SAN JOSE, Calif. – Nvidia Corp. rolled out a new generation of desktop and notebook graphics chips, fueling the next round of neck-and-neck competition with archrival Advanced Micro Devices. Nvidia claims its Kepler chips, its first to use a 28-nm TSMC process technology, have twice the performance per watt of its previous 40-nm Fermi graphics
A Kepler-based graphics card, the GTX 680, officially goes on sale today at retail outlets for $499. It sports 1,536 Nvidia proprietary CUDA rendering cores and 192 control logic cores on a GHz clock. That’s up from 512 and 32 cores in the Fermi parts.
Like Intel and others before it, Nvidia has learned more cores and a slower clock is the road to better power efficiency. Fermi ran its rendering cores at 1.544 GHz and the rest of the chip at a cooler 772 MHz. With Kepler, Nvidia takes a different tack, running the whole chip at 1.006 GHz but using three times as many cores.
The GTX 680 card rides a PCI Express Gen 3 bus and consumes 195 W, down from about 250W for Fermi cards. That means PCs require two conventional six-pin power supplies to support the card, rather than the one six- and one eight-pin supply Fermi cards needed.
Separately the Kepler chip is the basis for Nvidia’s new 600M family of graphics chips for notebook computers, already shipping to OEMs. The low-end versions should be suitable for $599 retail notebooks, an Nvidia spokesman said.
Acer is already shipping one notebook with a 600M chip. Nine other top notebook makers will follow suit when they ship laptops using Intel’s new Ivy Bridge processors
, expected to be formally released in late April.
According to Patrick Moorhead, principal analyst of Moor Insights & Streategy, the 600M has a unique capability to "dynamically shift power" from the CPU to the GPU. "OEMs use one heat spreader for the CPU and GPU so they know exactly how far they can go between combined CPU and GPU--this has never been done before," he said.
Independent testers have yet to publish benchmarks, but market watchers expect the Kepler parts will be fairly evenly matched with the 7970
, AMD’s 28 nm family that started shipping in volume in February.
“Based on the fact both companies’ parts are using GDDR5 memories at high clock rates, I expect they are very competitive,” said Dean McCarron, principal of Mercury Research (Bee Cave, Ariz.) “Both companies will be ramping products through the summer, so we will have some nice battles here,” he said.
“On paper Nvidia’s specs look great, but then so do AMD’s for the 7970,” said Jon Peddie of Jon Peddie Research (Tiburon, Calif.). “We won’t really know until we run tests in the same machine with the same programs on both boards,” tests that are not expected to be done for a few weeks, he said.
Nvidia commands a market share edge over AMD in the slower growing desktop discrete graphics segment at about 63 versus 36 percent, according to Peddie. But AMD still commands a narrowing lead in the faster growth notebook discrete segment at about 52 versus 48 percent.
A year ago, AMD had a larger lead at 62 to 38 percent, McCarron said. A total of 75 million discrete notebook graphics chips shipped in 2011, projected to grow to nearly 100 million in 2013, McCarron said.
Nvidia is getting out of the business of chip sets with integrated graphics as Intel and AMD move to put graphics cores on the same package as their latest processors. Due to this trend, Intel is the overall leader in graphics with a share of the $500 million combined integrated and discrete market at 59.2 percent. AMD stands at 24.7 and Nvidia at 15.7 percent, McCarron said.
AMD and Nvidia have been roughly matched in graphics performance for about three years. The current chips are not expected to shift the balance significantly.
Although AMD had a few weeks lead in shipping its 28-nm generation, both companies are comfortably ahead of the refresh cycle PC makers are in with the upcoming Intel Ivy Bridge processors, McCarron said.
The new chips are too power hungry for smartphones and tablets. Thus Nvidia’s graphics cores for its Tegra products for mobile systems have a separate lineage.