Post-Computex high end GPU analysis – More memory for Radeon, quad-GPU attempts for Kepler?
While the volumes aren't high, most vendors had enough to show at least a sample of GTX680 – and occasionally a GTX690 – on their booths, while AMD gunned for higher clock revisions and more memory.
So, you thought all was over this summer with Nvidia GTX680 gaming benchmark leads over AMD? Not exactly: first, even on the meetings behind closed doors, accusations flew – likely with some reason – of Nvidia's "little more than paper launch" with being extremely hard to get cards and limited volumes / yields / name it. Of course, if the GK104 is full of such problems, how would then the upcoming nearty double-sized GK110 die fare in the availability department?
Anyway, all major vendors had both top end AMD and Nvidia GPU boards displayed, aside from those vendor-specific GPU houses like say Galaxy and EVGA on Nvidia side, or Sapphire and Powercolor on the AMD side. Anything interesting worth highlighting?
Well, remember that the initial AMD HD7970 was very conservatively binned, with plenty of upwards potential even without the usual OC tricks? And, of course logically, the process would have been further fine tuned since then? So, the HD7970 GHz edition existence was pretty much confirmed by the vendors around, and yes its availability should be as good as – or even wider than – the GTX680. A, say, 1075 MHz default clock HD7970 would beat back the GTX 680 in even more gaming tests, not just the highest resolutions ones. Also, more importantly, it would reach an admirable 1.1 TFLOPs in double precision FP peak performance! Not far from the expected GK110 die expected half a year later, at much lower price for now.
This should be good news for GPU compute guys, who need all the double precision FP performance they can get. But they also need more memory to store the GPU compute datasets without going often across the slower PCIe bus to the system memory. So, in the absence of FirePro cards based on this chip (AMD workstation GPU division, wake up please??), Sapphire has its 6 GB HD7970 version finally shown here, running in CrossFire. With 1100 MHz GPU 'turbo button' and large memory to boot, this is an inexpensive serious GPU compute card for those willing to get into teraflop-class FP through the OpenCL programming model – yet still at gaming card prices.
Back to Nvidia, there was a not-so-secret' card and then another gem hidden inside You-Know-Who's showroom – a quad-GPU, dual PCB monster made out of four GTX680-class GK104 GPUs in one card. Aside from handling something like 600 – 800 W of consumption even if the clocks are slowed a bit, it remains interesting to see who would buy this – only the users keen on buying TWO of them, of course, for 8-way SLI.
Now, if such a card ever gets released, two big problems prop up: the thermals, of course, and the latencies: having TWO layers of PLX PCIe V3 bridge chips to accomodate four GPUs on one PCIe link adds something like 300 ns of extra latency for every single word transfer, which can impact both FPS gaming and GPU compute.
Also, if you're really into 8-way (compute), why not get four dual GPU boards with a mainboard having four native x16 PCIe V3 slots, like the Asus Z9PE-D8 WS or EVGA SR-X, and put in two Xeon E5 CPU to have 16 core feeding all those GPUs as well? After all, if willing to spend so much on 8 GPU units, may as well spend a bit on unreleased, unlocked CPUs to handle them, too…