A close look at the ATI 4850 and 4870

ATI LAUNCHED THE new R770 cards yesterday, and it will come as a shock to no one that they are called the 4870.

After the raw performance of the 4850 a week ago, here and here, you have to wonder what more they had up their sleeves And, in short, they deliver.

Visiontek 4850 and 4870 cards

A little over a week ago, we recieved four cards from the nice folk at Visiontek, two 4850s and two 4870s. That lead to a week of playing around with Crossfire and high end CPUs.

These cards are stunningly fast and inexpensive, but in the end, to me, the burning question was, will DAAMIT live up to this slide.

Five silicon generations

First, let's step back and tell you a little about what we tested on. The cards are the aforementioned Visiontek graphics board, and for the older 3870s and 2900XTs, generic ATI parts were used. The CPU was the fastest thing you can buy right now for single sockets, an Intel QX9770, 3.2GHz quad core and it was in an Intel X48 based DX48BT2. The memory was Kingston DDR3/13000 CL7 HyperX, and we have not gotten close to it's limits yet. Powering the beast was the trusty OCZ Silencer 750 Quad Crossfire edition PSU, basically your average everyday gaming rig.

Two tests were used, 3DMark06 to measure wattage, the benchmark itself is limited by CPU power. We have used 3DMark06 to test the power consumed because it is consistent and many sites run the same workload.

For raw rendering power, a new test was used, in this case based on UT3. Because we were more interested in the raw rendering power than gameplay and CPU loading, we used a flythrough demo without bots.

To accomplish this, we used the HardwareOC Unreal Tournament 3 benchmark on the Deimos and Torlan maps, picked by the scientific process of, 'We felt like those two'. The results were as you would expect, remarkably consistent. No bots, few random number, and in general, things on rails means FPS never varied by more than two, and even that was an anomaly. We set it for three passes and took the average FPS.

The net result was pretty smooth scaling from the UT3 tests, and consistent power measurements from 3DMark. Please note that all this testing was done on XP, the Broken OS was not used due to inbuilt malware and user antagonistic licensing. All software was patched to the latest and greatest available on June 20th. All of the UT3 tests were done in 2560*1600 resolution with 8x AA and 16x AF on.

The raw numbers are as follows, with wattage in (duh) Watts, 3DMark listed as it's score, and UT3 in FPS.

The raw numbers

The first thing you see is how flat the 3DMark scores are, and that is why we didn't use it for a rendering power test. In retrospect, looking over the final numbers, we should have run it at 2560*1600 with 8x/16x AA/AF on to avoid this bottleneck. Then again, that would have made it harder to cross-compare power numbers between sites.

The next thing you will see is that the 2900XT is consistently faster than the 3870, both in UT3 and 3DMark, not exactly a surprise considering that it went from a 512b memory interface to a 256b in the later cards. This has an effect on higher resolutions and workloads with more textures, but the 3870 uses half the power of the 2900XT.

Lets look at power, a fairly imprecise and variable thing on the best of days. All measurements were taken with an Extech True RMS Power Analyzer, and were measured at the wall. Keep in mind that this means the numbers are bumped up by the inefficiency of the PSU, so if you want the true DC power consumed, knock 20-25 per cent off the top of these figures.

If you subtract out the power used by the single cards from the same cards in Crossfire, you end up with roughly what a single GPU uses. More or less. Sort of. It looks like this, in Watts of course.

Measured wattages of the cards

Our first thought on this is to wonder how much more efficient, on wattage per unit die area basis is the 4850 over the 3870. The reason we picked these two cards is that they are both 55nm parts, and and the memory types, GDDR3 vs GDDR4 are closer than the GDDR5 of the 4870. If you take the wattage used by the 4850, and multiply it by the ratio of the two GPUs die areas, you should get about the power used by the 3870.

The RV670 is about 190mm^2 and the R770 is around 260mm^2. A little more maths tells you that the RV670/3870 is about .73x the size of the R770/4850. If you multiply that scaling factor by the measured power, 46W idle and 134W loaded, for the 770, you should get roughly the measured numbers of the 670. (.73*46) = 33.58 vs a measured 33W and (.73*134) = 98 vs a measured 109. Not bad, pretty close, so it doesn't look like there is all that much difference between the transistors themselves, somewhere between zero and 10 per cent overall.

Now that we know the transistors used are more or less the same between the cards, that would leave any performance differences between the two cards down to architectural improvements. Doing the same math for the UT3 tests, we get a scaling of (42/55) = .76 for Deimos and (40/53) = .75 for Torlan, mighty close don't you think?

That means the GPUs, without much done to them in the way of memory or other subsystems isn't all that much faster, mm for mm. People however don't buy mm^2 of GPU, they buy cards, complete cards. Those include memory, and if you compare comparable cards, the 4870 to the 3870, the picture changes a lot. For the Diemos test, you get a scaling of (42/76) = .56 and (40/58) = .69. For the math challenged, this means the 4870 is 1.8x and 1.45x as fast as it's immediate predecessor. Not bad at all.

Running the same numbers for Crossfire, you get 1.63x and 1.6x respectively. For scaling between single cards and dual, run a similar set of numbers. You get 1.63x scaling for 4870/Deimos, 1.93x for 4870/Torlan (leading me to believe the single 4870/Torlan score is a little low for some odd reason), 1.73x for 4850/Deimos and finally 1.70x for 4850/Torlan. These are in line with the same stats on the 3870.

OK, how about that graph we told you about earlier? It shows GigaFlops/Watt doubling from the 2900 to 3800 and doubling again to the 4800 series. It also shows GigaFlops per mm^2 doubling and doubling again. From what we know, do they?

If you normalise the single card scores on Deimos to the 2900XT = 1, the 3870 = .93 and the 4850/70 is 1.69. Die areas for the four cards are 420mm^2, 190mm^2 and 260mm^2, normalised to 1, .45, and .62 respectively. The normalised power figures are 1, .49, .61 and .86 in the same order. If you do a Deimos/Power and a Deimos/Area as wild stabs in the dark at the GF/Watt and the GF/Area stats, you get a chart that looks like this.

Normalised numbers

So, did they do it? In real world performance, no. If you take GFlops to mean theoretical work done by the GPU instead of a system test, it is likely, especially on the area graph. ATI did more than double the number of shaders, and their performance is about equal between generations, so theoretical power should go up by 2.5x over the R670.

In the end, the testing showed, on the hard numbers, what we already knew, theoretical potential is hard to reach. In the real world, things like busses, HDs, memory latency and CPU power get in the way. Another thing in their favour is that a lot of the effort in these parts went into DX10/10.1. I have a sneaking suspicion, backed up by the ATI documentation, that these cards scale way better on DX10 than DX9. Since that very nice and clean API is hamstrung by the malware and DRM infested Broken OS, we can't test that.

The hard numbers, however, don't begin to show off what these cards can do, so let's turn to the more subjective. For the past two weeks or so, I have pulled the 3870s out of my old gaming box and plugged in the Visiontek 4850s and 4870s. The two systems were the above testing machine and a Phenom Black 2.3 @ 2.8 on an Asus 790 board. Both were plugged into a Dell 3007 30-inch monitor and games were all run at 2560*1600.

The subjective opinion, playing Supreme Commander, Portal, Half-Life 2, Telengard, Dawn of War, CoD4 and UT3 is that these cards are damn fast. My informal game playing involves going into a game and turning everything up to the max, then playing for a bit. If it is playable, it is playable. If it starts stuttering, we back off a bit on a few things in a semi-scientific way until the low frame rates are gone.

With the 3870s, you would get some bogging here and there, not bad, but it did happen. None of it required backing off more than one notch on the settings tabs in any of the games. In a week playing with the 48xx cards, we got none of that. Even with the most torturous in-game settings, these cards never let us down. We could probably have gone into the ATI control panel and forced the issue, but with the in game settings, well, it just powered through things.

Top to bottom, 2900XT, 3870, 4850, 4870

Another thing to think about is power. No, not the usage numbers, but the connectors. The 2900 was the first to have an eight-pin connector, and the OCZ PSU was the first we got that could service that part. It has worked admirably since, but there has been no ATI card to date other than the 2900 that needed it. The 3870 only needs one six-pin as does the 4850. The 4870 on the other hands needs two six-pins.

If you are considering upgrading, you are probably going to be limited by the PSU configuration you have. Most modern PSUs have at least a single six-pin, and two of them are becoming quite common. This means the 4850 is a solid and doable upgrade, the 4870 is a little more of a stretch for most people.

The 4850s have only been out for one week, and the prices are pretty well pegged at the $199 price point, and will be stuck there until supply catches up with demand. This won't take very long for the 4850, and the 4870 will be a few weeks behind. Even with the newness of both cards, there are up to $30 rebates on many cards already, meaning you can get a 4850 for as low $169 right now. 4870s will assuredly drop from their $299 price point soon, they just started shipping a week ago. There will be 60K+ units in the channel before the end of June, so supply won't be a problem as soon as the boats are unloaded. For reference, we hear the number of 280s and 260s combined are 20K in total for launch with a syphilitic trickle of parts following the initial shipments.

So with the Visiontek 4850 and 4870, you have a simply amazing deal. For gameplaying, nothing else comes close. The 4850 lines up with the Nvidia 9800GTX+, something that is totally unavailable for another 3 or so weeks, and costs $30 more MSRP, likely $60 more at retail. The 4870 has no competition, the closest Nvidia can get is the $400 GTX (down $100 from what they told reviewers), and that is not worth the very slim performance premium, not to mention the power cost. A heavily overclocked 260 is just a hair faster than a $100+ cheaper 4870, and two 4870s in Crossfire crush a more expensive GTX280. Once prices come down, and they will shortly, it is going to be quite unpleasant for Nvidia.

We would unhesitatingly recommend the Visiontek 4850 and 4870 cards, they are amazing pieces of kit. If we had to pick one setup, we would recommend a pair of 4850s, for the $400 or less they cost, nothing can touch the performance. To put it in perspective, you can get two 4850s, an ATI 790 mobo, and a Phenom X3 for less than a single GTX280, with enough money left over for a meal at McDonalds, supersized, for both you and your online 'girlfriend'.

In any case, both Visiontek cards are wonderful GPUs. They perform well, never overheated once in two weeks of testing in 30+C weather, and had none of the mild occasional stuttering shown off by the 3870s.

At well under $200, there is nothing on the market that comes close to the 4850, and the 4870 simply has nothing against it at the same price point. These Visiontek 4850s are going in my permanent gaming rig starting today. ?