[hga]

I've been thinking hard about this lately, and the first question for me is "What would a 21st Century Lisp Machine mean?"

Lisp Machines were created in part due to the desire to get the most performance possible back in the days when CPUs were made out of discrete low and medium scale integration TTL (there were also ECL hot-rods, but their much greater costs across the board starting with design limited them to proven concepts, like mainframes of proven value, supercomputers, and the Xerox Dorado, after the Alto etc. had proven the worth of the concept).

Everyone was limited: maximum logic speeds were pretty low, you could try to avoid using microcoded synchronous designs, but e.g. Honeywell proved that to be a terrible idea, as noted elsewhere memory was very dear. E.g. the Lisp Machine was conceived not long after Intel shipped the first generally available DRAM chip, a whopping 1,024 bits (which was used along with the first model of the PDP-11 to provide graphics terminals to the MIT-AI PDP-10), etc. etc.

So there was a lot to be said for making a custom TTL CPU optimized for Lisp. And only that, initially: to provide some perspective, the three major improvements of LMI's LAMBDA CPU over the CADR were using Fairchild's FAST family of high speed TTL, stealing one bit from the 8 bits dedicated to tags to double the address space (no doubt a hack enabled by it having a 2 space copying GC), and adding a neat TRW 16 bit integer multiply chip.

The game radically changed when you could fit all of a CPU on a single silicon die. And for a whole bunch of well discussed reasons, to which I would add Symbolics being very badly managed, and LMI killed off by dirty Canadian politics, there was no RISC based Lisp processor, Lisp Machines didn't make the transition to that era. And now CPUs are so fast, so wide, have so much cache ... e.g. more L3 cache than a Lisp Machine of old was likely to have in DRAM, the hardware case isn't compelling. Although I'm following the lowRISC project because they propose to add 2 tag bits to the RISC-V architecture.

So, we're really talking about software, and what was the Lisp Machine in that respect. Well, us partisans of it thought it was the highest leveraged software development platform in existence, akin to supercomputers for leveraging scientists (another field that's changed radically, in part due to technology, in part due to geopolitics changing for the better).

For now, I'll finish this overly long comment by asking if a modern, productive programmer could be so without using a web browser along with the stuff we think of as software development tools. I.e., what would/should the scope of a 21st Century Lisp Machine be?

[quarterwave]

Thanks for the detailed perspective.

My limited & roseate view of a 21st century Lisp machine is based on an old theme - a massively parallel computing system using bespoke silicon logic blocks.

As you have noted below, not only are the cache sizes in a modern CPU monstrous, there's also the compilers optimized for these caches, instructions, branch prediction units, etc. No point in ending up with a chip that is much slower than an equivalent one running on a specially-designed virtual machine, which is itself much slower than MPI.

Dreaming on, such a Lisp machine would need a vast collaborative academic effort with substantially new IP design, in say the 32nm silicon process node. That's the most advanced node where lithography is still (somewhat) manageable for custom IP design.

[hga]

Well, there's the first Connection Machine architecture, very roughly contemporaneous with Lisp Machines (I had to regretfully tell my friend Danny Hillis that LMI wouldn't be able to provide Lisp Machines for Thinking Machines Corporation in time (which had to be formed because the project needed 1-2 analog engineers, which MIT was structurally unable to pay, no one gets paid more than a professor). He was really, legitimately pissed off by what Symbolics did with Macsyma, a sleazy licensing deal to keep it out of everyone else's hands (and they tried to get everyone in the world who'd gotten a copy of it to relinquish it). Later neglected, even when it became the Symbolics cash cow.)

Anyway, if you're not talking ccNUMA, the limitations of which has got me looking hard at Barrelfish (http://www.barrelfish.org/), e.g. if you're talking stuff in the land of MPI, again it's going to be very hard to beat commodity CPUs.

Although in that dreaming, look at lowRISC: http://www.lowrisc.org/ looking at things now, they propose taping out production silicon as soon as 2016, and say 48 and 28nm processes look good. From the site:

What level of performance will it have?

To run Linux "well". The clock rate achieved will depend on the technology node and particular process selected. As a rough guide we would expect ~500-1GHz at 40nm and ~1.0-1.5GHz at 28nm.

Is volume fabrication feasible?

Yes. There are a number of routes open to us. Early production runs are likely to be done in batches of ~25 wafers. This would yield around 100-200K good chips per batch. We expect to produce packaged chips for less than $10 each.

And with a little quality time with Google, the numbers look good. Ignoring the minor detail of NRE like making masks, a single and very big wafer really doesn't cost all that much, like quite a bit less than $10K.

And we now have tools to organize these sorts of efforts, e.g. crowdsourcing. But it's not trivial, e.g. one of the things that makes this messy is modern chips have DRAM controllers, and that gets you heavily into analog land. But it's now conceivable, which hasn't been true for a very long time, say starting somewhere in the range between when the 68000 and 386 shipped in the '80s.

[protomyth]

I've been wondering about it since like every other programmer I hit that time when I'm really looking at programming languages and VMs (in the "what would I design" sense). Looking to Lisp Machines to see what they were about leads me to the question: would concentrating on hardware memory management / garbage collection be a starting point to answer your question?

[hga]

One indication is that Azul, after 3 generations of ccNUMA systems with zillions of custom chips and a memory infrastructure that gives each one "mediocre" access speed to all the system's memory for running Java with gonzo GC ("generic" 64 bit RISC chips with things like a custom read barrier instruction), has given up and is doing their thing on x86-64 systems with their Zing product, albeit at least initially with tricks like kernel extensions to do bulk MMU operations before a single TLB invalidation. Look up their papers on the Pauseless and C4 GCs. The former was done in time to make it into the 2nd edition of sorts of the book on GC: http://www.amazon.com/Garbage-Collection-Handbook-Management...

Or to put it another way, without exhausting my bank account I could build from parts I can purchase today on Newegg a many CPUs 3/4ths TiB DRAM Supermicro system. Supermicro has standard boards with more memory, and has a monster you can only buy complete that'll hold 4 CPU chips and up to 6 TiB DRAM on daughter boards; I think based on some Googling that has a starting price of less than $35K.

Moore's Law is our friend. But its economics is not the friend of custom CPUs in competition with commodity ones.