AMD's Phil Hester Speaks: Next Generation Architecture, Sort of
by Anand Lal Shimpi on October 14, 2005 6:33 PM EST- Posted in
- CPUs
AMD has been unusually quiet since Intel announced their next-generation microprocessor architecture at IDF just a couple of months ago. AMD argued that they didn't have to talk about a new architecture, as Intel is just playing catch-up to their current architecture.
However, we look at it like this - AMD has the clear advantage today, and for a variety of reasons, their stance in the marketplace has not changed all that much. With a more competitive product, Intel could make it very difficult for AMD, which in our minds is even more reason for AMD to put their best foot forward today, letting the world know that when pushed, they will push back.
Regardless of what we think AMD should do, the fact of the matter is that we haven't heard much about what they are going to do, starting even before IDF. Could it be related to the departure of AMD's Fred Weber? We're not entirely certain, but while in Dresden for the Fab 36 grand opening we were able to speak with Phil Hester, AMD's new CTO.
AMD's new CTO - Phil Hester
Phil headlined a briefing on the future of AMD's microprocessor architecture, which piqued our interest, as it promised to answer questions we've had for a good part of the year. The briefing itself was much more of a Q&A, rather than a black and white outline of AMD's plans to counteract Intel's latest offerings. The one concrete piece of information we did get however, was this table below, outlining AMD's current architecture, what's coming very soon, and in the distant future:
Now | Coming Soon | Future Goals |
AMD64 Architecture |
Extensions to AMD64 | FPU Extensions to AMD64 |
Dual Core Architecture |
Multi-Core Architecture | Throughput Architecture |
Direct Connect Architecture |
Scalable SMP Architecture | On-chip Coprocessors |
Enhanced Virus Protection |
Pacifica Virtulization | Secure Execution |
HyperTransport 1.0 and 2.0 |
HyperTransport 3.0 | HyperTransport 4.0 |
DDR, DDR2 |
DDR3, FBDIMM | DDR4, FBD2 |
AMD PowerNow! Technology |
Partitioned AMD PowerNow! Technology | System Resource Management |
High Reliability RAS |
Mainframe-class Reliability | Best-in-class Reliability |
System Performance |
System Performance per Watt | Throughput per Watt per Dollar |
Note that the "Now" column does include a bit of what's coming soon, mainly support for DDR2. As we've mentioned before, DDR2 support is due for the Athlon 64 next year, along with the upcoming Socket M2 (940-pins) for the desktop and Socket F (1207-pins) for servers. Note that despite the similarity of the M2 socket to the first K8 socket, it will not be pin compatible with current Opterons and/or older Athlon 64s.
The Coming Soon column refers to features that could begin appearing as early as the next twelve months, but obviously items like DDR3 will not be shipping in the next year. Both the Coming Soon and Future Goals columns list extensions to the AMD64 instruction set as things coming down the pipe for AMD's architecture. The idea behind extensions to the AMD64 ISA is to improve performance in specific applications, much like how SSE/SSE2/SSE3 have done in the past.
It is no surprise that talks of multi-core are up next, although the desktop will remain predominantly dual core for the foreseeable future. AMD has often talked about a quad core Opteron, and we'd expect to see the first examples of that definitely by the time 65nm rolls around, although AMD has refused to comment as to whether or not we'd see a quad core 90nm Opteron.
In the next year or so AMD plans on truly showcasing the scalability of their Opteron architecture by providing platforms with support for up to 32-socket configurations, truly stressing the scalability of AMD's Direct Connect architecture.
Virtualization support is another common theme for upcoming AMD architectures, but once again this is no big surprise.
When referring to the future goals for AMD's architecture, the only example Phil Hester provided for FPU Extensions to AMD64 was the idea of introducing extensions that would accelerate 3D rendering. We got the impression that these extensions would be similar to a SSEn type of extension, but more specifically focused on usage models like 3D rendering.
Through the use of extensions to the AMD64 architecture, Hester proposed that future multi-core designs may be able to treat general purpose cores as almost specialized hardware, but refrained from committing to the use of Cell SPE-like specialized hardware in future AMD microprocessors. We tend to agree with Hester's feelings on this topic, as he approached the question from a very software-centric standpoint; the software isn't currently asking for specialized hardware, it is demanding higher performance general purpose cores, potentially augmented with some application specific instructions.
Obviously, as AMD's microprocessors get faster, wider and more powerful, faster interconnects (HyperTransport 4.0) and memory buses (DDR4/FBD2) are always necessary.
Hester also reaffirmed AMD's decision to produce a mobile-specific microprocessor architecture sometime in the future, much like what Intel did with the Pentium M. AMD is still probably years away from a shipping product, but the design teams are currently in a specification defining stage with regards to the project. So while Turion 64 today is hardly anything more than a rebadged Athlon 64 with a tweaked manufacturing process, there is continued support for Turion 64 to grow to be much different than its desktop counterpart.
We also asked Phil what his thoughts were with regards to CMOS based voltage regulators on-package, similar to what Intel demonstrated a couple of months ago at IDF. Much to our surprise, Phil talked about it being prohibitively expensive, and that the expertise required to manufacturer a voltage regulator in CMOS is very different than a microprocessor, thus it's not incredibly easy either. So while Hester didn't completely rule it out as an option, he did not seem nearly as gung-ho about it as Intel did at IDF. Intel did preemptively counter Hester's arguments back at IDF, saying that once you got the voltage regulation into silicon that you could drive costs down through mass production, and the normal scaling associated with Moore's Law.
While it was nice to see some sort of an idea of future direction for AMD, Hester didn't provide us with much more in the way of specifics. We couldn't help but feel that he was holding back on a lot of information, which has been the case with AMD for a lot of recent history. On the flip side, AMD's products are still incredibly competitive, and Intel's next-generation microarchitecture won't make its shipping debut until the second half of next year. That's a great deal of time for AMD to continue to rest on the success of the Athlon 64 and Opteron architectures, and hopefully more than enough time for AMD to give us something more concrete on what is coming.
We want roadmaps, with firm dates and code names and features, and we've yet to see or hear it. So the best we can report today, is what we told you back at IDF: with the exception of some minor updates as well as the migration to DDR2, the Athlon 64 micro-architecture will remain unchanged throughout 2006.
21 Comments
View All Comments
Marlin1975 - Friday, October 14, 2005 - link
Well socket 939 brought out the sempron in 754 for oems. Now there are some early socket 939 semprons, I am thinking sometime early to mid next year we will see M2. Only seems to follow suit.JarredWalton - Saturday, October 15, 2005 - link
AMD told me they expect DDR2-667 to offer about 10 to 15 percent more performance than the best official DDR. So, without other changes AMD was guessing 15% more from DDR2. That means a 2.4 GHz M2 would about match the FX-57, which is pretty good. I think that 15% might also be for dual-core multitasking scenarios, so it might be more like 8% for single core.Quanticles - Saturday, October 15, 2005 - link
Wouldnt that mean memory performance, not necessarily application performance?JarredWalton - Saturday, October 15, 2005 - link
No, AMD stated that overall performance would go up 10 to 15%. Basically, they said DDR2-533 didn't offer enough of a difference to make it worthwhile, but DDR2-667 started to show real benefits of 10% or more. Makes sense, as at 667+ MHz the bandwidth offered is far more than even the best DDR.PrinceGaz - Saturday, October 15, 2005 - link
If that's true, it does pose an interesting question regarding the PR of the new Socket M2 chips compared with existing S939 chips, specifically the Windsor vs the current Toledo-Athlon 64 X2 4800+ - Toledo (2x 1MB L2) clocked at 2.4GHz (12.0 x 200MHz)
Athlon 64 X2 5000+ - Windsor (2x 1MB L2) clocked at 2.667GHz (8.0 x 333MHz)
Now if the DDR2-667 memory was only equal to DDR400, and not 10-15% faster, you would expect the 2.667GHz part to have a rating of 5300+ (or at least 5200+) rather than 5000+. If there is a real-world 10% performance improvement with DDR2-667, then the rating for it should actually be cranked up to about 5800+, instead of lowered to 5000+.
So either the M2 parts will be slightly lower performing because of DDR2-667, or AMD have decided to re-align the PR again. I suspect they are re-aligning the PR though it seems rather pointless doing so, even counter-productive for dual-core chips, as the rating corresponds to nothing that can be compared with Intel parts.
johnsonx - Saturday, October 15, 2005 - link
I don't think the model number of X2 chips is supposed to be a 'PR Rating' at all. What would the rating compare to? It's just a model number where a bigger number = higher performance. I don't think it's supposed to scale linearly either (ie a 5% increase in model number isn't supposed to mean a 5% performance increase... it may be only 2%, or it may be 15%). AMD doesn't need a 'PR Rating' anymore anyway, now that Intel has stopped clockspeed marketing. You're reading too much into the choice of model number, at least for X2 chips.JarredWalton - Saturday, October 15, 2005 - link
The PR numbers come from a battery of benchmarks that AMD runs. The tests have been updated a couple times over the years, but most of the applications are getting a bit old (i.e. little SMP support). So they are relative to performance on other CPUs, though I'm not sure what the baseline is. Some "rounding" does occur, of course. Basically, 4200+ scores about 10% faster than the 3800+ (X2) on the benchmark suite.As for the M2 chips, I hadn't heard the 2.66 GHz figure or model number. Is that a guess, or did I just miss that one? If so, 2.66 GHz X2 would have to rate higher than a 5200+ (unless AMD is just realigning PR again).
PrinceGaz - Sunday, October 16, 2005 - link
Just do a google for "X2 5000+". You'll get tons of hits from all the major hardware-review sites. Oddly enough it didn't appear in AT's news at the time (around Sep 26-27). VR-Zone is a good example.http://www.vr-zone.com/?i=2755&s=1">http://www.vr-zone.com/?i=2755&s=1
----
"AMD Athlon 64 X2 5000+ Info
AMD has confirmed on the Socket M2 design and will be unveiling the Socket M2 processors in Q2 2006. Single-core Athlon 64/FX "Orleans" and Dual-core Athlon 64 X2 "Windsor" will be based on 90nm processor technology and will support DDR-II 667 memory bus, bi-directional 1GHz HT bus, Presidio Security as well Pacifica Virtualization. The Athlon 64 X2 5000+ is based on Socket "Windsor" core and is clocked at 2.667GHz (8.0 x 333.3MHz) and contains 2 x 1MB L2."
ElFenix - Tuesday, October 18, 2005 - link
sounds like the processor i'm waiting forashegam - Friday, October 14, 2005 - link