After issuing a statement talking about the differences between the GTX 980 and GTX 970 over the weekend, Nvidia (NASDAQ:NVDA) is now clarifying where the problem originated in terms of the VRAM allocation issues faced by several owners of the GTX 970.
Error in communication
The issue arose when Nvidia’s technical marketing team — which sent out reviewer’s guides to members of the press — was unaware of the changes in the Maxwell architecture, mainly the ability to individually disable ROPs. This meant that the actual ROP count, which was 56, was miscommunicated as 64. The L2 cache was also stated as being 2MB, when in fact it was 1.75MB.
The difference in ROP count is not due to the fact that they are unavailable, but because they are not used due to limitations with the crossbar. The crossbar – which essentially routes data from the processing clusters to the memory controller – utilizes the L2 cache. The GTX 980 features eight memory controllers of 32-bit bus width each, with each controller allocated a certain segment of the cache for access to the crossbar. By limiting the L2 cache to 1.75MB, Nvidia had to disable access to the cache for one controller, instead routing it to a neighboring memory controller’s cache, making it slower for that controller to access the crossbar.
To then facilitate usage of the entire 4GB VRAM, Nvidia sectioned the memory into two parts, with 3.5GB comprised of memory controllers that had direct access to the L2 cache, and a 512MB segment that had to interface with another memory controller to access the cache. The way Nvidia segmented the VRAM meant that a program could only access one section of the VRAM at a time; meaning that a game could either access the 3.5GB block or the 512MB block, but not both simultaneously. In a scenario when both segments are being read simultaneously, there would be a noticeable decline in performance. This is what led to the abnormal results users noticed in situations where the entire 4GB memory was being accessed.
This leads to a correction in the bandwidth on offer as well, as the GTX 970 cannot deliver 224GB/s due to the slower 512MB memory segment. At best, the 3.5GB section clocked at 7.00 GHz can attain a bandwidth of 196 GB/s.
Nvidia failed to catch the error during the review of the document, which was then disbursed to the press. In the ensuing months, the manufacturer did not notice any issues, as the engineering teams believed that the technical marketing team knew about the changes introduced in the Maxwell architecture. It was not until users started highlighting the VRAM allocation issues that the technical marketing team went back to the reviewer’s guide to check the specifications of the card.
Doing the right thing
While Nvidia was quick to clarify the issue once it was discovered, several users are demanding refunds for what they claim to be is “false advertising.” At this stage, it looks like there is no mention as to how Nvidia will handle the returns, but the vendor is claiming that it will assist users in getting refunds from retailers in situations where they’ve crossed the standard return time window.
For its part, Nvidia has mentioned that it will be rolling out a driver update that will “tune what’s allocated where in memory to further improve performance.” That doesn’t mean that the card itself will not get any faster, seeing as how the memory segmentation is done at the hardware level, but at least Nvidia will be looking to clarify that in a better way to users. The vendor is still stating that the GTX 970 is the best value for money card in the mid-tier segment, and while there’s hardly anyone who would question that statement, it is entirely possible several users would be reticent in buying the card following this debacle.