Now that several outlets have released both the performance numbers for the ATI HD 5870 (see Anandtech, Tom’s Hardware and [H]ardOCP) and the official specifications for the Fermi, we can draw some conclusions regarding their expected performance.
Despite having double the Stream Processors (SPs) as compared to the HD 4870 (1600 vs. 800), the HD 5870 does not perform 2x faster. It runs neck and neck with the 4870X2 which also sported 1600 SPs, although on two separate GPUs. On the memory front, there is roughly 30% more bandwidth available to the HD 5870 compared to the HD 4870. All in all, decent improvements, and about as expected for a GPU that is essentially a 4870X2 on a single die. Yes, there are DX11 instructions and other tweaks, but the HD 5870 only offers minor performance gains in the latest games.
ATI has managed to beat NVIDIA’s top single GPU card by ~30%, depending on the game, resolution, and settings. Great news. Not so great news is the crossfire scaling. A pair of GTX 285s will run only slightly behind a pair of HD5870s in most games, and sometimes even performs better than the Crossfire setup. Maybe this can be fixed with drivers, and maybe we’ve simply reached the limit of the level of GPU performance that current CPU technology can support. Time will tell with the release of new drivers and as new CPUs become available.
The problem presented by this poor scaling lies in wait for the dual GPU HD 5970. If the new monster card scales as poorly as a pair of HD 5870s, there is a very real chance it will not outperform the single GPU Fermi offering. With lower core and memory clocks than the HD 5870, it seems unlikely that an HD 5970 would run faster than a pair of GTX 285s. Perhaps ATI has a trick up their sleeve to make the HD 5970 scale better than a pair of HD 5870s in Crossfire configuration and traditional single card dual GPU offerings. Time will tell on that front as well.
We’re going to see an increase in SP count from 240 in the current GTX 285 to 512 in the top Fermi offering, a ~113% increase. The switch from the GTX 285’s 512-bit memory bus paired with GDDR3 to a 384 bit bus width paired with GDDR5 should result in a ~50% gain in memory bandwidth, setting aside any clock changes. Those numbers add up to a card that should handily outperform a pair of GTX 285s in SLI, and by extension, outperform the HD 5970. The generation gap between Fermi and RV870 has the potential to be as large as the RV770 to RV870.
There are a few implications for the consumer who just wants a fast gaming card:
Price to performance ratio is always a primary concern to gamers with a limited budget to spend on graphics hardware. HD 5870 runs on a 334mm2 die fabricated on a 40nm TSMC process. We can assume that HD 5970 will include two of these dies, for a total area of 668mm2. Current estimates for Fermi’s die size range from 460mm2 to 530mm2, also on TSMC’s 40nm process, with the truth probably lying somewhere in between. Ignoring yield differences between the 334mm2 dies and NVIDIA’s slightly larger ones, this means the Fermi single GPU product should cost less to produce than the two GPUs for each HD 5970. A single GPU PCB is also cheaper to produce. Ultimately, this might not translate to lower prices for the consumer, but it could.
NVIDIA has invested a lot of time and die space on features that will have little impact on gaming performance, and those extra non-gaming features must be paid for, even if someone just wants to max out Crysis. This could result in a $600 video card that just sits on shelves next to the $600 HD 5970, while the 5870s fly into gamers hands at $300-$350. They should settle down, too, once supply can meet demand. On the other hand, it might result in a massively powerful gaming card with a lower price due to a hefty development subsidy courtesy of the HPC customers. What remains to be seen is if anyone actually needs the level of power Fermi should be bringing to the table. It will be a sure bet for HPC and folding type applications, but for gamers, it might be overkill.
ATI is focusing more squarely on the gaming market, and should be able to keep their prices down due to a lesser focus on GPGPU-type functionality and smaller die size. The problem now is yields. Can they be improved to the point that the HD 5870 becomes widely available again, before the Fermi launch blunts demand? Everyday there are no cards available is another day closer to the Fermi launch, and gamers willing to spend $350-$400 on a video card may also be more willing to wait and see what the competition has to offer.
Discuss this article on Overclockers Forums.