5 ms·
Also: "Didn’t have the resources to support a customised register file. A synthesised register file resulted in huge congestion when routing the wires from the
by ajross 9y ago
Also: "Didn’t have the resources to support a customised register file. A synthesised register file resulted in huge congestion when routing the wires from the flip-flops. Instead, black-boxed the register file and hand-wrote some Verilog to instantiate specific flip-flops, muxes, and tri-state buffers. Effectively hand-crafting their own bit block out of standard cells."
That just sounds insane to me that the tools can't handle synthesis of straightforward stuff like this. My personal fantasy startup is a next generation, mostly open-source EDA simulation, synthesis/optimization and place & route toolchain. You can expose the analog cell design to the world, but generate revenue from the per-fab process-specific partnering required to work the customers through the process of making real hardware. Who's got funding for me?
- phkahler 9y agoMy comment on the register file is this: instead of making it 6r3w (6 read 3 write), just make it 2r3w and plop down 3 copies of it - each copy has 2 read ports feeding a local ALU and writes go to all instances. Then to add another ALU you make a 2r4w register file and plop down 4 instances of it - one connected to the new ALU. This is still only 6 ports total compared to 9 on the one they've got. Routing within the file should be much easier. The tradeoff is multiple instances of it.
- sitkack 9y agoArea is the new clock speed.
- StringyBob 9y agoAlthough not obvious to consumers, area has always been the bottom line - Gordon Moore used to call the biz 'expensive real estate' (about $1bn/acre)
- ajross 9y agoOh sure. But the text as I read it seemed to imply that the issue was the library they were using had hugely wasteful flip flops or something, or was unable to route above them, or something crazy. It seems (and I'm not expert or professional in this space) that everyone who's used these EDA tools has experiences like this, and hates them. Seems like a great target for innovation at a lower margin, especially now that ASIC development is becoming commoditized.
- _chris_ 9y agoHi phkahler, the ultimate critical path is the load data coming back to the register file. Replicating the register file doesn't end up being a win (there would be so much area to broadcast it out to), otherwise, everybody would do it.
- phkahler 9y agoHi Chris! But it must alleviate the routing problem within the register file no? And the scalability would be better right? I mean, you can't possibly expect to add another ALU to Boom and require 4 write & 8 read ports can you? Does AMD use 12r6w It seems there has to be a better way ;-)
- jabl 9y agoSplit up the physical register file so each pipe has it's own fraction of it, and have the register renaming logic "route" to the correct fraction? Well, that's what BOOMv2 partially does by splitting the INT and FP register files, so I guess I'm proposing splitting the INT file further [1]? [1] This has probably been thought of and dismissed a million times over by actual experts. Say, losing whatever advantages it may have due to extra micro-ops needed to move stuff back and forth to the correct register file fraction?
- _chris_ 9y ago> Split up the physical register file so each pipe has it's own fraction of it, and have the register renaming logic "route" to the correct fraction? Have I got the paper for you! From my adviser's previous student: https://dspace.mit.edu/handle/1721.1/34012 https://dspace.mit.edu/handle/1721.1/34012 But you're right -- the win isn't obvious. There's pain and advantages in both directions. I've always wanted to build it, but I only have so much time in the day. :(
- deepnotderp 9y agoIf you've actually got the EDA expertise for something like that, I can introduce you to people who can get you the funding. Check my profile and email me if you were serious.
- ajross 9y agoHeh, no, it was a joke. Rather, the idea is serious, but my own background wouldn't drive any interest in funding absent a team and a product. I threw it out there half hoping to hear a "me too!" from someone already working on something like it.
- krupan 9y agoCheck these guys out: http://www.symbioticeda.com/ http://www.symbioticeda.com/
- ajross 9y agoCool, thanks. Though they seem focused on the digital side: synthesis and formal verification of HDL code. IMHO that world is served fairly well by existing tooling, as it's fundamentally a software problem that looks like other software problems. Where what I want to see (or do, at least in my dreams) is closer to the analog end: design of parametrized cells, GPU-parallel large circuit analog simulation, circuit-dependent optimization (i.e. choose "speed" if the cell is on a critical path, "area" or "power" if not, optimally choose drive strength and buffers based on latency needs of the path, swap flip flop implementations likewise, etc...) place and route, eventually mask file generation, reverse synthesis and design rule checking (though obviously that part becomes fab-specific). None of this stuff is "hard" in a fundamental way, but it's been hidden behind bad tooling for so long that no one seems to have tried to innovate much in the past few decades. Anyway, I got ideas aplenty.
- deepnotderp 9y agoWell if you do find someone who is working on that please let me know, the future wave of semi startups will thank you :)