NVIDIA's 64-bit "Denver" Tegra K1 Should Be Amazing

[89c51]

Probably yes, since 32 bit version works already.

Didn't know that. How well this works?

[emblemparade]

Nvidia will likely use the same core in successive chips - I'm guessing a quad core next. Then there might be a chip with a Maxwell-based GPU.

Well, maybe. The problem with Denver is that the die size is huge ... quad core will have to be quite enormous, maybe too big for mobile applications.

[ravyne]

Denver core size isn't an issue.

Well, maybe. The problem with Denver is that the die size is huge ... quad core will have to be quite enormous, maybe too big for mobile applications.

I wouldn't call it enormous, not for the high-end devices this is intended for. The Denver 64-bit CPU core area seems to be a little larger than twice the area of the CPU cores in the quad-32-bit K1. Given the larger caches (4x instruction L1, 2x data L1), and the move to 64 bits from 32 (where everything -- registers, datapaths, etc) are literally twice as wide, it makes perfect sense for them to be precisely as large as the 64-bit Denver cores are. In fact, choosing their software code optimizer instead of typical OoO logic has stopped them becoming significantly larger and more complex.

Besides, while the CPU cores are twice as large, they still account for a fairly paltry percentage of the whole chip. The GPU portion is several times larger than all CPU cores combined in any version.

More likely is that they simply want the two versions to be drop-in compatible (save software, of coarse), so they needed to have the same or very similar physical footprint and external memory bus / memory controller. Since the two Denver cores can issue up to 7 instructions with two of those being load/stores each, and the 4 32-bit cores can issue only 3 instructions with one of those being a load/store, they're actually pretty evenly matched in terms of potential of-chip-memory operations. Its likely that simply adding additional Denver cores would become memory starved without a significantly beefier memory interface.

[wizard69]

"Denver" is just the name of the core(s), not the whole chip. Nvidia will likely use the same core in successive chips - I'm guessing a quad core next. Then there might be a chip with a Maxwell-based GPU.

It will be interesting to see how well these cores actually work. The good numbers NVidia seems to be posting are related to executing microcode produced by tracing apps in execution. This leads me to believe that performance will be all over the board.

Oh, and one reason that Denver is so powerful is because its instruction decoder/scheduler is 7 operations wide, compared to Cortex-A15's 3-wide (and Apple's Cyclone at 6-wide). This is a really powerful core, at least as far as ARM processors go. It will be interesting to see how much electrical power it draws though.

There is a bit of strangeness in this architecture. It will be interesting to see how code executes before it hits the microcode cache.