AMD Radeon HD 7950 Launch Review



Product: AMD Radeon HD 7950 Video Card
Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: January 31st, 2012

Tahiti Pro

There are four fundamental differences between the Tahiti Pro and XT specifications. The full complement of ALU's is reduced from 2048 to 1792; four less Compute Units attached to the dual asynchronous compute engines. Each compute unit is not only comprised of four SIMD-16 vector units, but also a scalar unit; both vector and scalar registers, a branch and message unit, scheduler, local data share, 16KB L1 read/write cache, and texturing components which are comprised of four texture filler units and sixteen fetch load/store units. As CU's are disabled the performance is affected, reducing the texturing and pixel fillrate. With 8 Render Back Ends (RBE) serving 32 Compute Units in Tahiti XT, the next logical grouping is 28 CU's/7 RBE's.

GCN Compute Unit

Other than dropping the CU count by 4, and thus the texture unit count by 16 (one RBE), the engine clock is dropped by ~15% from 925MHz to 800MHz. On the memory side of things the full 384-bit bus with 3GB of VRAM is present, with the GDDR5 base clock rate reduced by 10% from 1375MHz to 1250MHz, resulting in a 5.5Gbps Quad Data Rate (QDR). These changes result in pixel fillrate dropping by ~15%, texture fillrate dropping by ~32%, and memory bandwidth by 10%.

Tahiti XT vs. Pro

The maximum board power of the AMD Radeon HD 7950 is given as 200W, 50W lower than the big brother HD 7970/Tahiti XT. This is enforced by AMD's PowerTune technology which adjust engine clock speed dynamically on the fly when the workload is calculated to exceed that limit for thermally significant period of time. In practice, AMD's PowerTune limit is set high enough that the only applications that will trigger PowerTune are ones that aren't much fun to play - FurMark, OCCT; the 'Power Virus card killers'. Naturally, this option is adjustable in AMD's control center software, Vision Engine Control Center (VECC, which is branded the Catalyst Control Center on non-AMD processor PC's). As part of the warranty-void-if-you-use-it-but-it's-an-advertised-feature AMD overdrive, you can adjust engine clock, memory clock, and PowerTune limit. The limits on PowerTune are generous, +- 20%, or 180W to 220W. The card features two PCI-Express 6-pin power inputs, rated at 75W each, and is specified to draw 150W from the PCI-Express slot for a total of 225W. AMD got the balance of potential draw and specification rating spot on. The reference AMD design for the 7950 reduces the number of GPU power phases by 1 to a 4 phase configuration (+1 memory phase). Interestingly, the reference 7900 series PCB features an spot for an additional power phase, for a total of 6 - and the PCIe power input connectors can be both 8-pin 150W inputs. An extra phase for 6GB memory on FirePro cards? Design headroom for a >300W Tahiti monster clocked SKU?

PowerTune

Zero Core Power is the desktop implementation of Bus Alive Chip Off, where the card is powered down to as little active as possible - the GPU cores are off, the fan is off, only the bare minimum necessary for a fast resume is kept working for an incredible ~3W idle when the display is blank. Zero Core Power works on all cards present in the system, not just additional GPU's to the primary. When the display is blank, all GPU's can enter ZCP state, as well as if an application fires up and only uses a single GPU, additional GPU's can be in ZCP for the lowest possible power consumption.

ZeroCore Power

As just noted, the AMD Radeon HD 7900 series are PCI-Express 3.0 cards, and backwards compatible with 2.0/1.1 equipped mainboards. While PCI-Express 3.0 offers double the bandwidth of PCI-E 2.0, only early adopters of Intel's X79/Sandy Bridge-E platform have it yet, and it's likely the 'good enough' point is going to remain x8 PCIe 2.0 lanes per card despite AMD's claims that their dual DMA (direct memory access) engines can saturate the 16GB/s PCIe 3 x16 bus.