Pentium 4 6XX Sequence and New EE P4s Launched

Mojo-Dog writes "Today Intel took the wraps off their new Pentium 4 Processors with EM64T extensions for 64-bit computing. The Pentium 4 6XX Sequence and Pentium 4 3.73GHz are based on Prescott 2M cores with a full 2MB of on-chip L2 cache as well. HotHardware.com has a full review with benchmarks posted of these new P4s, many of which also offer Intel's SpeedStep technology for power savings and improved thermals, which has been available in Pentium Mobile CPUs for some time now."

Comments

Re:2MB Cache?

[Grounded0]

MIPS R12000 system that's sitting on my desk has 8MB of L2 cache. And yes, it's circa 2000.

Re: EM64T Extensions - 64-bit computing?

[Grounded0]

Basically they're just IA-32 architecture without it's most worst design errors.

1. 8 registers increased to 16 (it still sucks compared to SPARC's 128).

2. Larger addressing width (eg. can allocate more than 4GB of memory limited by 32-bit architectures). Alpha and MIPS had this capability in 1992.

3. NX bit (can prevent buffer overflows). Has been available for ages on good CPU architectures.

Intel fanboys? Biased Journalism?

[inflex]

Normally I don't pay much attention to these reviews, but damn this review smacked of Intel fanboyism and anti-AMD'ism. In summary, the comments fell into two catagories:

1. If Intel beat the AMD in a test
"Once again it's game over for AMD"

2. If AMD beats Intel in a test
"AMD struggles to keep ahead of Intel in this test"

I thought at first it was just a one off comment - but the almost all of the evaluations were like that.

Obviously we each tend to have a preference for one brand over another but please can we have consistent commenting.

Paul.

Re:64-bit GPUs

[cnettel]

You're wrong. Or, rather, "bitness" is a very silly measure. In a general purpose chip, you can measure the maximum word size for single operations.

Then, you realize that the current SSE/3Dnow etc stuff will actually handle 128-bit data.

Then, you can think that you should measure the bandwidth of the memory bus. With dual channels, that's generally 128 bits now for CPUs, but for Intel, the memory bus is of course still a part of the chipset. Most GPUs top out at 256, with lower counts and basically the same architecture for the cheaper models. The front-side bus in Intel chips is 64-bit, but running on a higher frequency. Also, most accesses, IIRC, are aligned to be the size of one cache line - 64 bytes or 512 bits. Also, note that the 8088 was an 8-bit CPU and the 80386 sx a 16-bit CPU by this definition. Obviously not what we want.

Finally, we can measure it by the addressing model. This makes some sense and then we also get to the result that AMD64 was the first x86-like ISA to achieve 64-bit flat space addressing. The "flat space" requirement is important, as we want to consider the 8086 (/8088) 16-bit and not 20-bit (16-bit segment + 16-bit offset with locked segment spacing). In this area, many GPUs are tailored to their actual memory capacity. Why should we waste addressing bits and consequentially lines on stuff we can't use?

By this definition, a modern GPU isn't "even" 32-bit, but why the heck should we care. The number of bits as a performance metric is stupid unless one has to take extra measures to avoid the boundary. That was the case in 16-bit x86 code, and is currently the case in some heavy-iron 32-bit code. The number of bits "of" a GPU is not a relevant metric.