Cell Server (IBM)

[sunscar]

Current semantics aside... including the fact that we know of nothing on the DCC side that's been ported to work specifically with CELL (things are ported to work fine with Linux, though, even if not my prefered OS), and the fact that we have no actual CELL CPUs in front of us (they exist, but you're not going to just go to Newegg and get a few, ATM)... On paper, CELL CPU clusters Could be good for movie rendering, double precision FP or no, it's never stopped me or anybody else from rendering footage on FP32-only CPUs. Sure, there are reasons why you'd want faster and more precise CPUs, but it's not impossible.

Later

 

[Laa-Yosh]

double precision FP or no, it's never stopped me or anybody else from rendering footage on FP32-only CPUs. Sure, there are reasons why you'd want faster and more precise CPUs, but it's not impossible.

Are there any FP32-only CPUs on the market nowadays? x86 is FP80...

 

[Farid]

Are there any FP32-only CPUs on the market nowadays? x86 is FP80...

By X86 you mean since the Pentium MMX and its 80bits FPU registers, don't you.

 

[Colourless]

x87 has always had 80 bit FPU registers

 

[sunscar]

Yeah, but to what end? I didn't know they could process 80-bit FP numbers directly, thought it had to be done via software abstraction.

 

[aaaaa00]

Yeah, but to what end? I didn't know they could process 80-bit FP numbers directly, thought it had to be done via software abstraction.

The very first 8087 FPU Intel built way back in 1978 supported FP80 and performed full 80-bit floating point computations in hardware. FP80 is there to give the developer support for extra precision when an application demands it.

Intel was behind the push for the IEEE 754 standard, which defines both the well known 32/64 bit and less common 43+/80/128 bit formats for floating point numbers.

Today, no CPU used in any serious floating point application only supports FP32.

 

[Farid]

This X86 family having 80bits FPUs since the beggining stuff had me googling this sunday morning.

http://www.cpu-world.com/CPUs/8086/index.html

Intel 8086 microprocessor is a first member of x86 family of processors. Advertised as a "source-code compatible" with Intel 8080 and Intel 8085 processors, the 8086 was not object code compatible with them. The 8086 had complete 16-bit architecture - 16-bit internal registers, 16-bit data bus, and 20-bit address bus (1 MB of physical memory). Because the processor had 16-bit index registers and memory pointers, it could effectively address only 64 KB of memory. To address memory beyond 64 KB the Intel 8086 used segment registers - these registers specified where code, stack data and extra data 64 KB segments are located within 1 MB of total processor memory. To accommodate this awkward memory addressing many 8086 compilers included 6 different memory models: tiny, small, compact, medium, large and huge. 64 KB direct addressing limitation went away with the introduction of the 32-bit protected mode in Intel 80386 processor.

http://www.cpu-world.com/CPUs/8087/

Intel 8087 is a numeric co-processor for Intel 8086, 8088, 80186 and 80188 processors. The 8087 has 8 80-bit general registers implemented as a stack. All floating point operations are performed with data from the stack (usually the data at the top of the stack) and data from external memory. Effective address calculation for external memory accesses is performed by the main processor. Both the main processor and the 8087 decode floating-point instructions (they all start from ESCAPE bit pattern). If the decoded instruction references the memory then the main processor calculates memory effective address and performs "dummy read" of memory. The Intel 8087 captures the memory address and, possibly, data, and uses the captured address to read more data or write data to memory. After that the main processor may continue to execute integer instructions without waiting until the 8087 completes execution of the FP instruction. To synchronize execution of multiple FP instructions the main processor used WAIT (FWAIT) instruction to wait until the co-processor becomes available. Because the integer instructions and floating-point instructions could be executed in parallel, it was common to see integer and FP instructions intermixed in x86 programs.

Intel was not the only manufacturer of this co-processor. Soviet Union produced clones of 8087 FPU. AMD published information about their version of 8087 in one of their databooks, but it's not clear if any AMD co-processors were produced or not. There are 8087 co-processors stamped "IBM", but it seems that IBM didn't manufacture the chips and only re-branded Intel chips.

 

[PC-Engine]

Yeah but wasn't the math coprocessor integrated into the CPU core later on in the form of the 486DX chips?

 

[aaaaa00]

Yeah but wasn't the math coprocessor integrated into the CPU core later on in the form of the 486DX chips?

Yup. As of 486DX, 8087 FPU capability was integrated onto the main CPU itself.

8087 was architecturally part of x86 since nearly the beginning, but most people didn't need or couldn't afford FPUs back then, so it was an optional add-on.