Chatter around GPUs for RISC-V is growing The activity around creating a legit graphics processor for RISC-V chip designs, an emerging competitor to x86 and ARM, is gaining steam. Special interest groups at RISC-V next year will expand the focus on extensions for shaders and advanced matrix operations, which is important for artificial intelligence and machine …
> A developer board with a RISC-V chip and a GPU went on sale, but sold out quickly
I'd pick the quad core HF105-000 with 16GiB of RAM on Mouser for €591.50 by SiFive which is expected to ship on the 11th of February (They even have a datasheet for the SiFive HiFive Unmatched). And for it you plug your GPU card into the PCIe Gen3 x8 slot, which in my mind is a better option.
You make it sound like it will be insanely complex.
It is not like a packet router will end up with the Standard Extension for Vector Operations (V) which adds 186 additional Instructions. That would be a total waste of silicon on the chip.
Or that a desktop RISC-V CPU would not have the Standard Extension for Control and Status Register (Zicsr) which adds 6 instructions some of which allow you to query registers to find out who made the chip, what the model is, and which extensions and revisions are supported among other things.
At the end of the day it will be no more difficult than historical software being able to support the x86,286,i386,i486,Pentium, i686, x86_64 and all the various Streaming SIMD Extensions (SSE), SSE2, SSE3, SSSE3, SSE4, Advanced Vector Extensions (AVX), AVX2, AVX-512. The programming is simple, query what is supported and then use whatever gives the optimal performance for the desired application.
Who or what is doing the querying?
Genuine question, I always assumed that support for the various CPU extensions was tied to the OS version not the application? E.g. Windows8 onwards would only be supported on CPUs that had e.g. AVX512?
But are you saying that the OS dynamically intercepts instructions? Or the application queries on installation?
> Who or what is doing the querying?
Does it matter, ultimately the CPU is doing the querying!. RISC-V has three software privilege levels (in increasing order of capability): user-mode (U-mode), supervisor mode (S-mode), and machine mode (M-mode). If one privilege level does not have access, the function can always passed using a secure mailbox to a higher level to run and send back to the lower privilege level what that level is allowed to know depending on the code that was implemented by the designer of the system. What one OS chooses to do/allow may be different to others.
> But are you saying that the OS dynamically intercepts instructions?
That would be slow, it is a valid option, but not necessarily the best.
> Or the application queries on installation?
Travel far back in time when there was a chip called the 8087 which if present it could allowed floating point operations to be done in hardware.
Back in the 80's there were four options for programs that needed floating point:
option 1: Check if the chip was present, and if it was use it.
option 2: Check if the chip was present, and it it was not then exit with a error message that floating point hardware required was missing.
option 3: Check if the chip was present, and it it was not then use a software library to emulate what the chip could do only much much slower.
option 4: Check if the chip was present, then be lied to and told that the hardware is there even though it was not by having the check intercepted by a TSR (Terminate and Stay Resident) ISR (Interrupt Service Routine) , that captured all calls to use floating point hardware and emulated what the chip would do only much much slower.
For programs that needed floating point initially only the first two options existed, some companies added the third option, and others did not. And eventually the fourth option was added.
Me predicting what will happen in the future, I have honestly have no idea. I would expect the option as to what happens at run time to have previously been decided by the compiler options when the program is converted into machine code. I would expect if floating point hardware present in the CPU at runtime use that, if it is missing emulate the functionality in software.
e.g. RP2040 in hardware only supports integers, so the raspberry pi foundation have licensed a software floating point library ( https://www.quinapalus.com/qfplib.html ). In that case the compiler does all, you do not need to worry that the hardware is missing floating point functionality.
...I need an effective and inexpensive PCIe x16, 75 Watt (No external power connectors needed), 64-bit OpenCL processor capable of working on two or more jobs at time. Science needs it even more than I do.
Screw the Graphics processing. Let's build the giant distributed Climate Cruncher.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2023
