AMD needs investors to murder all the ostriches

A new road needed to be chartered for it to come out of the doldrums more or less intact - especially for Hector Ruiz and his cohorts' pay packets, of course. A new road needs a new map - so, as expected, AMD came out with a shiny, new "codename blizzard" road-map!

And we don't even dare think of how a single day would look under renewed Intel total dominance, if AMD went down.

So, beside wishing AMD the best of luck to succeed with this extremely aggressive roadmap, let's see what could really be in it from a competitive point of view. And if there are any pitfalls that AMD could face in completing it. The focus will be mostly be on the high-end.

Let's start from the only major hardware thing that AMD actually showed off - the 3GHz quad-core PhenomFX based on "Agena" core. This is, I believe, a carefully hand picked part, seemingly running the CPU-loaded games well on a triple-card CrossFire GPU setup. OK, this is a BIIIG improvement over those Computex demos seven weeks ago, where the CPU cores of Phenom's metropolitan twin sister, Barcelona, ran at roughly half the speed.

On the other hand, game CPU loads in 32-bit OS are still not near to theh non-stop 100 per cent of, say, high-performance Linpack GFLOPs benchmark under 64-bit Linux or Win64 - I had experience earlier this month where well-overclocked CPUs that could stand such game runs for days, failed that Linpack within few hours, and this is more of the ultimate pre-release stability test.

And those 3.7GHz ++ quad-core Intel CPUs didn't manage it despite water cooling combined with high-RPM fans blowing over the chipset heat pipes, optimised voltage and PLL settings and so on. I had to lower the speed to about 3.4GHz at a voltage higher than usual (1.475 volts) to successfully run that code over a day or two. So, if AMD shows me now 48 hours of Linpack loop running non stop, 64-bit code on 64-bit OS on that 3GHz Phenom, I'll say: "ready to roll, guys!".

And, yes, if AMD really delivers it this year, at least it will have a proper fighting chance agains 3.33++GHz "Yorkfield" Penryns which, as you know, will start shipping well before the end of this years' last quarter - that 3.33GHz is without Intel calling in their 10 per cent or so "frequency headroom reserves", by the way.

If AMD really got the high-frequency leakage solved, as I assume they have been focusing on the past few months, those 2.5++ GHz grade rare limited quad-core yields this year will focus only on the Opteron and PhenomFX, that's workstation/server and top-end enthusiast PC, ranges - the same target as Intel's initial Penryn runs.

A number of completed shipments of such CPUs this year should recover AMD's image somewhat in the eyes of vendors and users, and also get it ready for two particular 2008 exercises, both critical for the company's very survival.

The first one - a fresh performance, price and market-share battle with Intel - not an easy endeavour against a 10x stronger yet nimble and fast-reacting adversary which has learned its lessons well from the NetBurst outburst.

AMD will need the magic combo of performance equal to Intel's top parts, its on-time volume delivery and attractive pricing, not to mention pleasing its miffed early-adopter channel partners, to gain some serious market share here. The first two are particularly important, since Intel could well launch its first Penryns today - the parts are ready, and fabs don't face any delays in ramping up the 45nm lines.

What about the second one? Well, there we come to the possible real main purpose of the Analyst Day paperware codename barrage: scouting for investors. I feel that AMD presentation felt like promises of a company in need of ca$h, and lots of it. The aim is, I feel, to get that cash by early 2008.

Since it would be a long-term funding with no quick returns expected, there must be a convincing roadmap in place, defined on a longer-than-usual timeline, to justify it. The slides, plus of course the expected press droid coverage afterwards, are to play that role.

To keep Fabs 36 and 38 spick and span over the next four or five years - two more process refreshes - AMD needs that cash. And, if you look at those "Asset Lite" slides, the two fabs still figure prominently, despite the increasing number of Chartered and TSMC logos seen on them. As I said before, while there's nothing wrong having the combined fab and foundry model, AMD needs full control over the two fabs if it wants to reliably keep the maximum performance for the flagship CPUs.

All those 45nm enhancements to the ultra-low-K SOI process, followed by high-K metal gate in late 45nm and early 32nm runs, may be developed together with IBM, but I'm not sure that IBM will fund AMD fab retrofitting with these, unless there's something I don't know.

What would be produced on those new lines, the prospective investors would ask? We already know about the 45nm Shanghai single-die 4-core and Montreal dual-die 8-core server CPUs, and Leo single-die desktop entry, all with 6MB L3 cache per die, aimed at sometime in (late?) 2008, but that's just an early appetiser.

Here it is, says AMD - we will Bulldoze(r) the competition into submission! And yes, Bulldozer looks promising - a brand new core, modestly branded as the "fastest in history for both single-threaded and multi-threaded applications", with from 1.7x up to twice the throughput of a Barcelona core, and up to eight cores per die, sharing a single multichannel memory controller, not to mention four HT 3 links to the other chips.

Well, talking about very fast core for single-threaded apps, Alpha EV8, or 21464, was to offer 8-wide out of order superscalar issue on 12 execution units, supporting either 1-thread or 2-thread or 4-thread multithreading on such a wide execution resource. And, yeah, in the old Alpha days it was supposed to offer some 1.7x to 2x the IPC of the already fast 4-wide EV6 core.

Knowing that quite a few Alpha designers still may be sitting in AMD (Dirk Meyer, the EV6 designer, comes to mind first), could the Bulldozer be the X86 equivalent of EV8 Alpha? To achieve AMD's performance jump goal, the new core will have to provide much wider instruction fetch and execution, both integer and FP. More rename registers, executions units, wider internal paths, will all come into play - resulting in around double the core logic size compared to Barcelona core on the same process.

Then, we got to pack up to eight of them on a single die - I'm sure the eight-core "Sandtiger", the very first 45nm Bulldozer-based CPU, aimed at servers and workstations for late 2009 launch (read: 2010 delivery), will be a large die, on a par, if not larger, than 290mm2 of Barcelona.

It may be worth it, though, as Sandtiger would bring along some interesting newbies, like G3MX memory extender for extra DDR3 capacity, and IOMMU (input output memory management unit). Now, IOMMU is interesting, the only similar thing I'm aware of comes from a British product, Quadrics QsNet network - the supercomputer interconnect has a 64-bit network processor sitting on PCI-X bus that integrated a, 64-bit too, MMU, to map the memory space of the whole 1000+ node cluster as the virtual memory of each node. Nifty to, say, do a shared-memory Oracle data mining exercise on a cheap cluster? Well, now this may sit in the CPU.

To win market share, Sandtiger will have to cross the Sandy Bridge - Intel's new name for 'Gesher' Nehalem follow-on 32nm core, expected to be available at about the same time in 2010. Running at well above 4GHz, Sandy Bridge is supposed to be quite massive - we're looking at up to eight out-of-order cores per die, each with two dedicated cache levels, massive instruction-width improvements per core too (7 DP FP ops per cycle for instance, I don't know the integer numbers yet), 64 GB/s memory bandwidth per chip, augmented by up to four CSI links to other CPUs. A 200 GFLOPs/chip monster - I guess Sandtiger will be looking at the similar performance numbers.

If I were IBM, the Bulldozer architecture coupled with improved HT3 and other enhancements like IOMMU could be enough of justification to put Power7 into that same socket, and, quite possibly, shell out some dosh to keep that socket platform "financially healthy". Not to the point of actually buying a huge stake, as it could jeopardise AMD licensing arrangements with Intel, but still chip in for the chips.

Now, if I was Samsung, which previously did try to go into a 64-bit CPU foray with the Alpha, and still may harbour some of these ambitions, on top of excellent process technology - this roadmap would be interesting. Knowing both Samsung and AMD, it would take a hefty sacrifice in both corporate cultures to make anything beyond R&D or foundry partnerships happen.

In the end, the impressive slides with the new semicon processes and even more impressive products to make on them, may just be the reading material for private fund investors, already busy deciding between yet another new tower in Dubai or palace in Shanghai, since London riverfront lots are now mostly taken.

A private investor would, I assume, want to see a decisive product advantage to justify investment in that product. In AMD slides you can see that 45nm "Shanghai" Barcelona follow-on seems to be timed right in front of the first Intel Nehalems, and Bulldozer-based "Sandtiger" is also placed right before Intel's Sandy Bridge. If we follow the past year track record, Intel is expected to come with its lot on time, but AMD may take an extra three or six months more than its roadmaps say - yet that is OK for Hector, by that time the investor's money would be tied in, so they'll have no choice but to stick with it.

In summary, I am impressed with the newest AMD future product slides, and believe that, even if they just come anywhere close to what these slides promise (like, say, being six months behind that schedule), AMD will have a good chance to stabilise and slightly strengthen its position. Again, its has to deliver on the promise, and that assumes certain financial injections. As I mentioned before, to raise the investor's confidence, I believe this will still require a major top management change, very soon.

What about other stuff like new instructions? AMD is preparing quite a few extensions in computing, multimedia and security areas. But, again, these will be of little use from the software support point of view if Intel CPUs don't use them as well. The first two types of extensions will be useful in their Fusion line where, for the first time, they showed the slide of "Falcon", a single-die proposed platform with all the stuff in - Bulldozer CPU core, 3-D GPU and video engine, North Bridge and PCI-Express lanes - a SoC basically.

But that's another story. µ