vacancies advertise contact news tip The Vault
GLOBAL CONTEST: Win a Samsung Galaxy S5 with Tech21! [x]
facebook rss twitter

Review: NVIDIA's GeForce Go 6600 and GeForce Go 6200 Launch

by Ryszard Sommefeldt on 8 February 2005, 00:00


Quick Link:

Add to My Vault: x

NVIDIA's GeForce Go 6600 and GeForce Go 6200 Launch

And GeForce Go 6600, but NVIDIA didn't supply a logo for that. Skills.

Check this out. "The TurboCache technology and NV44 definitely need to make their way into notebook or IGP solutions provided by NVIDIA, for integrated and portable (not DTR) graphics that start to make good performance sense, rather than passable performance sense."

Go Rys! Go Rys! It's your birthday!

Ahem. The reason for my excitement is because my assertion rings true today. Yes, anyone with a couple of functioning brain cells and an eye on how the 3D graphics market works could have predicted it, but this is my article and it was actually my birthday recently. Along with NV43M, a mobile version of NV43, NV44M featuring TurboCache makes its debut. TurboCache is particularly pertinent to portable systems, so I'll focus on that.

TurboCache and HyperMemory, schemes where the GPU's rendering pipelines can write data back out to system memory after processing it, as well as reading from it, are somewhat suited to being used in a laptop. AGP lets you read from system memory, mapped by the AGP host and accessible by the GPU. PCI Express lets you do the same in the other direction, writing back to the PCI Express graphics (PEG) host and therefore back into system memory. The driver, GPU (with its integrated MMU) and PEG host all conspire to let it happen.

The end result is you need less memory local to the GPU. The TurboCache 6200s I had a look at had 16MB and 32MB local, respectively. You need at least 16MB because the final buffer swap to get a picture on the display device is done entirely from local memory. For a large resolution, like those on recent laptop screens, you may even need a 32MB minimum.

Secondly, and it's a minor point in reality, the minimum bus width you need to satisfy to the GPU is 32-bit which is the bus width of a single GDDR3 DRAM. So at least one DRAM device that's preferably 32MB in size on a laptop.

You'll remember NV44 was the GPU used for the TurboCache 6200 boards. Some tweaks to NV44 and a bit of software on top give you NV44M.


M denotes mobile, as you expect. The tweaks are almost entirely related to power management on the GPU, or at least that's what the interesting differences are. There's likely to be no actual difference betwee the NV44 and NV44M ASICs. That is to say that NV44 will support the GPU features that NV44M does, just the driver and card BIOS will collaborate to disable them.

Much like you'd enable Cool 'n' Quiet, PowerNow!, SpeedStep or other dynamic clock changing schemes for your processor, NV44M supports those too, using a combination of the card's on-die thermal diode, the GPU's BIOS and the driver installed under the operating system. Temperature reported by the diode and the load on the GPU monitored by the driver will cause the driver to call GPU BIOS functions that change frequency and voltage on the fly.

During idle periods, frequency and voltage are reduced to slow the GPU down and have its heat output and power draw drop, saving on the battery and producing less heat. Load it up and the driver and BIOS will scale it back up to full power. If on the battery, the scaling will be less agressive to suit with it all controlled by the software interface in the driver.

The collective name for the technology is PowerMizer, which I've touched on briefly in the past. The ability for the GPU to dynamically adjust its clock and voltage, aften many times per second, is the key ingredient for successful operation of what's essentially a discrete desktop GPU in the first instance, as a mobile part. There are other factors involved, like optimising out GPU hotspots using new silicon spins, but when that's all been completed, clock and voltage adjust is the mechanism used.

NVIDIA quote between 22W and 40W for Go 6800 (NV41M), 12W to 20W for Go 6600 (NV43M) and less than 10W at all times for Go 6200 (NV44M), all depending on their clock speed and voltage.

A lot of the same technology, bar TurboCache, applies to NV43M too.

New Products

GeForce Go 6200 was actually launched last month to muted reception. It's NV44M featuring a single fragment quad (group of four pixel fragment pipes, each with a texture sampler) at 300 or 400MHz. 2 vertex units supply the fragment units with rasterised fragment data. On-package or on-module memory is likely to be 32MB at all times using a single DRAM device, with up to 256MB managed by the GPU's MMU with 32MB of local memory, and 128MB managed with 16MB of local memory. Frequency for the memory is 300MHz (600DDR) or 350MHz (700DDR) depending on the base frequency of the GPU.

GeForce Go 6600 is NV43M (two quad) with two vertex units and up to 350MHz (700DDR) memory, leaving the mobile versions of 6200 and 6600 with scant vertex power compared to the desktop variants. NV43M isn't a TurboCache product so has full local memory and higher performance.


MXM rears its head again with NV43M and NV44M, NVIDIA offering the GPU as a PCI Express MXM module design, or indeed discrete if the vendor prefers. MXM is NVIDIA's mobile form factor, used to create a standard physical package and cooling unit for mobile devices that use it. A muted reception for MXM initially has turned into a small but steady trickle of MXM design wins, with at least a few of NVIDIA's launch partners for NV44M/Go 6200 and NV43M/Go 6600 using MXM in their designs.

Target systems

NV44M and NV43M are targetted as the graphics unit for mid-level performance laptops in the thin and light category. Think something like a 1.7GHz Dothan Pentium-M weighing about 2Kg/4lbs, with 512MB of memory and a widescreen LCD, in a thin chassis. Ever seen anything like a Dell Latitude D600? It'll make its way into notebooks that around that size.


NVIDIA's presentation documents for the new products talk about performance per watt, and while they're happy with the TDP of these new parts, it's a bad metric to convey to you. instead, if you have a look at the benchmarks for the desktop TurboCache 6200s, which start here, you'll have a decent idea of Go 6200's performance. For NV43M, find a review of a PCI Express 6600 non-GT with 128MB and you have a ballpark area for the very fastest Go 6600s, guessing a bit. NVIDIA assure us some samples are on the way to Chez Rys, so I'll confirm or crap on my guesswork in due course.

They are GeForce 6-series GPUs

Finally, remember that it's a GeForce 6-series processor. That means bilinear texture samplers, floating point fragment units, HDR rendering, decoupled ROPs, the video processor and all the other good stuff that makes up NVIDIA's recent graphics accelerator products. I'll go over all that in more detail should a proper sample show up. Check here for the basics and remember that it only has two vertex units instead of three.

Sum it up

They represent relatively speedy mid-range graphics with a power-optimised bent, in skinny notebooks, for gaming and good video on the move. NVIDIA have brought the entire stable of discrete products to the portable market, with them all shipping as I type this, in a bunch of designs from the likes of Clevo, Uniwill, Wistron and others.

ATI are doing much the same with PCI Express, HyperMemory and optimised mobile versions of desktop ASICs. 3D gets ever more powerful on the move, and it's good to see NVIDIA push their latest architecture onto devices you can use to play Half Life 2 on the bus. More as and when I get my grubby paws on a sample, and my apologies for the press regurgitation.