Nvidia Tesla is reborn as ATI Streams Fire

BOTH NEW GPUS from both Nvidia and ATI - sorry AMD, Daamit, claim teraflops-class performance: OK, for the generic US$ 500+ GTX280, you'll need to clock it at 650MHz or above GPU and corresponding increased shader clock, around 1.5GHz, to get that performance.

As for the HD4850, it is advertised as the first graphics card with 1tflops single precision theoretical peak performance for less than US$ 200 - its faster cousin, HD4870, is rated at 1.2 tflops for less than US$ 300.

The "professional high performance computing" versions, the Tesla 10-series for Nvidia and Firestream for ATI, will differ by larger on-card memory to fit in the large HPC datasets, possibly slightly higher FP performance, and - of course - lack of graphics outputs. And higher price, natch.

Today, let's look at how the two compare on raw hardware capabilities and scaling. We'll be covering more on the increased application usage of GPGPUs despite their limitations, the programming pros and cons, as well the competition, in subsequent GPGPU-watch stories. Yes, ultimately, it may be more important who will run the new Photoshop filters or GPGPU antivirus first.

So, the landmark 1 Tflop milestone has been passed: how much of it do you really get at the end? A big chunk is determined, of course, by optimised libraries and the programming environment - what about the hardware limitations?

First, GPUs don't have nearly as much cache or other memory on chip as typical CPUs. The sustainable speed depends more on the memory bandwidth and the memory controller efficiency for long bursts, where the initial latency can be hidden - clever pre-fetching techniques can help, too.

Let's even say that either of the GPUs mentioned here could sustain 100 GByte/s read or write speed to their local memory when using those long burst transfers. To sustain, say, half a teraflop and just sending out one 32-bit single precision FP number per flop to the memory - not even counting the reads for the operands - would take two terabytes/s sustained memory speed if trying to sustain that speed across data in local memory. Doable? Maybe, with those proposed Rambus next-gen memories, but not with the current stuff. Working on loops within GPU chips internal memory would alleviate this.

Secondly, talk about the double precision FP: after all, while you can use single-precision in some tasks, or parametrisation for the bigger jobs, IEEE standard DP FP is still the mainstream of most scientific and engineering codes.

Now, Nvidia GTX280 (and equivalent Tesla card) offers DP throughput at 1/8 of the SP peak, i.e. just over 125 GFLOPs for the new Tesla 10 series or the GTX280 OC. The ATI 4800 series offers DP at 1/4 the SP peak performance, i.e. 300 GFLOPs on the HD4870. In either case, if solely dependent on the card memory throughput, it wouldn't be easy to get anywhere near that peak. But of course, hundreds of those stream processors in GPUs can hold some data in their local registers and shared memory, to be processed at full speed.

At the recent Tesla briefing, Nvidia suggested that their four-card, 16GB and four single precision tflops slim rackmount box should be able to get somewhere around 350 Linpack Rmax GFLOPs (measurable maximum) in double precision. For a box costing somewhere around US$ 9,000, that is a great number - as long as your app can get anywhere near that number. It is quadruple the speed of a dual-CPU overclocked 4GHz Skulltrail in that same Linpack DP - at about the same cost. The problem? Unless your app is CUDA coded for the GPU support, there will be far more software that can make use of 80 Gflops on that Skulltrail than 350 GFflops on the custom Nvidia box.

The GTX280 chips' wide 512-bit memory bus, considered a burden among gaming GPUs as it complicated both the die and board design, is a huge plus in technical computing use. Simply, for a given memory technology, when you need to max out the capacity, you'll at anytime have double the possible capacity - and bandwidth - with green goblin's cards. The new Tesla cards have 4GB GDDR3 RAM per card - even though more conservatively clocked due to the dual-rank mounting and higher loads, it still allows packing that much more data into the fast local memory rather than losing 10x performance when going over PCI-E to the system memory.

The ATI side compensates for the narrower 256-bit bus with faster GDDR5 memory, however, not only it cuts the maximum capacity by half, but also requires - still very rare - higher capacity GDDR5 memories if needing to go to 2GB or more memory.

So, from the raw hardware point of view, ATI offers higher peak SP and much higher peak DP flops, but its narrower memory bus could turn it into a little bit of a capacity expansion 'flop' for those computing apps in need of more on-board memory. How do the two, Tesla and Firestream compare internally at the chip level? What about the software? You'll have to wait to read all about that soon. µ