I get your point. Compatibility is definitely important (to the people who approve big purchases of tech), because app vendors are stupid slow to react to a new architecture, and port their software to it.
IA64 was faster over all than x86 clock for clock… Unless you had to run non-natively compiled code. The x86 emulation layer in it was not great. Anything compiled for the architecture ran faster and better, but the only software vendor who seemed to even attempt to release anything for it was Microsoft. There was a full IA64 build of Windows server, and I believe they even built exchange for it too; possibly more, like mssql…
But as far as I’m aware, they’re the only mainstream vendor who tried. So when someone wanted the IA64 server to run the QuickBooks server components and everything sucked harder than a $2 hooker on crack, the problems became immediately evident with the platform.
It’s actually really good that we’re moving away from natively compiling software for CPU architectures. Yes, it may run slightly faster that way, but CPU speed is rarely the problem with modern computing. It stopped being the primary motivator for CPU purchasing around the core i-series, 4th Gen. I would maintain that a 4th Gen platform would run fine today despite the obvious deficiencies, if we could cram it with enough RAM that goes fast enough to keep up, and a quick SSD. In fact, I have a handful of systems that are still running on that platform today and they’re working great. With the extensions included in more modern CPUs, the need for compute speed is even less. AES is now handled in hardware. Which is one of the major slowdowns for older CPUs on the modern web. I could go on, but I think I’m making my point quite well.
Except for a handful of edge cases where every ounce of performance matters, it’s basically irrelevant to have more compute power. A lot of the bottlenecks of those systems are because of their interconnects. Fix that and you would have a viable platform.
If anyone needs proof of this, we need not look any further than the trend towards SBCs. A raspberry Pi can run a lot of things far better than older hardware at a fraction of the power consumption. With trends towards architecture agnostic software, it’s becoming more and more viable to use smaller, more power efficient systems to do the same work.
There will almost always be a place for big iron, but for most, smaller is better. It does what you need it to do, and that’s it.
Frustrating part A is that we have a universal binary format… and it’s HTML5. Frustrating part B is that nobody with a purchasing department wants to admit it. Slack ships with its own browser like you don’t have one. Modern web games can run on a sufficiently fancy Amiga, yet there have been Electron apps without a Linux version. That Amiga’s gonna suffer overhead and unstable performance, but I mean, so do native Unreal 5 games.
The good ending from here would be a period of buck-wild development. RISC-V, MIPS, finally doing that guy’s Mill CPU. I was gonna say that neural networks might finally get high parallelism taken seriously, but no, optimized matrix algebra will stay relegated to specialty hardware. Somewhere between a GPU and an FPU. There’s server chips with a hundred cores and it still hasn’t revived Tilera. They’re just running more stuff, at normal speed.
The few things that need to happen quickly instead of a lot will probably push FPGAs toward the mainstream. The finance-bro firehose of money barely splashed it, when high-frequency trading was the hot new thing. Oh yeah: I guess some exchanges put in several entire seconds of fiber optics, to keep the market comprehensible. Anyway, big FPGAs at sane prices would be great for experimentation, as the hardware market splinters into anything with an LLVM back-end. Also nice for anything you need to happen a zillion times a second on one AA battery, but neural networks will probably cover that as well, anywhere accuracy is negotiable.
Sheer quantity of memory will be a deciding factor for a while. Phones and laptops put us in a weird place where 6 GB was considered plenty, for over a decade. DRAM sucks battery and SRAM is priced like it’s hand-etched by artisanal craftsmen. Now this AI summer has produced guides like ‘If you only have 96 GB of VRAM, set it to FP8. Peasant.’ Then again - with SSDs, maybe anything that’s not state is just cache. Occasionally your program hitches for an entire millisecond. Even a spinning disk makes a terabyte of swap dirrrt cheap. That and patience will run any damn thing.
I get your point. Compatibility is definitely important (to the people who approve big purchases of tech), because app vendors are stupid slow to react to a new architecture, and port their software to it.
IA64 was faster over all than x86 clock for clock… Unless you had to run non-natively compiled code. The x86 emulation layer in it was not great. Anything compiled for the architecture ran faster and better, but the only software vendor who seemed to even attempt to release anything for it was Microsoft. There was a full IA64 build of Windows server, and I believe they even built exchange for it too; possibly more, like mssql…
But as far as I’m aware, they’re the only mainstream vendor who tried. So when someone wanted the IA64 server to run the QuickBooks server components and everything sucked harder than a $2 hooker on crack, the problems became immediately evident with the platform.
It’s actually really good that we’re moving away from natively compiling software for CPU architectures. Yes, it may run slightly faster that way, but CPU speed is rarely the problem with modern computing. It stopped being the primary motivator for CPU purchasing around the core i-series, 4th Gen. I would maintain that a 4th Gen platform would run fine today despite the obvious deficiencies, if we could cram it with enough RAM that goes fast enough to keep up, and a quick SSD. In fact, I have a handful of systems that are still running on that platform today and they’re working great. With the extensions included in more modern CPUs, the need for compute speed is even less. AES is now handled in hardware. Which is one of the major slowdowns for older CPUs on the modern web. I could go on, but I think I’m making my point quite well.
Except for a handful of edge cases where every ounce of performance matters, it’s basically irrelevant to have more compute power. A lot of the bottlenecks of those systems are because of their interconnects. Fix that and you would have a viable platform.
If anyone needs proof of this, we need not look any further than the trend towards SBCs. A raspberry Pi can run a lot of things far better than older hardware at a fraction of the power consumption. With trends towards architecture agnostic software, it’s becoming more and more viable to use smaller, more power efficient systems to do the same work.
There will almost always be a place for big iron, but for most, smaller is better. It does what you need it to do, and that’s it.
Frustrating part A is that we have a universal binary format… and it’s HTML5. Frustrating part B is that nobody with a purchasing department wants to admit it. Slack ships with its own browser like you don’t have one. Modern web games can run on a sufficiently fancy Amiga, yet there have been Electron apps without a Linux version. That Amiga’s gonna suffer overhead and unstable performance, but I mean, so do native Unreal 5 games.
The good ending from here would be a period of buck-wild development. RISC-V, MIPS, finally doing that guy’s Mill CPU. I was gonna say that neural networks might finally get high parallelism taken seriously, but no, optimized matrix algebra will stay relegated to specialty hardware. Somewhere between a GPU and an FPU. There’s server chips with a hundred cores and it still hasn’t revived Tilera. They’re just running more stuff, at normal speed.
The few things that need to happen quickly instead of a lot will probably push FPGAs toward the mainstream. The finance-bro firehose of money barely splashed it, when high-frequency trading was the hot new thing. Oh yeah: I guess some exchanges put in several entire seconds of fiber optics, to keep the market comprehensible. Anyway, big FPGAs at sane prices would be great for experimentation, as the hardware market splinters into anything with an LLVM back-end. Also nice for anything you need to happen a zillion times a second on one AA battery, but neural networks will probably cover that as well, anywhere accuracy is negotiable.
Sheer quantity of memory will be a deciding factor for a while. Phones and laptops put us in a weird place where 6 GB was considered plenty, for over a decade. DRAM sucks battery and SRAM is priced like it’s hand-etched by artisanal craftsmen. Now this AI summer has produced guides like ‘If you only have 96 GB of VRAM, set it to FP8. Peasant.’ Then again - with SSDs, maybe anything that’s not state is just cache. Occasionally your program hitches for an entire millisecond. Even a spinning disk makes a terabyte of swap dirrrt cheap. That and patience will run any damn thing.