Of Last-Minute BIOS Updates & Factory Overclocking

As I mentioned earlier, the launch of the Radeon RX 5600 XT is turning into quite a competitive event between AMD and NVIDIA. After initially announcing the card two weeks ago, we’ve seen NVIDIA start to cut prices on the RTX 2060. And that seemingly has elicited a reaction-reaction from AMD as well.

Midday Wednesday, less than a day after AMD made the drivers available to the press, the company sent word that they would be sending out new BIOSes for the cards they had sampled, Sapphire’s Pulse Radeon RX 5600 XT. And sure enough, on Thursday morning new BIOSes arrived, which had a rather material impact on testing results.

I’ll get to the content of the BIOSes themselves in a second, but even before we go that far, I want to note how unusual it is to receive last-minute BIOS changes. For brand-new architecture launches this isn’t unheard of, especially to fix last-minute bugs. But for GPUs that have been out for months – and for cards that are already on the way to retailers – last minute BIOSes are almost never necessary. So when it does happen, it generally means that a vendor is making some last-minute adjustments for competitive reasons, and that’s exactly what’s happening here.

All told, AMD and Sapphire’s last-minute BIOS adjustments seem to have done two things to Sapphire’s card.

  1. Further extended the voltage/frequency curve from 1630MHz to 1760MHz
  2. Increased the memory clock on the factory overclocked mode from 12Gbps to 14Gbps

Indeed, if we take a look at the voltage-frequency curves on the Pulse RX 5600 XT’s quiet mode BIOS, we see the following.

While the display in AMD’s drivers is a bit rough, we can clearly see that AMD/Sapphire has extended the curve, allowing the card to boost to higher clockspeeds at a higher voltage. And the situation is pretty much the same for the performance BIOS as well, with a similar extension of the voltage-frequency curve.

Reference Clocked-Cards

Now, to be sure, AMD has not changed the reference specifications for the Radeon RX 5600 XT. So despite what’s going on, the baseline hasn’t changed. Case in point: the Pulse 5600 XT’s quiet mode BIOS has the same 135W TGP (~150W TBP) power limit both with the old and new BIOS, as well as the same 1460MHz rated game clock. And indeed, adding a bit more to the end of the voltage frequency curve generally doesn’t do much for the card, as higher frequencies are quite expensive from a power standpoint, and the card is generally bottlenecked elsewhere.

Furthermore it’s not clear right now whether similar BIOS updates are going to be issued for lower-end cards. The Sapphire Pulse is a (or at least was a) mildly factory overclocked card, with overbuilt cooling and power delivery. So AMD and its partners may not even want to issue a BIOS update to baseline models, where power and thermals would be more restricted.

That said, this entire update situation gives me the distinct impression that AMD’s plans for the RX 5600 XT changed in the last two weeks. The company was seemingly originally intending on giving the card a relatively low frequency cap, which would improve overall energy efficiency by lowering average power consumption, (still) easily beat the GTX 1660 series, and keep the RX 5600 XT from encroaching on the RX 5700. Instead, the frequency cap has been lifted, which gives the quiet-mode Pulse RX 5600 XT the slightest of performance bumps.

Factory Overclocked Cards

Unlike quiet mode, however, AMD and Sapphire’s BIOS changes are far more significant for the card’s performance mode, which is its default and full factory overclocked mode. Not only has the updated BIOS extended the frequency cap, but it has also increased the card’s power limits, leading to Sapphire adjusting the rated game clock from 1560MHz to 1650MHz. Finally, as the GPU clockspeed improvements on their own aren’t especially potent, the companies have raised the raised the memory clock as well. As a result, the Sapphire Pulse RX 5600 XT is a good bit faster now than it was 5 days ago.

Sapphire Pulse Radeon RX 5600 XT BIOSes
  Shipping Performance BIOS Updated Performance BIOS Shipping Quiet BIOS Updated Quiet BIOS
Base Clock 1235MHz? 1290MHz 1235MHz? 1290MHz
Game Clock 1375MHz 1615MHz 1460MHz 1460MHz
Boost Clock 1560MHz 1750MHz 1620MHz 1620MHz
Memory Clock 12Gbps GDDR6 14Gbps GDDR6 12Gbps GDDR6 12Gbps GDDR6
TBP ~160W ~180W ~150W ~180W
GPU Power Limit (TGP) 140W 160W 135W 135W

As a quick reminder, AMD tends to throw around both Total Board Power and Total Graphics Power. The latter is the value reported by their drivers and only measures the GPU power. TBP includes the memory and everything else

As a factory overclocked card, the Pulse 5600 XT was already shipping with a slightly goosed clockspeed and TBP, making it a 160W card in its default performance mode. The new BIOS has pushed that further north, giving it a TBP of around 180W, or around the same TBP has the Radeon 5700.

More significant than that, however, is that AMD and Sapphire have also increased the memory clock from 12Gbps to 14Gbps. This is very important, because by its very design, the RX 5600 XT’s biggest bottleneck is memory bandwidth – it has the RX 5700’s GPU throughput, but ~35% less bandwidth – meaning that increasing the memory clock delivers a major boost to the RX 5600 XT’s performance.

Case in point, here are benchmark numbers for the Sapphire card clocked at AMD’s reference clocks, AMD’s reference clocks with a 14Gbps memory clock, and finally with Sapphire’s full factory overclock.

Sapphire Pulse RX 5600 XT Performance Scaling (1080p)

The bulk of the performance uplift we see comes from the memory clock increases, not the power limit + clockspeed increases. This underscores the importance of memory bandwidth to this part, and furthermore the significance of this last-minute BIOS change.

The net impact of this change is that the performance delta between a reference-clocked RX 5600 XT and the factory overclocked cards that AMD is sampling to the press is quite significant. In its full factory overclocked mode, the Pulse RX 5600 XT ends up being 8% faster than a reference clocked card, which is far from the largest factory overclocks we’ve ever seen, but larger than the kind of modest overclocks that Pulse cards normally see. It’s also not too far away from the kind of performance delta that separates entire tiers of cards; if it were a 10% boost AMD could call it an RX 5600 XTX and it would fit right in to normal video card product stacks.

Rationale & Results

The big question, of course, is why AMD and Sapphire decided to make such a last-minute change. To answer that, I’ll start with AMD’s official line:

Based on ongoing testing with our board partners, we have raised the GPU core and memory frequencies for overclocked Radeon RX 5600 XT SKUs to take advantage of increased thermal and electrical headroom built into partner’s custom designs. The updated VBIOS has been made available to our board partners for inclusion in select OC SKUs at launch. AMD is dedicated to disrupting the market with industry-leading compute products, and the new VBIOS makes the Radeon RX 5600 XT an even more powerful contender for high-performance 1080p gaming. Previously announced product specs are unchanged, as they remain AMD’s recommended reference design specs.

Meanwhile, I suspect AMD’s motive may be a bit more straightforward: beating the now-$299 RTX 2060. As it so happens, the factory overclocked Sapphire Pulse RX 5600 XT pulls ahead of NVIDIA’s reference RTX 2060 by 2% – which is essentially a wash in benchmark terms – but it allows AMD and its board partners to weave RTX 2060 performance into their product messaging. If AMD can position the RX 5600 XT as an RTX 2060 competitor rather than a GTX 1660 Ti competitor, then it’s all the better for them.

Ultimately, whatever AMD’s original plans might have been, it looks like AMD’s new plan is to take advantage of partners overbuilding some of their RX 5600 XT SKUs. Case in point, if a factory overclocked card ships with 14Gbps memory and the VRMs necessary to do a TBP of 180W – say, if it’s reusing a Radeon 5700 design – then AMD and the partners have gone ahead and adjusted their cards to get more performance. To reiterate, these are factory overclocked cards that are not representative of baseline performance. And at least in the case of Sapphire’s card, these overclocked cards carry a price premium. But it creates a bit of a two-tier situation, with 12Gbps reference cards delivering a certain level of performance, and 14Gbps factory overclocked cards delivering a decent bit more.

Meanwhile, although Bill of Material (BOM) costs are a matter for the board partners rather than something for consumers to be concerned about, I suspect these factory overclocked SKUs are going to wipe out most of the BOM gains from the RX 5600 XT SKU. 14Gbps memory, better PCBs, higher capacity VRMs, and better cooling all run counter to the cost-cutting that the RX 5600 XT would otherwise allow. And as we’ll see in our benchmark results, it also wipes out most of the energy efficiency gains. So AMD is once more trading off higher power consumption for better overall performance.

All told, it reminds me a bit of the Radeon R9 290 and the GTX 460; both cards used last minute tweaks and factory overclocks to try to close a performance gap with the next tier. These kinds of actions certainly make a mess of consumer messaging, but as we’ve seen before, they can be quite effective in very close fights.

The AMD Radeon RX 5600 XT Review Meet the Sapphire Pulse Radeon RX 5600 XT
Comments Locked

202 Comments

View All Comments

  • Ryan Smith - Thursday, January 23, 2020 - link

    Unfortunately Blender doesn't play nicely with new hardware. Or with AMD's currently buggy OpenCL drivers.
  • ozzuneoj86 - Wednesday, January 22, 2020 - link

    The stagnation in the sub-$300 video card market is getting pretty tiresome. I was unimpressed when the GTX 1060 6GB came out in 2016 and was barely faster than the GTX 970 from 2014 (which I bought new in early 2015 for around $250 on sale). Now, 3 1/2 years later we're getting only marginally faster products in the low $200 price range (1660, 5500xt). If you already have a card that was in the $200-$250 price range any time within the past *5 years*, you have to spend $280-$300 to get any kind of noticeable upgrade

    As a comparison, that'd be like if the GTX 970 I bought on sale for $250 in 2015 (an admittedly great price, but not unheard of) had performed no better than a GTX 460... or even a GTX 470. That sounds absurd now, and yet that's what the mid range market has turned into.
  • philosofool - Wednesday, January 22, 2020 - link

    This seems like a strange analysis to me. This card is a legit entry level 1440p card, which has never existed in the sub-$300 range before.
  • cmdrmonkey - Wednesday, January 22, 2020 - link

    nVidia is charging more and giving us less than they ever have in the past because they have no meaningful competition from AMD.
  • Spunjji - Thursday, January 23, 2020 - link

    It's true that Nvidia haven't offered anything like the value proposition that the GTX 970 was on its launch since then, and things have definitely slowed down in the GPU arena. I'm not entirely on board with this criticism overall, though.

    First off, it's a bit unfair to compare the price of a card you got on sale with launch pricing. The 970 launched at $330, which was an absolute steal but still more than $250.
    Second, the 1060 provided performance that was better than a GTX 980 (and about 20-40% better than a 970, depending on the game and resolution) for $250. AMD countered with the 580 and, well, to be fair that was pretty much that until now.

    That's why it confuses me that you'd complain now, when the 5600XT (and the price drops it inspired) means we can *finally* get performance that's 50-100% better than the 970 at a lower price. It took about twice as long as it used to, for sure, and that just seems to be how things are now.
  • ozzuneoj86 - Thursday, January 23, 2020 - link

    Sorry, I wasn't trying to make an unfair comparison. I was just thinking more in terms of time... 5 years, which used to be an incredibly long time in this industry. If we're comparing launch dates and pricing, then it has taken six years to get a large upgrade for a GTX 970 at a lower price... though arguably the RX 5700 fit that bill last summer when it was often available on sale for $300 or a little less. To me, that makes the 5600XT with less memory a lot less interesting for only $20 less. These cards are fine if people have the money for them, but the slow progress is what is getting to me. Compared to the massive changes we've seen AMD bring about in the CPU market, the GPU market is very stale. There aren't any no-brainer purchases at any tier if you have a mid-range GPU from within the past 5 years. This is probably the closest we've come, as you said, but its by such a small margin. If we had performance like this for closer to $200 it would have shaken things up and made GPUs interesting again. Instead, we have the same back and forth about whether it's worth it to spend another $20 and get last year's 2060, or to buy a 4 year old used 1070 for $190 on eBay, or to simply lower the settings a couple notches and stick to the 6 year old GTX 970.

    This isn't really relevant, but... I guess my 970 actually ended up being more like $220, because I got a $30 check from nvidia due to that memory settlement. And then, well, I did sell the DLC codes that came with the card so it was closer to $200. That ends up being like $40 per year... thanks nvidia! :P
  • peevee - Wednesday, January 22, 2020 - link

    Ryan, because you mention all the time that 6GB of VRAM might not be enough soon, can you write an article explaining the major uses of VRAM by various applications?

    It seems like neither compressed textures nor 3d models of everything needed at the same time (or within a few seconds) could take as much, and everything else can be preloaded quickly on the fly, especially with PCIe4x16... as it allows to update half of that 6GB VRAM every 1/10th of a second.
  • cmdrmonkey - Wednesday, January 22, 2020 - link

    Looks okay, but nobody is going to buy it because nobody actually buys AMD video cards. If you doubt this look at the Steam hardware survey.
  • Korguz - Thursday, January 23, 2020 - link

    " look at the Steam hardware survey. " and that is 100% reliable ? BS. not every one has and uses steam, so no.. NOT a reliable metric. those that i know.. dont all run nvidia cards, some have radeons, and they dont have steam...
  • cmdrmonkey - Thursday, January 23, 2020 - link

    Steam has over 1 billion accounts and 90 million monthly users. The hardware survey has a sample size in the millions. Medical and psychological studies don't even have sample sizes like that. I'd say it's a damn good indicator of what most people are using.

Log in

Don't have an account? Sign up now