Leak: ATI Radeon 4800 Gets 480 Stream Procs

Chicago (IL) – Next month will be more important for AMD’s graphics division than any other month of this year: The company is set to release the (RV770) 4800-series of graphics cards, including the models 4850, 4870 and 4870 X2. While the launch is still very much a secret, we were able to receive details and the specifications of the new cards – which will put ATI back into the ring with NVIDIA.

The product launch will be significant not just because of the availability of chips, but also because of the fact that AMD/ATI is returning to a true six-month refresh cycle of its graphics products. Looking back, the 2900/3800 series was not really a six-month refresh, since the 2900XT was eight months late. The Radeon 3800 series was launched in November of 2007 and since the Radeon 4800 is pretty much set for a May 2008 launch (cards went into production in March), the ATI team has returned to its traditional cycle. There is a good chance that the acquisition troubles have been digested and new products will be surfacing in a more predictable fashion from now on.

Specifications: 480 stream processors, 800+ million transistors

Let’s have a look at the general specifications of the Radeon 4800 series, which is based on the RV770 graphics processor. The new GPU has 480 stream processors or shader units (96+384), 32 Texture units (up from 16), 16 ROP (same as 2900/3800), a 256-bit memory controller and native GDDR3/4/5 support (expect GDDR3 and GDDR5 memory versions of RV770 cards).

The transistor count has jumped from 666 million in the RV670 to more than 800 million transistors in the RV770.

The manufacturing process has been carried over: TSMC is producing the GPUs in 55 nm (the process itself is called 55GC).

The Radeon 3800 series had a serious flaw called texture low fill-rate, which was addressed by ATI with an increased number of TMUs (Texture Memory Unit) from 16 to 32. The specifications indicate that 16 TMUs can address 80 textures on the fly, which means that 32 units should be able to fetch 160 in the RV770: This should allow the new GPU to catch up with NVIDIA’s G92 design. However, the G92 has 64 TMUs that were enabled gradually (some SKUs shipped with 56), resulting in a fill-rate performance that beat the original 8800GTX and Ultra models.

ATI’s RV770 will be rated at a fill rate of 20.8-27.2 GTexel/s (excluding X2 version), which is on the lower end of the GeForce 9 series (9600 GT: 20.8; 9800 GTX: 43.2 9800 GX2: 76.8).

New for the 4800 series is AMD’s decision to split the clock of the GPU and shaders, following a move that NVIDIA made with the GeForce 8800: Back then, the shaders were clocked at 1.35 GHz, while the rest of the chip ticked at 575 MHz.

The Radeon 4850 will debut with a GPU clocked at 650 MHz, while the shader array will be clocked 850 MHz. The 4870 has an 850 MHz core and a 1050 MHz shader clock. The fill-rate is closely tied to the GPU clock: The 4850 sports a fill-rate of 20.8 GTexel/s (32 TMU x 0.65 GHz), while the 4870 achieves a 27.2 GTexel/s (32 TMU x 0.85 GHz) performance.

The RV770 GPU is equipped with a 256-bit memory controller (512-bit for the Radeon 4870 X2: The R700 represents just two RV770 GPUs slapped together). 4850 GDDR3 models will come with 256 MB or 512MB of GDDR3 memory, clocked at 1.14 GHz (2.29 GTransfer/s), resulting in a memory bandwidth of 73.2 GB/s. 4850 GDDR5 versions integrate 512 MB of GDDR5 memory clocked at 1.73 GHz (3.46 GT/s), supporting a bandwidth of 110.5 GB/s. 4870 GDDR5 models will get 1 GB of memory clocked at 1.94 GHz (3.87 GT/s), achieving a maximum bandwidth of 123.8 GB/s.

Flagship dual-GPU 4870 X2 cards will include 2048 GB of GDDR5 memory clocked at 1.73 GHz. The Radeon HD 4870 X2 will be introduced at a later date (and could see spec revisions).

You may be wondering why ATI decided to go with "odd clocks" for the video memory part and the answer may surprise you: According to our sources, AMD is no longer clocking the cards towards performance, but towards power efficiency instead. This happened first with the Radeon 2900XT (which was bleeding lots of current) and continued with the 3800 series.

Overclockers should have no problem reaching the magic 4 GHz mark for the GDDR5 memory (1.94 to 2.0 GHz), but that, of course, would break the sub-150 watt thermal-envelope these cards were designed for. More than ever before we can envision AMD and add-in-board vendors offering a “Radeon HD 4870 TDP - Saves you $$$ on Electricity Bills!" or a "4870 Lots of Horsepower and Power Hungry" edition of these cards.

The 4850 in fact is in a 110 watt power envelope (the 4850 256MB GDDR3 is a sub-100 watt card). There is, by the way, a big difference in power consumption between GDDR3 and GDDR5 versions. 512MB GDDR5 memory will consume 34.5% less power than a model with 512MB of GDDR3 memory.

As you can guess, a 110 watt power envelope means that 4850 boards will have a single-slot cooler, while the 150 watt part (4870) should be available as single-slot part as well (from at least one partner, 1-slot vapor chamber part). Most of manufacturers will, however, offer dual-slot cooling for the 4870.

4870 X2 is an interesting version. AMD did not send out any specs to its partners and it is expected the board will be a bit more than just a 3870 X2 two RV770 GPUs. ATI is said to be making some changes, most notably in the PCI Express department (3870 X2 is actually a PCIe 1.1 part). The product will keep the aggressive pricing in place, and, according to our sources, will scale much better than 3870 X2.

The 4850 256MB GDDR3 version will arrive as the successor of the 3850 256MB with a price in the sub-$200 range. The 4850 512MB GDDR3 should retail for $229, the 4850 512MBGDDR5 will set you back about $249-269. The big daddy, the 1GB GDDR5 powered 4870 will retail between $329-349.

When it will become available the 4870 X2 will hit the market for $499.

Our sources indicate that the launch is only weeks away, so expect your usual hardware sites to feature dozens of reviews of 4850 and 4870 with GDDR5 memory, since this memory type will be a hot topic in months to come.

  • caamsa
    I am in the market to pick up a new video card in the near future like maybe this fall. Hopefully these will be decent cards and will not be over priced.
  • korsen
    Those idiots will never learn... 16 ROPS? why the hell did everything else increase except that?

    ATI keeps screwing up royally and unless the 48x2 can even make a dent vs the 9900's, AMD is going to drown in it's already screwed up debt.

    The same way people get pissed off for companies releasing beta software as GOLD, i'm sure most of us tech geeks are getting tired of NOBODY TESTING THEIR OWN CRAP FOR PROBLEMS.
  • lopopo
    This is kinda upsetting. Here is this product that sounds like it will preform well and its released at a time when AMD's future is dubious. Even if it gets rave reviews, I would be scared to buy.
  • Gravemind123
    Since they are opening up all of their documentation necessary to make 3rd party drivers for linux, I would assume that someone would make a windows driver with the data if AMD/ATI went under.

    I suppose we will see if it is the lack of TMUs or the lack of ROPs that was holding back the 3800s, hopefully increasing the shaders and TMUs will increase performance enough to make it a viable competitor to the 8800GTX, I still can't believe that the 8800GTX has hardly been surpassed in almost a year, graphics companies both need to get off their asses and make something new.
  • korsen
    but you also need to consider that gfx increases come out every 6 months as opposed to every two years for processors. We're looking at huge jumps here. I say they should go into processor schedules and start making something that's worth upgrading to.

    It's like, meh, this new generation is only slightly better than the last, i'll pass. It's like comparing an old celeron to a new celeron instead of an P4-HTT vs an Athlon vs a Core2 vs a Nehalem vs AMD's brain-child. Celeron on Celeron action is a stalemate and LAME.
  • sailer
    If AMD/ATi can get these cards out and they perform close to promised, I can see a couple of them going into my computer. But I'm not holding my breath about it, as I can remember the anticipation I had last year for the 2900, after which I bought a 8800 GTS 640.
  • Andrius
    Either ATI has accepted it's role as "the (class) second GPU provider" or something went wrong with the other 16 ROPs.

    If something bottlenecks an architecture the logical step is to increase it until it stops being the bottleneck. It would seem ATI thinks 64x5 stream processors was the bottleneck not the 16 ROPs.

    The strange decisions at AMD/ATI continue ever since AM2 came out 2 years ago :heink:
  • homerdog
    9079553 said:
    Those idiots will never learn... 16 ROPS? why the hell did everything else increase except that?
    Because ATI is doing AA in the shaders instead of the ROPs (they call them RBEs by the way). 16 should be enough.
  • thomasxstewart
    Before 4x, there was dynamic controllers on mainboards, they where 80X2 then 80X4 & so on?. Dynamic controllers are more situation specific? & 8 x80 or today 16x80?. Isn't game card little mainboard, with each unit little transistor controller or effector. 8 x80 being half? total potential.Or is it TOP at ~640, as seems design here. It is great improvement from much less assembly machins on board, i'm sure. Is this multiboard thing forced upon US simply because No one has designed more complex mainboard Host controllers for years?6&7/XX

  • eodeo
    The new GPU has 480 stream processors or shader units (96+384)

    What does this mean? 96 shaders (same as old g80 gts) + 384 vertex processors? Or what I know that 2900 and 3870 have 3x better geometry performance than nvidia g80/92, but I never did understand how come that its shader performance is so low compared to nvidas solution. Does anyone have an explanation for this?

    And 32 TMU?. What the bleep is that? Texture units?? what part of UNIFIED processors am I not understanding right?

    And another thing- 3870 supposedly has 64x5 processors making it x5 wasted as it competes to 9600gt which has 64x1 processors. What is ATI doing that it needs x5 to get same results as nvidia with x1?