An interview with ATI - Dave Baumann - Interview with ATI's Dave Baumann - Page 1

Interview with ATI's Dave Baumann -  Page 1

AMD recently launched a new high-end graphics card. At the time of compiling this Q&A session, the card was not yet released, hence some of the questions are a little in the past tense.

The ATI Radeon™ HD 4890 (RV790) and is based on technology introduced in the ATI Radeon™ HD 4850 and ATI Radeon™ HD 4870 graphics cards (RV770). We figured it be a good time to sit down with a product manager of both that RV770 and RV790 chipset.

Our Q&A will be conducted with Dave Baumann. A name that might sound familiar to some of you. Dave Baumann was, at the time, the chief editor for Beyond3D. Back in 2006 he joined ATI (now part of AMD) as technical marketing manager.

Dave now is Product manager and both the RV770 and RV790 fell under his responsibility.

Hello Dave, thanks for the time to answer some of our questions in this Q&A.

Dave, recently RV790 aka the ATI Radeon™ HD 4890 was launched. Judging from the looks, feel and specifications the RV790 ASIC seems very similar to the RV770, with perhaps merely an increased clock frequency and small voltage tweak. If not, can you tell us a little about any architectural changes and improvements made ?

Architecturally speaking ATI Radeon™ HD 4890 maintains the exact same design as the rest of the ATI Radeon™ HD 4800 Series of products - functionally they are identical. This is one of the reasons why we decided to continue this ASIC as an HD 4800 series product, rather than move to, say, HD 4900 branding.

The chip that powers the ATI Radeon HD 4890 is, however, a new one, with changes to the physical design to facilitate higher clock speeds. This is why the chip has the new designation (RV790).

When we look at PCB design, any significant changes there ?

There are a few changes, largely to facilitate higher power and higher clock speeds. In conjunction with the work done on the ASIC we've also changed the board design to facilitate better idle power on ATI Radeon HD 4890. In doing so we've managed to pull the idle/desktop power load down by about 30W relative to the ATI Radeon HD 4870.

Our work on the ATI Radeon HD 4870 gave us an in-depth understanding of GDDR5 memory. With ATI Radeon HD 4890 we've been able to use that knowledge to make some tweaks to the memory interface. This has allowed for more stable, and consequently faster, memory operations in the newer model.

Looking beyond the PCB, there are also changes to the reference heat sink. Although the shroud design is similar, underneath we moved to a triple heat pipe design with an additional fourth heat pipe directly over the regulators.

All of this is obviously to cope with the higher power loads but will also help with overclocking.*

Additionally, we listened to the feedback from end users and the technical press community with regards to the idle temperatures of the ATI Radeon HD 4850 and ATI Radeon HD 4870. Although the boards are perfectly comfortable running at the temperatures reported by various reviewers, forum posts suggested that end users preferred lower overall temperatures. To this end, we tuned the ATI Radeon HD 4890 to run the ASIC cooler at idle. Generally, this results in ASIC temps in the region of 10-20C lower when compared to the previous models. Qualitatively, when running at full load, the ATI Radeon HD 4890 boards do feel cooler to the touch relative to the ATI Radeon HD 4870.

* AMD's product warranty does not cover damages caused by overclocking, even when overclocking is enabled via AMD software.

Here's the point where I like to step back in time a little. When we go back to the R600/RV670 (ATI Radeon™ HD 2900 XT) series GPU design, ATI made a bold move using a 512-bit memory controller. Next to using a LOT of pins and wires the decision was made to step back to a 256-bit memory bus, and pursued GDDR5 memory on the higher spec'd products. This brought ATI back on top for sure and when introduces it certainly delivered a "can of whoopass" to the competition.

Looking back at the lovely yet zealous designed 512-bit ring bus memory controller can you explain a bit why that was e.g., moving back to 256-bit memory and leaving that 512-bit memory bus for what it is ?

Bus interface width is just a means to a particular bandwidth target. When the ATI Radeon™ HD 2900 was designed it depended on a combination the best available memory - at the time GDDR3 - and memory bus to reach the bandwidth goal.

ASICs have what's known as a pad ring - this is the area around the edge of the chip where the physical interconnections with the chip are routed out to the package. Because the pad ring is difficult to shrink, and contains many analogue parts of the chip which don't always get the benefits of a process shrink, this plays a very big part in the size (and hence cost) of the chip. This means that more connections equate to a larger pad ring.

If you are designing a very big chip in the first place then having a larger bus interface can come a little cheaper, because it's not adding too much die space. Conversely, a large pad ring typically dictates a large overall chip size.

During the design process of the R600, because of the size of the chip in the first place, the 512-bit bus was relatively cheap. One of the important design criteria for the RV770 was obviously a laser focus on particular price targets (aka the Sweet Spot strategy). The pad ring required to support a 512-bit interface would clearly make the chip far too big and costly, hence the 256-bit interface. Furnishing the RV770 with a similar amount of bandwidth meant focusing on newer memory technology.

Here our work with JEDEC paid off, accelerating the design of GDDR5 memory technologies. By the time the ATI Radeon HD 4870 was launched we [had] an infrastructure in place that allowed us to go to market with roughly the same bandwidth as R600 had, but in a hugely cost effective manner. The prices people are seeing now for ATI Radeon HD 4870 and ATI Radeon HD 4890 can very much be attributed to a shift in the memory industry to GDDR5, which AMD helped make happen by virtue of working proactively with JEDEC.

Next page please.