stated in the analyst meeting. It may well be that a 4 core version will be released in 2007 which won't be the k8l but it will be a 4 core revg.

7:02 PM, August 25, 2006  
Anonymous Anonymous said...

There are some speculations that K8L can do only one 128bit store per two cycles. At least that is what AMD data says but the data might be wrong.

Yeah your wrong. It has dual 128bit cyclers. It says nothing of doing 2 cycles per 64-bit. Looks like it was made for 2 128bit data packets per clock. It never said anything about having 1.

It will use similar ways of dealing with data the same way conroe does in FULL 128bit data packets, not 2 64-bit data packets like K8's. This is not K8. This is K8L. Your getting your facts mixed up with another ark.

7:49 PM, August 25, 2006  
Anonymous Anonymous said...

Edward said...

"Love number 52- Voltage Issue
Translation - Your AMD desktop will heat up and explode at will. :)"

You didn't read the erreta, did you? The processor "stays in the elevated voltage" longer than expected, until the next software state transition kicks in. That doesn't imply heat up as you suggested - only that Cool&Quiet won't be as effective as intended (well, it's already better than Speedstep).

Besides, Athlon 64 do not explode, even under high load and extremely poor cooling. We tested it ourselves. We stopped the fan and ran games on an Athlon 64 X2 4200+, whose temperature went higher than 80C, then we kept it on for 3 minutes or so until the processor shut itself off.

The best part is, after that we put back on proper cooling and the 2.2Ghz X2 4200+ runs flawless since then. It even still overclocks to 2.6Ghz with air cooling. The quality of that AMD CPU is just one word: perfect.

Sorry for this divergence to the topic, but I have to dispell your FUD that Athlon64 could explode. That's plain false.

I agree. I have 2 AMD systems running 24/7, use them as servers. Never one flaw like I hear in the Intel parts but then it was too late. D: I'd never go intel again... after that bad experionce with my laptop... bad memories.

7:54 PM, August 25, 2006  
Anonymous enumae said...

Anonymous said...

"anyway try reading by yourself.
cause you didnt seem to do so"

Actually I did, and like the article stated, "the errata lists always get overblown by the fanboi set, and this one is no exception".

Well Sharikou is a fan-boi, so I will give it no mind.

Also it doesn't seem uncommon for this to happen, AMD or Intel.

[Link]

9:27 PM, August 25, 2006  
Anonymous Anonymous said...

Dont believe the misinformation the inqiuer writes without naming sources. Remember the reverse hyper threading BS they were bs'ing about. AMD clearly said revg will be a die shrink this q4 this year. quad core in 2007 will be k8l. They are playing with names and dates to make it seem like things are slipping.

10:06 PM, August 25, 2006  
Blogger Kalle said...

is it just me again or the link to the fault-analysis does not work? It redirects me to a page saying " We can't process your request. You may also want to check the URL and try again."

As for the listed erratas...
"AI3. Address Reported by Machine-Check Architecture (MCA) on Single-bit
L2 ECC Errors May be Incorrect
When correctable Single-bit ECC errors occur in the L2 cache, the address is
Problem:
logged in the MCA address register (MCi_ADDR). Under some scenarios, the
address reported may be incorrect.
Implication: Software should not rely on the value reported in MCi_ADDR, for Single-bit L2
ECC errors.
"

Seems like software that relies on the MCi_ADDR is bugged. It shouldn't relie on it. If it does it is its own fault when things go bad.

About AI20, does it occur in SIMD or onlyt x87? If latter then its quite pointless since everyone are moving away from that awful stack based thingie to way faster SSE stuff. Also, as was said most software just ignores these exceptions.

AI22 shouldn't happen if your code is bug-free.

AI28. Sleeping is no problem, just that some special-purose diagnostic software might get wrong results.


They are not planning to fix 3, 20, 28, 32 and 36. Most likely because behaviour in those situations is not defined anyway and software should not be counting on undefined behaviour. Another reason is either that OS or microcode update can fix those things and no new revision is needed.


"Besides, Athlon 64 do not explode, even under high load and extremely poor cooling"

I think last CPU that could burn was AthlonXP. I've run an old 90nm Netburst at full load with 2mm air-clearance between the IHS and heatsink base for 24h. It was in constant throttle around 105C but after reinstalling the heatsink it worked perfectly and it still does more than a year later. Thermal protection in current CPU's is rather good.

12:00 AM, August 26, 2006  
Blogger pointer said...

Besides, Athlon 64 do not explode, even under high load and extremely poor cooling. We tested it ourselves. We stopped the fan and ran games on an Athlon 64 X2 4200+, whose temperature went higher than 80C, then we kept it on for 3 minutes or so until the processor shut itself off.

The best part is, after that we put back on proper cooling and the 2.2Ghz X2 4200+ runs flawless since then. It even still overclocks to 2.6Ghz with air cooling. The quality of that AMD CPU is just one word: perfect.


I agree that nowadays CPU will not smoke. In quite some years back, There was a similar test in Tomshardware with both Intel and AMD CPU. The end result was that AMD CPU smoke and stop functioning after the fan put back, while intel's CPU just halt (without smoke) and run again when the fan is put back. AMD learnt from there and now they should have the proper thermal throtling and i guess that's the reason Edward has to run the test to confirm it :)

5:14 AM, August 26, 2006  
Anonymous Anonymous said...

AMD seems to be having some difficulties with achieving proper speeds at set voltages.

http://www.theinquirer.net/default.aspx?article=33964

10:12 AM, August 26, 2006  
Anonymous Gim Leong said...

"Someone analysed the bugs list of Intel's Core 2 Duo (Woodcrest, Conroe). One word, unreliable."

The link to the Yahoo message board seems to be stale.
Can you give an updated link or a copy of the page?

Thanks!

Gim Leong

6:54 AM, September 03, 2006  

Post a Comment

Links to this post:

Create a Link

<< Home