Hacker News Reader

Comment by ebenupton | Hacker News Reader

Nope (though note, I'm biased for several reasons). In terms of performance per unit area and performance per Watt, VideoCore IV is about as good as you're going to get. There are chips with higher performance GPUs out there, but they get there by brute force: throwing area (and therefore cost) and/or power at the problem.

View on HN

This is confirmed by the recent move of AMD towards extremely hot GPUs. You already see that in the most recent 290x, but they're talking about making hardware that is comfortable running at >100C.

I'm not sure how relevant that is with regards to performance per watt, but I do know that heat is one of the most expensive (by-)products of electricity.

That's not really relevant. Broadcom's VideoCore competes against the likes of PowerVR, ARM's Mali, and Qualcomm's Adreno. Those GPUs are all designed for mobile use, and cannot easily scale up to compete with desktop/workstation-class GPUs due to architectural differences and fab process differences.

AMD's chips aren't egregiously power-hungry, they just sell into a market where the product segments are defined by the thermal and electrical limitations of the PCIe expansion card form factor. It will always be the case that the top of the line card draws as much power as can be delivered through the slot and two extra connectors, and as much as can be dissipated through a two-slot cooling system. If AMD can safely operate their chips at a higher temperature, then their cooling system will be a bit more efficient, so their thermal envelope expands a bit and they can increase performance by a few percent.

Early indications are that NVidia's upcoming Maxwell architecture achieves significant efficiency improvements due to being designed with mobile use in mind, but it hasn't yet been demonstrated in a large enough configuration to compare against high-end desktop graphics chips.

I stand corrected.

Isn't there some theorem of thermodynamics that says your device gets more efficient as the temperature difference between the hot and cold parts of the device increases?

AFAIK the only reasons devices aren't run at arbitrarily high temperatures is that material properties change unfavorably, or you risk fires.

The efficiency thing is more about heat engines. When we're just dissipating heat, the relevant relationship is that the rate of heat transfer is proportional to the temperature difference. This is why a hot object that isn't producing heat will cool exponentially. In the case of a steady-state source of heat, the temperature difference will be inversely proportional to the thermal conductivity of the connection between hot and cold reservoirs. Heatsinks and fans are used to increase that thermal conductivity so that the temperature difference stays small enough that the absolute temperature is safe. If you can double the tolerable temperature difference, then you can cut your fan speeds by more than half, or even get by with passive convection.

Or, approaching things differently, if you're only going to keep your graphics card for two or three years, there's no reason to let the fans spin up to an audible level until the chip reaches at least 80 degrees C.

Addressing a common misconception: fan speed/noise is not directly related to how much the graphics card is heating up the room by. Whether you run the fans at full tilt or you let the chip almost melt, the heat output will depend only on the workload.

The thermodynamics you are remembering mostly refers to Carnot engines which convert temperature differences to work.

I just switched my GPU to an nVidia GTX 750 for precisely this reason. I have been near exclusively using Radeons for ten years, but the last few generations just run too hot and loud.

Waiting for the VideoCore V :) I noticed that the Broadcom LTE offerings such as the M320 do not use VideoCore and instead use PowerVR 5XT. Is it because they were originally Renesas designs? Hoping to see more VideoCore action in the future :)

Two years ago (or whenever the BCM2835 itself came out on the market), sure, but surely things have improved since? We've at least jumped a technology node, and hasn't nVidia improved?

Less than you'd think. Other vendors have improved somewhat, but from often from pretty poor starting points. Jumping a process node is a "brute force" approach: we can all do that. I'm more interested in apples-to-apples comparisons on a given node.

Had you worked on it directly, or were working on other parts of the chip while at Broadcom? I've always been hoping to see some hard benchmarks. I had worked on a video codec IP for a competing (but defunct) chip, hence my curiosity :)

I'm curious about the technological history of the VideoCore. The wikipedia page [1] and Broadcom marketing page [2] don't give a lot of information to distinguish it from competing cores (such as ARM's own Mali or nVidia's Tegra line). In fact, the information I see makes it seems pretty run-of-the-mill tech of 5 years ago.

I guess I could RTFM now that it's available :D

Edit/afterthought: none of this really matters actually, as there is no other chip on the market with the BCM2835's capabilities and that is more open-sourced.

[1] http://en.wikipedia.org/wiki/VideoCore [2] http://www.broadcom.com/products/technology/mobmm_videocore....

I did work on it directly. James (now working solely for Pi as HW director), Gordon (now working solely for Pi as SW director) and I (still working for Broadcom, and for Pi) were members of the VideoCore IV design team. James and I were responsible for the QPU (quad processor unit) design and implementation in the V3D block.

There's no one thing about V3D which makes it superior to other cores. Just lots of attention to detail at the design, RTL implementation and layout stages.

Reply on news.ycombinator.com