RV740 – first to 40nm
RV740 features 128 five-way superscalar shader units or 640 stream processors, depending on which metric you prefer to use – we prefer to use the former, because not all of the stream processors are equal.
Regardless of that though, it’s double the amount of shader units in the Radeon HD 4670’s RV730 GPU and 20 percent fewer units than there are in the excellent RV770 chip, which featured 160 shader units or 800 stream processors.
The Radeon HD 4770 makes up for its shader unit deficit with higher clock speeds – the core is clocked at 750MHz, compared to 625MHz and 575MHz on the 4850 and 4830 respectively.
That’s how it delivers a similar theoretical throughput to the Radeon HD 4850 and quite a bit more performance than the 4830.
Click to enlarge
As the texture units were reorganised in AMD’s RV7x0 series architecture and are now allocated on a per-SIMD basis, the number of texture units has been reduced in proportion to the number of active SIMD cores. RV730’s texture units were slightly chunkier than RV770’s, with two texture units per SIMD core, but AMD has gone back to the original ratio with RV740 – there’s one texture unit per SIMD core here, meaning a total of eight units or 32 individual texture samplers.
Like the compute performance, the Radeon HD 4770's texture sample rate follows closely behind the Radeon HD 4850's 25 GTexels/sec fillrate – it's just four percent lower at 24.0 GTexels/sec, despite having 20 percent fewer texture units. An area where the new RV740 chip excels over even the Radeon HD 4850 is in pixel fillrate.
Click to enlarge
RV740 features 16 ROPs like RV770 and they’re clocked 20 percent higher on the Radeon HD 4770. The result is that the pixel fillrate has increased from 10 GPixels/sec on the Radeon HD 4850 to 12 GPixels/sec on the 4770. With that said, the 4850 wasn’t exactly fillrate limited, so we’re unsure exactly how much of a performance difference this might make.
It will be stopped short by the 128-bit memory interface though, which is half of bus width in RV770 – AMD has adopted four 32-bit memory channels in RV740 instead of the four 64-bit channels in RV770. When questioned, AMD would not disclose details on any bandwidth efficiency optimisations it might have made to the ROPs at the same time – it seems pretty obvious that some will have been made to better utilise the available memory bus.
AMD has made up for some of the bus width deficit by coupling RV740 with 512MB of GDDR5 memory, which hums along at 3,200MHz (effective) and delivers 51.2GB/sec of memory bandwidth. That’s still some way off the 64.0GB/sec offered by a Radeon HD 4850 fitted with 2,000MHz GDDR3, but it’s not as far behind as it might seem at a first glance.