AMD should have beaten Intel to every multi-core launch

An AMD dual-die quad-core story from last year generated some letters expressing concern about how that device would work, which made an explanation necessary.

But having done that, it made me conclude that there was an easier, first step approach that should have been done before a dual-die design - the packaged CPU insert card.

Also, now that AMD has got over its quad-core delivery problems and is now shipping that device in volume made it opportune to look back at what the company might have done differently.

I said in that piece that AMD should have launched a dual-die quad-core in December 2006, when the company moved to 65 nanometre manufacturing technology. But I now believe that its dual-core, quad-core and octa-core introductions should have happened much, much earlier.

Market a metric - not the underlying technology
I was a believer in the AMD64 native processor strategy. But AMD was and still is so wedded to it. That strategy would have been fine if the product had arrived and performed as advertised. But it didn’t in the case of Barcelona, and as a consequence the company continues to lick its own inflicted wounds.

With its Netburst processor, Intel had great sales success because the company exploited the frequency advantage that was inherent in the design. Even though AMD had comparable offerings that operated at lower frequency, the chip giant used its marketing clout to convince the market that frequency really was the yardstick.

But the 2001 AMD Athlon XP launch re-introduced the model number nomenclature, which eventually convinced Intel that frequency really wasn’t representative of performance. So as a marketing tool for mainstream CPUs, frequency went the way of the Dodo.

Because the AMD64 architecture easily lends itself to packaged CPU card solutions, as Intel had exploited Netburst ’s frequency advantage, shouldn’t AMD have exploited the core advantage that a packaged card design should have delivered?


Putting the core into AMD64
Putting discrete chips onto an insert card and marketing it as a dual-core, quad-core, octa-core or a hexadeca-core device (16 cores) would have run in the face of the company’s head-in-the-sand native processor strategy. But if the packaged architecture had delivered, the underpinning design really wouldn’t have mattered.

I’m confident that Intel would have said that it didn’t. The chip giant has been selling packaged dual-die quad-cores (two dual-core dies located on the same substrate) in volume for what seems like an age.

From the beginning, AMD should have offered two versions of its top-of-the-line Athlon 64 FX enthusiast CPU - single-socket (FX) and dual-socket capable (FX2). Later on, if the market was ready for more cores AMD could have introduced the FX4 part, which would have supported four sockets.

AMD ’s dual and four socket Opteron chips would have been differentiated from their enthusiast cousins as later models wouldn’t have supported registered memory.

HP, Sun and others may have supported an Opteron multi-socketed insert card. But if the tier one OEMs hadn’t played ball, AMD could have released a reference design for the channel to exploit.

Time to market advantage
A packaged CPU insert card is generally avoided because of cost. But as these multi-core modules could have arrived much earlier than the competition, the additional cost shouldn’t have been a problem.

With standard multi-socket AMD64 motherboard support inherent in the design, why should AMD have offered multi-socket card support as well? Because a packaged CPU insert card offers greater marketability and design flexibility.

The packaged CPU insert card would have had no performance advantage over the regular motherboard types. So this would have been purely a marketing play to be first to market with multi-core technology.

But of course, if AMD had wanted to it could have differentiated the platform in a number of ways. To focus attention on the module architecture, the chip maker could have launched its latest and greatest in a packaged version first. The company could also have increased the processor’s level two cache. If it wasn’t too complicated it may have been possible to introduce an off-chip level three cache as well.

Using a dual socket design, if a packaged CPU architecture had been part of AMD’s platform strategy, dual-core, quad-core and octa-core packages could have been brought to market when the single-core, dual-core and quad-core devices were launched.

Using a quad-socket design, the quad-core part could have launched at the same time as the single-core introduction. The octa-core part could have arrived with the dual-core introduction, and the hexadeca-core package could have debuted with the quad-core launch.

Now I don’t know if the four socket design would have proved a commercial success. So the reason for its inclusion here is to show what was technically doable and to demonstrate the early launch time of these parts.

Intel entered the PC market with the first dual-core device in April 2005. AMD could have been selling a packaged two socket version 20 months earlier - at the September 2003 Athlon 64 launch. The quad-core device, using four sockets, could also have launched at the same time, which would have been three years and three months before the competition - November 2006 Intel quad-core launch. The company’s octa-core and hexadeca-core introductions would have again been first to market.

With such a huge time to market advantage, conceivably, from a mind share standpoint, Intel should have been blown out of the water, and AMD64 would still have been the talk of the town. So the $64,000 question: why didn’t AMD follow such a strategy?

Now people will say that these multi-core introductions may have been far too early for the market to exploit. So what is also being illustrated here is the availability time, not necessarily the launch time, that AMD could have had.

Looking after the enthusiast
As well as off-the-shelf packaged CPU card solutions, AMD should have offered bare insert cards that the enthusiast would have populated.

In the dual-socket case, if this card had been populated with a single-core processor, it should have been possible to first install another single-core device - to make the package dual-core - and later a dual-core part to enable a triple-core system.

It should be noted that the HyperTransport (HT) interface link that AMD processors use comes in two flavors. Regular HT and cHT. Coherent HT technology (cHT) is required for multiprocessor support. Single processors that don’t support multiprocessing only ship with plain vanilla HT. So in the dual-socket case above, cHT devices would be required.

The BIOS would report the number of cores. So if the module only had a single-core device that is what the BIOS would report. For two single-core devices the package would report a dual-core device. For a single and dual-core device, a triple-core device would be reported and so forth.

Buyers that had invested in, for example, a socket 940 single-core processor shouldn’t have had to discard it when the dual-core version entered the market.

AMD should have engineered and validated the parts to be compatible. If the dual-core part was higher in both frequency and HT bus speed than the single core device that it would partner, the dual-core device should have automatically defaulted to the frequency and HT speed of the slower device.

How many enthusiasts who had this triple-core system would have forked out the money to upgrade to an Intel quad-core environment, especially when they still had the option to upgrade their single-core device to dual-core and enable a quad-core system at much less cost than a brand new platform?

Giving real socket and CPU longevity, just like the company did with the Socket 462 platform, would have given AMD customers a good reason to not jump ship to Intel.

The modular, packaged approach
The packaged CPU insert card is nothing new. HP used this design for its DL585 four-way server. Sun used it for its eight-way X4600 counterpart. ASRock sold a motherboard called the K8 Upgrade 760GX. This Socket 754 based design allowed for Socket 939 upgradability by simply inserting an upgrade card and moving over some jumpers.

AMD should have developed two modular designs. One socket based and the other integrated using soldered down devices. Both versions would have had module installed DIMM slots.

Because an insert card offers generous real estate, a dual socket solution would have been possible. As an example, the length of the card above could be doubled. Alternatively, a dual-socket solution could have been developed which used two single-socket cards back to back, which wouldn’t have increased length but thickness would have doubled.

Either design would have allowed the company to offer a dual-core device using single-core devices, a quad-core device using dual-core processors, and an octa-core device using quad-core chips. As discussed earlier, triple-core offerings should have been a possibility as well.

It should have been possible to link two full length cards back to back, which would have doubled the number of devices. So using single, dual or quad-core devices, the four-socket module would have had a respective tally of four, eight or sixteen cores.


Making it all fit
A full length card would have been pretty long and heavy, and two of these back to back in terms of space (not use) would have taken up at least four PCI card positions. So a form factor rethink would have been in order.

The full length card would have had two connector positions per PCB to spread the module weight. Two of these cards back to back would have had four connector interfaces, which would not only provide power, but would be the HyperTransport interface as well.

If the connectivity of a four socket design would have necessitated the use of a four layer plus motherboard, the card to card CPU connections could have been linked directly.

Because of length and thickness requirements, AMD could have gone down the road that Apple took with its Power Mac G5. This would have provided an isolated cooling zone for the CPU module, which would have been cooled by fans. AMD could have given that design the full Apple treatment by liquid cooling the CPUs as well.