Were VR graphics ahead of their time?
A headset owes the eye a new picture on every turn of the head, finished before the lag shows. Flight-simulator practice had put the noticeable lag at about 50 milliseconds. Frederick Brooks, looking at the systems of 1994, found the usual lag at 250 to 500 milliseconds. At a typical turn of about 50 degrees a second, that delay makes the scene swim. The same brief frame had to contain the world as well as arrive on time. The fastest commercial image generator he cites for that year managed about 600,000 polygons a second, about 30,000 polygons inside one twentieth of a second. A kitchen brought to quarter-inch accuracy took his group several person-years, and the spice rack and the stove knobs remained a texture on a block. From a model to a pixel ran on a model that had already given up most of its detail. In that year the picture was ahead of the machine on both counts, the lag and the polygons.
The polygon count caught up first. By 1999 one pipe of an SGI Reality Monster could draw up to 180,000 polygons in that same twentieth of a second. Worlds past 500,000 polygons still needed extra measures to stay in real time, so the heaviest models had a price, but Brooks judged the limit to be money more than the technology. Cards paid for by games were enough to build a workable system from parts on the open market. For many applications, drawing the picture had ceased to dominate the cost. Resolution and lag stayed behind that new rate. A 1280 by 1024 projector filling a 10-foot cave resolved about 4 minutes of arc, where acuity of 20/20 is about 1 minute, and the industry median for a head-mounted display was a full-overlap field of about 45 degrees. Brooks still put the lag first. On a headset, 150 to 500 milliseconds of delay made the scene swim after the polygons had multiplied. The machine could draw more of the world. It still could not match the eye, and it still could not keep the picture with the head.
A card that could draw the frame left the room in another price band. Anthony Steed, writing in 2017, puts the high-end installations of that earlier period at hundreds of thousands of pounds. The sum made sense where the training was dangerous, or where a design review would save a cost the firm could already name. Those are the two uses Brooks had treated as established, design review and vehicle training, and they were the uses that could carry a room of that cost. Firms selling complete immersive systems in the 1990s, Virtuality and Division among them, faded with the excitement. The work that stayed was the engineering that could still justify the room.
The lag came within reach of an ordinary computer later, and the picture left the special room with it. Steed places that point in the two years before his article, when a high-end desktop could drive real-time graphics fast enough for a low-latency immersive picture. At the start of 2017 Samsung had shipped more than five million GearVR units, a phone held in a visor, and the phone was the image generator. A survey dated 17 June 2026 estimates that more than 171 million people use VR, about 70 percent of VR consumers using the headset for games. Inside that use, the dedicated visor is the part that shrank. The survey cites IDC: Meta's Quest shipments fell 42.3 percent on the year, and in 2025 Meta's lead in the combined headset and glasses market was coming from smart glasses. Those figures describe machines long past a frame of 30,000 polygons. In 1994 the processor could not finish the picture. After the polygon rate was caught, the headset rose or shrank as something worn on the face.