Original Link: https://www.anandtech.com/show/198

Overclocking the TNT

by Anand Lal Shimpi on September 13, 1998 8:53 PM EST


On the phone with a good friend of mine, Manveer Wasson, joking around about the usual an idea hit us both that would prove to be a challenging yet extremely beneficial experiment. 

AnandTech Exclusive

At the time of testing the Canopus Spectra 2500 I found myself admiring the excellent design of the heatsink/fan combo Canopus outfitted their first Riva TNT based card with, Manveer poked a little fun at my extreme admiration of the fan by telling me to develop my own cooling solution.

This is when the idea hit us, nVidia developed the Riva TNT with the intention of running the chip at a 125MHz clock speed, so why not give it a try at that clock speed? Due to the price of 0.25 micron silicon real-estate, and the increasing pressure to release a competitor to 3Dfx's Voodoo2, nVidia was forced to ship the TNT manufactured on 0.35 micron silicon wafers.  The greater the thickness of the silicon manufacturing process, the more heat the processor will generate, and therefore nVidia managed to eliminate the possibility of running their highly anticipated TNT chip at the 125MHz they originally proposed.  Here's where the problems begin.

Proposing an Acceptable Alternative

You must keep in mind that nVidia built the TNT to run at 125MHz, unlike the Voodoo2 which was designed to run in the 90 - 95MHz range and unlike most system microprocessors which are designed to run at their official clock speed alone, the TNT was manufactured with the intention of being run at 125MHz.   Once the first revisions of the literally flaming 0.35 micron silicon hit the hands of the engineers it was obvious that a change in the recommended clock speed had to be made.  Initially the change was supposed to be from 125MHz down to 100MHz, however instability at 100MHz required that the drop from 125MHz down to double digits be made in order to ship a reliable product.  Just about a week before releasing the TNT chipset to the public, nVidia changed their recommended clock setting once again, this time to 90MHz to increase stability.

A Lower Clock Speed Means Lower Performance

As you might be able to guess, this drop in clock speed also resulted in a drop in performance.  The original specifications for the Riva TNT chip indicated a 250 million pixels per second fill rate which would put it over the top of the Dual Voodoo2 SLI setup that had reigned for the past few months.   However that 250MP/s fill rate was calculated using a TNT clocked at 125MHz, not at 90MHz.  At 90MHz, the fill rate drops to 190MP/s, conceptually still much higher than a Voodoo2 but in reality not all that much better.  In an attempt to explain the lack of the TNT being the proposed Voodoo2 SLI killer it was initially dubbed, nVidia reassured the public that early in 1999 they would make the move to a 0.25 micron manufacturing process for the TNT and release a revision of the chip designed to run at full speed, being 125MHz.  In theory, once nVidia makes that jump, we should have a definite Dual Voodoo2 SLI killer on our hands...right?  Well, what nVidia wasn't counting on was an actual test to prove this theory, hang on to your seats, as AnandTech managed to develop a custom made cooling system made specifically for the Riva TNT chipset which allowed it to be clocked at 125MHz and beyond!  How were the results?  Is the next revision of the TNT worth waiting for?  Did nVidia "bend the truth" with the possible performance of their TNT chip?  Let's find out as we push the limits of nVidia's 2D/3D wonder in an AnandTech exclusive, Overclocking the Riva TNT - Current & Future Expectations.

The nVidia Riva TNT - Current Expectations

It is obvious a great portion, this editor included, of the PC enthusiast population became preoccupied with the search for the elusive Voodoo2-killer.  This whale hunt ended in much disappointment, but fortunately, it managed to provide for a wonderful combatant against the hype that many manufacturers had to generate to promote their products over 3Dfx's reigning king, the Voodoo2. 

A Direct3D masterpiece, the Riva TNT chipset already provides performance greater than that of a single Voodoo2 and comparable to a Dual Voodoo2 setup under Direct3D situations on virtually all processor platforms, Super7 excluded.  The TNT's OpenGL performance will be improved in the near future with better drivers and a more thoroughly optimized ICD, however currently, the TNT already provides with performance that is around the mark of a single Voodoo2 while offering clearly superior image quality.  All of this can be achieved for a price tag falling in the $169 - $199 range.  The extremely competitive 2D performance of the TNT rounds off the extremely positive outlook the TNT will give even the most pessimistic user.   You want to keep in mind that the Riva TNT is still the best option for a well rounded 2D/3D accelerator, and definitely gives 3Dfx a run for their money, but is it a sin to want the performance nVidia originally assured we would be getting?  Not at all in the mind of this overclocker...



The Cool Solution

Sitting at home, playing around with our newly developed cooling method, it was time to put it to the test.  If you recall, nVidia claimed that their TNT would be a Voodoo2 SLI killer if it operated at its full potential, that being 125MHz.  The problem at hand was overclocking a 0.35 micron processor running at 90MHz to 125MHz.  In terms of system processors, this could be compared to overclocking an original Pentium II 233 to around 333MHz, and if you've ever had any experience with the first 233's that hit the streets, that wasn't too realistic of an overclock from a reliability perspective. 

By constructing a cooling system designed to meet the needs of the task set before us, we managed to put together a crude system that would accomplish the cooling necessary to get the TNT to produce benchmarks at 125MHz, our goal.  While the specifics of the cooling system cannot be disclosed as of now, we are currently working to see if it is feasible for mass production and you will see up to date information on the topic once the word is in.  If it can be manufactured for use in normal home systems, you can expect it to retail for under $30. 

Using a direct method of modifying the clock settings of the Riva TNT chipset, AnandTech ran tests at the standard 90MHz as well as the following overclocked settings: 110MHz, 115MHz, 120MHz, 125MHz, and 135MHz.   Remembering that 125MHz is the sweet spot for the TNT, we should expect it to become a Voodoo2 SLI killer at that speed if nVidia's claims were well founded.

The Test

The Slot-1 Pentium II Test System AnandTech used was configured as follows:

  • An Intel Celeron 300A, and an Intel Pentium II 400 on an ABIT BX6 Motherboard

  • 64MB Mushkin SEC PC100 SDRAM

  • Western Digital 5.1GB Ultra ATA Hard Drive

  • AOpen 32X IDE CD-ROM Drive

  • Windows 98 with all of the latest patches/drivers installed

The benchmark suite consisted of the following full version game titles

  • Forsaken - Running the Nuke Demo

TNT vs Voodoo2 - OpenGL - Quake 2

Celeron 450A

Timedemo - 800 x 600
- demo1.dm2 % Change crusher.dm2 % Change
nVidia Riva TNT @ 90MHz (Default) 60.3 0% 40.8 0%
nVidia Riva TNT @ 110MHz (Overclocked) 68.4 +13% 40.9 +0.2%
nVidia Riva TNT @ 115MHz (Overclocked) 69.4 +15% 41.6 +2%
nVidia Riva TNT @ 120MHz (Overclocked) 71.5 +19% 41.6 +2%
nVidia Riva TNT @ 125MHz (Overclocked) 72.4 +20% 41.6 +2%
nVidia Riva TNT @ 135MHz (Overclocked) Failed - Failed -
3Dfx Voodoo2 12MB @ 90MHz (Default) 60.3 - 36.6 -
Dual 3Dfx Voodoo2 12MB SLI @ 90MHz (Default) 87.6 - 39.8 -

Quake 2 Comparison

Using the demo1.dm2 scores as the realistic high end of the performance spectrum, and the crusher.dm2 scores as the lowest realistic performance scenario AnandTech ran tests at clock speeds ranging from 90 - 125MHz on the specially cooled TNT board.  The results were amazing, at 125MHz, the performance increase under Quake 2 was a full 20% on the high end!  While this doesn't bring a single TNT up to par with a Dual Voodoo2 SLI setup per the original claims of its performance, it does give you a considerable advantage over a single Voodoo2 at a much more reasonable cost.  You decide whether or not this is worth waiting 5 - 6 months for.  On an unrelated note, Canopus is promising new drivers on the 28th of September that will increase performance by around 20%, if Canopus claims are as great as they say they are, then you can except an update to this comparison in the near future with some benchmarks possibly closer to those of a Dual Voodoo2 SLI setup on a single TNT board.  

q2-graph2.gif (12476 bytes)

The crusher.dm2 scores show very little improvement with the increased clock speed, the TNT still holds the advantage over even a Dual Voodoo2 setup due to its excellent AGP 2X implementation which allows for much data to be transferred over the bus in the most taxing of situations such as the one illustrated by the crusher benchmark.

TNT vs Voodoo2 - Direct3D - Forsaken

Pentium II 400

800 x 600
- Nuke Demo % Change
nVidia Riva TNT @ 90MHz (Default) 130.88 0%
nVidia Riva TNT @ 110MHz (Overclocked) 138.93 +6%
nVidia Riva TNT @ 115MHz (Overclocked) Failed -
nVidia Riva TNT @ 120MHz (Overclocked) Failed -
nVidia Riva TNT @ 125MHz (Overclocked) Failed -
nVidia Riva TNT @ 135MHz (Overclocked) Failed -
3Dfx Voodoo2 12MB @ 90MHz (Default) 96.26 -
Dual 3Dfx Voodoo2 12MB SLI @ 90MHz (Default) 177.13 -

The Direct3D tests would only complete on the TNT clocked at 110MHz, Forsaken would consistently crash at any higher clock speeds.  Judging by the scores AnandTech obtained from running Quake 2 at 125MHz and comparing to the increase achieved by using a Direct3D benchmark, you can expect a 125MHz TNT to bring in Forsaken scores in the 140 - 150 fps range, still under that of a Dual Voodoo2 setup, but definitely above that of a single Voodoo2 card and definitely worth the money.  Do remember that this performance isn't available now, and unless the cooler we constructed can be manufactured at a reasonable cost, you will have to wait another 5 - 6 months before you can get this performance increase. 



Even at 125MHz, the TNT is no Voodoo2 SLI killer, but you have to keep in mind that you're only spending $199 on a TNT card compared to the $300+ on a Voodoo2 SLI setup, not to mention the cost of coupling those Voodoo2 cards with a 2D accelerator even half as good as a TNT.  For the price, the TNT is a great product, however it isn't worth waiting 5 - 6 months to get a 10 fps increase when the TNT is clocked in at 125MHz.  If you're looking to buy something now, don't go the Voodoo2 route with the intent of getting a TNT later, either go one way or the other.  As more news about the AnandTech cooler develops you will see updates about the progress on the frontpage, until then, happy quaking.



After the initial success with overclocking nVidia's combination wonder, the Riva TNT, there were quite a few requests to perform a more in-depth analysis of what the future may hold for nVidia and their flagship product.   The overclocking experiment was a startling success, however many questions remained unanswered. 

Exclusive - Continued

Next year, when nVidia does chose to launch their TNT product at 125MHz, what will the speed of the on-board RAM be?  Will the 0.25 micron design allow for greater overclocking possibilities?  How about greater resolutions at the 125MHz clock speed?  Taking a second look at the AnandTech Exclusive, let's answer these questions as we bust out the cooling system for another trial run.

125MPH in a 110MPH Zone

That was the problem with the original tests, if you notice the results overclocking the Canopus Spectra 2500 produced, the performance barrier seemed to be at around the 110MHz limit.  The 20MHz increase in clock speed resulted in a tremendous 13% performance increase, however bumping the clock speed up to 125MHz, a full 15MHz increase in clock speed, only resulted in a 20% improvement over the scores at the default 90MHz clock speed, and a meager 7% increase over the scores at 110MHz. 

The reason?  The memory clock of the TNT card remained at 110MHz, the default set forth and recommended by nVidia.  By overclocking the chip to a level above 110MHz we were essentially ruining the balance of power between the speed at which the TNT can process the data and the speed at which the RAM can retrieve it.  Normally, with the TNT clocked at 90MHz and the memory clocked at 110MHz, the TNT can process all geometry data and retrieve data from the local [graphics] memory in an efficient manner.  When the overclocking was performed, the speed of the processor was increased, without compensating for this performance increase by upping the clock speed of the memory.  This explains the diminishing returns once the TNT's clock speed surpassed the speed of the memory. 

Removing the Bottleneck

With the memory clock established as the bottleneck in this case, what is there left to do than remove it?  That is easier said than done.  When the TNT is released at full speed in February of next year you can expect it to ship with considerably faster SDRAM or even SGRAM depending on the cost of the chips at the time of production.  While the actual clock speed may vary, we should be able to expect that the memory will run at a speed considerably greater than 125MHz, therefore eliminating the bottleneck we just discussed.  Unfortunately, the availability of such high speed SDRAM is close to nothing for the price point most TNT cards will be selling for, meaning it is unlikely that we'll be able to overclock the RAM on most TNT cards too far beyond the recommended levels.



Two types of TNT

The Canopus Spectra 2500 which was used in the original TNT Overclocking tests featured 16MB of Hyundai SDRAM running at 110MHz.   The test card AnandTech used wouldn't reliably accept any memory clock speeds greater than 115MHz, a fairly limiting factor in removing the bottleneck recently set before us. 

An interesting thing to note is that the Creative Labs and STB TNT cards don't use the same Hyundai SDRAM Canopus seems to be shipping with their Spectra 2500.  Instead they use Fujitsu SDRAM which seems to have a better time overclocking, in fact, the Creative Labs Graphics Blaster TNT AnandTech just recently started testing would allow for its memory to be taken up to 130MHz without nearly as many crashes as the Spectra 2500 experienced at 120MHz. 

Canopus either chose to go with the cheaper SDRAM to keep the costs of their luxurious TNT card down to a competitive level, or has yet to experiment in the Fujitsu arena.  Either way, it looks like that alone takes Canopus out of the running for the most overclocking friendly TNT board. 

Cooling the TNT

Here's the big question, how did Anand do it?   Unfortunately the specifics of the cooling system used cannot be disclosed at this time, however the chip itself was cooled to below the freezing point.  Since we've managed to eliminate the Spectra 2500 as a possibility for the true overclocker's test, how do we get the Graphics Blaster TNT and STB Velocity 4400 to run at higher clock speeds without the cool fan Canopus outfitted their TNT with? 

Fortunately, the heatsinks used on the CL and STB cards is quite reminiscent of the older 486 heatsinks.  Low and behold, a quick trip out to Radio Shack for a 486 fan will reveal that virtually any small fan will easily attach to either of those two cards.  Considering that the Creative Labs card sells for $169 retail, an extra $7 can bring you one step closer to the $199 Canopus Spectra 2500 without too much of the added cost. 

AnandTech is still working on feasibility tests for the cooling idea, however as of now the outlook doesn't look too incredibly promising.  More information about the cooling methods will be disclosed at a later point in time.

The Difference is at 1024 x 768

Each test run performed at an overclocked clock speed with the Riva TNT must be monitored carefully as much instability can result from a single mistake in the setup process.  An individual test can sometimes take up to three or four runs to complete properly due to a number of outside factors, this complexity limits the amount of tests than can be run at a given time.  For this reason, the original article only featured benchmarks at 800 x 600.  Logically, one would assume that increasing the resolution to 1024 x 768 would produce a greater increase in performance over the default clock speed in comparison to the increases seen at 800 x 600.  Following up that statement with benchmarks proved just that, the increase from 90 to 120MHz at 800 x 600 proved to be 19% while at 1024 x 768 the increase was 23% showing the trend towards producing a chipset that can up the resolution recommendation to 1024 x 768.  Once the TNT makes it out at 125MHz, regardless of the speed of the memory, you can expect it to be a fairly decent performer at 1024 x 768 on most of the mid-range processors as well as an excellent companion for high-end processors at that resolution. 

5MHz Can Make a Difference

Since the Graphics Blaster TNT was still under initial evaluation at the time of the follow-up to the TNT article, it couldn't be used for the benchmarks, however as the Canopus Spectra has been able to show us, a 5MHz increase in memory clock does make a noticeable difference in overall overclocked performance.  The 5% increase in clock speed results in an increase in performance of about 3 - 4%, judging by this one would be led to assume that running the SDRAM at a greater frequency (130MHz for example) would yield a much higher overall performance increase, possibly bringing us one step closer to that desired Voodoo2 SLI performance.   For those results however, we'll have to wait for the Creative Labs TNT Review which is nearing it's completion as this article is being written.

So once again, armed with the Canopus Spectra 2500, and the AnandTech cooling system, let's put the chip to the test as the TNT is overclocked once again for an even greater trial of its true power.



The Test

The Slot-1 Pentium II Test System AnandTech used was configured as follows:

  • An Intel Celeron 300A clocked at 450MHz on an ABIT BX6 Motherboard

  • 64MB Mushkin SEC PC100 SDRAM

  • Western Digital 5.1GB Ultra ATA Hard Drive

  • AOpen 32X IDE CD-ROM Drive

  • Windows 98 with all of the latest patches/drivers installed

The benchmark suite consisted of the following full version game titles

  • Forsaken wasn't used as it remained too unstable at overclocked CPU/RAM speeds

  • Quake 2 v3.17 using Demo1.dm2

TNT vs TNT - OpenGL - Quake 2

Celeron 450A

800 x 600 1024 x 768
- demo1.dm2 % Change demo1.dm2 % Change
nVidia Riva TNT @ 90MHz MCLK = 110MHz 60.3 0% 38.6 0%
nVidia Riva TNT @ 110MHz MCLK = 110MHz 68.4 +13% 43.9 +14%
nVidia Riva TNT @ 115MHz MCLK = 110MHz 69.4 +15% 45.3 +17%
nVidia Riva TNT @ 120MHz MCLK = 110MHz 71.5 +19% 46.2 +20%
nVidia Riva TNT @ 125MHz MCLK = 110MHz 72.4 +20% 47.1 +22%
nVidia Riva TNT @ 115MHz MCLK = 115MHz 71.2 +18% 46.1 +19%
nVidia Riva TNT @ 120MHz MCLK = 115MHz 72.7 +21% 47.4 +23%
nVidia Riva TNT @ 125MHz MCLK = 115MHz 74.1 +23% 48.3 +25%

Conclusion

Not much explaining is necessary here, at 125MHz, with a memory clock (MCLK) speed of 115MHz, the Riva TNT can really give you a nice performance increase, especially at 1024 x 768.  This performance trend would lead one to believe that being able to run a TNT at 125MHz with a memory clock speed of 130MHz would provide a performance increase great enough to make the TNT a strong enough competitor to 3Dfx's empire to tilt the favor of quite a few Voodoo2 supporters.   Just imagine the performance increase a true 0.25 micron TNT running at 125MHz with a 200MHz memory clock would yield...

To be Continued...

Log in

Don't have an account? Sign up now