Category: AI

Ben Thompson of Stratchery in his recent blog on Intel Split prompted me to coin the word “Intelligrated“, which is a counterpoint to his thesis. – No, its not in the dictionary. Before we get to that, lets start with one topic he brings up as it is near and dear to me and many of my old fellow chip nerds from that time (1987-2003) which I would call as EDA 1.0 era.

EDA changed microprocessor roadmap starting Circa 1987 and continued through late 1990s: Ben references Pat’s paper on Intel EDA methodology which scaled design methodology to track moore’s law. Intel invested heavily in EDA internally as the industry was immature. Around the same time Sun Microsystems which built its business selling EDA/MCAD workstations was changing the industry EDA landscape (methodology and eco-system). [An aside: Would not be surprised if x86 CPUs till Pentium IV, were designed using Sun workstations]. Both companies had parallel but different approaches.

EDA 1.0: Intel vs Sun approach: Sun’s approach was industry tools and if it does not exist enable the industry eco-system to be built. It perhaps started in 1989 when a motley crew (25) of engineers (including yours truly) built the first CMOS SPARC SOC (Tsunami – referenced here) with no prior experience in custom VLSI. We all came out of a cancelled ECL SPARC microprocessor where none of us had done any custom VLSI design. The CAD approach was…

Necessity is the mother of invention. Sunil Joshi captured the EDA methodology then in the MicroSPARC HotChips presentation. 486 (Pat’s chip) had 1.2M transistors (100+ engineers perhaps) vs the more integrated MicroSPARC at 800K transistors that came 2 years later (same time as Pentium which had 200+ engineers) but a full SOC. AS noted, we had only 2 mask designers and every engineer had to write RTL, synthesize, verify, do their own P&R, timing analyze and two of us built the CAD flow so that it can be push button. Auto-generated standard cells were automatically P&R using compiler tools. That was not the norm for ‘custom VLSI’ circa 1991 for that scale.

That eventally got the name ‘Construct by correction vs Correct by construction’ and throughout the 1990s, this evolved, scaled and made the processor design competitive as well raised a lot of boats in the EDA industry that evenutally got Intel to adopt industry tools with a healthy mix of in-house tools. With no in-house EDA teams, we creatively partnered (Synopsys, Mentor), invested (Magma, mid 1990s), helped M&A (gateway design – verilog, Cooper and Chyan – Cadence), spin-out (Pearl timing analyzer to replace motive). At the same time IBM EDA tools which were superior to both Sun’s approach and Intel, but was locked up inside IBM until later in the decade, when it was too late.

In parallel, there was a great degree of systems innovation (SoCs, glue-less SMP, VIS, Ethernet, graphics accelerators, multi-core+threading) that was enabled by EDA 1.0 and CMOS custom VLSI by the industry at large with Sun leading the parade. Allow me to term it as ISM 1.0 (Integrated Systems).

Now IDM 1.0 is what made Intel successful to beat all the RISC vendors. We (Sun and compatriot RISC vendors) could beat Intel in architecture, design methodology and even engineering talent and in some cases like Sun which had OS and Platform built a strong moat. But, we could not beat Intel on manufacturing technology. Here’s a historical view of tech roadmap.

Caution: Given dated history, some data from 1990s could be incorrect – corrections please notify.

In a prior blog I have called out how Intel caught up with TI and IBM on process technology by 1998 (they were the manufacturing leader but not the technology leader w.r.t xtor FOM or metal litho until 1998). TI Process technlogists used to complain ‘At Intel design follows fab and you folks are asking fab to follow design’ as we demanded xTOR FOM and Metal litho more than Moore in the 1990s. By 1998 with coppertone, Intel raced ahead with both litho as well as xtor FOM (60% improvement in 180nm with coppertone to boost Pentium MhZ). So Intel was not xtor FOM leader in early 1990s, but they pulled ahead by 1-2 generations by late 1990s. It has been done by Intel. When they did go ahead is when IBM and TI became non-competitive (starting 1998) for high end microprocessors and the beginning of our end and my departure from microprocessor (unless I go to Intel). (Side note: Both TI and Intel bet on BiCMOS till 650nm). Unlikely history will repeat itself as the dynamics are different today with consolidation, but it has been done before by Intel.

Intel’s leadership with IDM 1.0 and co-opting ISM 1.0 architectural elements (by 2002 – multi-core, glue-less SMP, MMX, integrated DRAM controllers) into its processors made it difficult for fabless CPU companies to thrive despite having systems business to fund – which was not sufficient by 2002. Even 500K CPUs/year (Sun/SPARC) was not economically justifiable. IBM, SGI, HP and many more esp. dropped as cost of silicon design and tech went up. [Side note: I am not sure on a standalone basis Graviton is economically viable for Amazon – if 500K CPUs was not viable in 2000, 50K is certainly not viable in 2020 – but sure they can wrap other elements in the TCO stack to justify for a few generations – not sustainable over 3-5 generations). Regardless, 20 years later…

IDM 2.0 is a necessary first step and good that Pat & Intel are putting that focus back. But IDM 2.0 needs ISM 2.0 and the same chutzpah of late 1980s of design EDA innovation but this time perhaps own the ‘silicon systems platform SW stack’.

ISM 2.0 is ‘Integrated Systems 2.0’. If SoC were the outcome of Moore’s law in the 1990s, SoP (Systems on Package) is the new substrate to re-imagine systems as the platform is becoming heterogenous for compute (CPU, DPU, IPU, xPU), Memory (DRAM, 3D xpoint, CXL memories) and Networks (Ethernet and CXL). There will always be a CPU (x86 or ARM), but increasingly we will find a DPU/IPU/XPU in the system that sweeps all the new workloads. The IPU/DPU/GPU/XPU will increasingly be an SoP with diversity of silicon types to meet the diversity of workload needs. But it will need a common and coherent software model to be effective in enabling platforms and workloads including low level VMs or run-time with standard APIs (e.g. P4/Sonic, Pytorch+TF and others).

On economies of scale which killed the RISC revolution (amongst other reasons), I have written about SoC vs SoP in a different blog here, but its important to consider the diversity of customers from OEM platforms to cloud platforms, emerging telco/edge service providers, and emerging ML/AI or domain specific service providers that have a large TAM. Each one needs customization i.e. no more one size fits all platforms, its multiple chips into multiple SoPs to different ways to package and deliver to these new channels of delivery – OEM (Single System), Cloud (distributed systems) and emerging decentralized cloud. But to retain the economies of scale of chip manufacturing while delivering customized solutions to the old and new category of customers, we are moving towards disaggregate at the component level and aggregation at the platform software level.

Just to get a better sense of varied sales motion- Nvidia is a chip company, a box company (mellanox switches and DGX) as well as a cloud company (delivering ML/AI as a service w/Equinix).

This is more than Multi-core and Virtualization that happened in Circa 2003 (VMware). An entire new layer of software will be a key enabler for imagining the new ‘integrated systems’ and delivery of them. For lack of a better TLA , let me call it EDA 2.0. The design tooling to assemble these variety of SoP solutions requires new design tooling to enable customization of the ‘socket’. The old mantra was sell 1 chip SKU in millions. That is still true. But we will have multiple SoP SKUs using the same multi-million unit chip SKU. The design tooling to assemble these SoP has not only manufacturing programmability but in the field as there will be some FPGA elements in the SoP as well as low level resource management functionality.

Hijacking the OSI model as a metaphor to represent the infrastructure stack…..

The homogenous monolithic CPU has now become heterogenous CPU+DPU/IPU/XPU/FPGA. Memory from being homogenous DDR to DDR + 3Dxpoint on DDR bus and CXL bus.

So ‘Integrated Systems’ is an assembly of these Integrated chips based on target segments and customers, but have to manage the three axes of flexibility vs performance vs cost. While silicon retains that one mask set for most target segments (economics), the customization at the packaging level enables the new ‘integrated system’ (the bottom three layers in above visual) . This new building block will become complex over time (hardware and software) thus value extraction (simplification of complexity results in value extraction) but requires capital and talent. Both exists ironically either with the chip vendor or with the cloud service provider, the two ends of the current value chain.

The pendulum is starting swing back from the vertically integrated chip company (1980-2000) to the era when OEMs owned Silicon (Sun, HP, IBM) or chip companies owned fab (Intel), to Horizontalization with the success of fabless chip (Nvidia, Broadcom……) + TSMC (2000-2020) to again vertical integration at sub-system level for certain segments or markets.

Back to Intel Split vs Intel Integrated. In this era, if there is any lesson learnt from the EDA 1.0 era, it would be smarter to do a build, buy and partner i.e. build the new tooling (EDA 2.0 and BIOS 2.0) in a smart combination of build, buy, partner and expand into invest, open source and build a moat around that eco-system that will be hard for competitive chip companies to compete. EDA 2.0 is not the same as EDA 1.0 – its both design tools pre-manufacturing and low level programming and resource management frameworks. Directionally some of it is captured here by Chris Lattner ( MLIR & CIRCT). We have a chicken and egg situation that to create that layer we will need new silicon constructs, but to get to the right silicon, you will need the new layer (akin to how Unix+C enabled RISC and RISC accelerated Unix and C eventually referenced here..)

Coming back to Intel v TSMC and splitting, TSMC is good at manufacturing – but has not (yet) built eco-systems and platforms at higher levels. Its their customers (fabless). Intel knows that and done that many times over. I make the case that Intel Integrated with IDM 2.0 and ISM 2.0 and being flexible in delivery SoC, SoP and even rack level products to emerging decentralized-edge cloud providers will the emerging opportunity.

Splitting the company will devalue the sum of parts than being whole in this case. While, the point of split (fab) driving accountability and customer focus and serviceability is there, there are perhaps other ways to achieve the same without a formal split while retain the value of integration of the parts.

Smart integration of the various assets and delivery. Creatively combining design engineering, EDA 2.0, BIOS 2.0 and linking it to SoP and SOC manufacturing including field level customization will be a huge value extraction play. The Apple vs Android model of market share vs margin share. IDM 2.0 with ISM 2.0 will get market share and margin share for dominant compute platforms.

A POV – Intel has do 3 things. IDM 2.0 (under way), ISM 2.0 (will elaborate that in a future note) and something else ( aligned with Intel’s manufacturing DNA) truly out of the box before 2030 when its going to be economically and physically hard to get more from semiconductors as we know. That has to be put in place between now and 2030…

————————————————————-

References to EDA at Sun…

• http://www.cecs.uci.edu/~papers/compendium94-03/papers/1995/dac95/pdffiles/02_4.pdf

Historical CPU/Process Node Data..