This is a dual GPU engineering sample of

an AMD ATI HD 4870X2

that the company shipped back in 2008.

We bought this from a viewer and

although AMD didn't ship a variant of

what this eventually became, we've never

seen an engineering sample of it pop up

online, and this one's got some

interesting differences with it. The

retail version of this was the single

fastest graphics card when it launched

and forced Nvidia to slash prices of its

GTX 280 by $220 and its GTX 260 by $110

to contend with AMD's offerings

according to Techreport and Anitech from

the time. The release of this dual GPU

or the retail version of it came at a

pivotal moment for AMD less than two

years after acquiring ATI and only two

months before spinning out its foundry

business. Andy hadn't yet begun to

contend with the aftermath of the great

recession because it was the beforemath

or the middlemath of the great

recession. The 2008 financial crisis was

only really just beginning and that

would kick off one month after in

greater form when the layman brothers

collapsed a month after the dual GPU

launched for AMD. And just to be clear,

this didn't cause the layman brothers to

collapse. Although in a certain way that

would be sort of funny except for the

the part where the the financial crisis

happened, but otherwise it would be

funny. AMDy's stock was worth about $5

per share in August of 2008, falling to

$3.50 per share just 2 months later

after this card launched, setting it up

for failure long term. Following October

of 2008, the price plummeted to around

$2 per share by February of 2009. The

GPU that we're looking at today was part

of AMD's last flagship launch before it

entered a freef fall. It was also around

this era that AMD dumped its real estate

in the lease back deal, signaling what

almost became a death spiral. So, this

dual GPU video card that we're working

on today that our viewer found in I

think was a thrift shop or something

like that. uh this would have launched

right at the beginning of a long period

of turmoil for AMD the industry in

general but in particular for AMD long

before Ryzen was ever even someone's

idea that would only come to save AMD

nearly a decade later from the launch of

this card. We're fortunate enough to

have been able to buy this from a

viewer. We've actually got a lot of

really unique hardware from viewers over

the years. So, if you find stuff that

you think we might find interesting for

videos and you want to loan it to us or

sell it to us, you can email our tips

line, tips gamers.net. But our thanks to

uh Muzz for sending this over to us. We

bought it from them. And we're going to

get started with a tearown of this

engineering sample dual GPU card from

AMD to see what the differences are from

retail and try to get it to turn on.

Before that, this video is brought to

you by ID Cooling and the Frozen A720

cooler. The A720 air cooler performed

well in our testing last year. The A720

is a relatively high-end dual 140mimeter

air cooler with seven heat pipes. We

found the use of larger fans can be

beneficial to acoustic performance given

the thermals. Although, you'll want to

check your case for compatibility given

the taller nature of the cooler and its

fans. ID Cooling uses an all black look

for its A720 and includes mounting

hardware for all modern sockets. Learn

more at the link in the description

below. So, we had a lot of trouble

getting this card to work. We're going

to talk about that in a little bit here,

but today what we're focusing on is some

of the history of this card. A lot of

nostalgia from this era for a lot of

people. Personally, I had a system with

two 4870s. So, they were in Crossfire

and I think it was called the Ant

Skeleton. I think it was the big version

of the case. They had a mini and a large

one. That was my build for that era. I

remember playing Battlefield Bad

Company, whatever number it was back uh

shortly after in the 2010 2012 era with

those two 4870s. And uh this is kind of

familiar to that except it's one card

and an engineering sample which I

personally could have never dreamed of

having back at that time. So we're going

to try to get it to boot. We ran into a

lot of problems though. We'll talk about

that. Uh this is also a good opportunity

though for the history and the tear

down. Back when AMD's website looked

like it cared about gaming, rather than

using AI to spit out maps about Nebraska

and Texas during keynotes about AI for

education, the company announced its ATI

Radeon HD 4870X2.

The site feels like a throwback.

Gradients used for the buttons on the

website, a borderless styling, red and

black everywhere, and simple navigation.

It was a simpler time. It's a throwback.

AMD released the ATI Radeon HD 4870X2 in

August of 2008. Some things don't change

apparently, including naming schemes for

computer hardware. AMD marketed the card

towards what they said were ultra

enthusiasts and positioned its $550 dual

GPU against Nvidia's 9800 GX2 and GT200

series lineups. Today, that'd be about

$820 with inflation. This followed

Nvidia's famous 8800 GPU series, best

known for being able to run Crisis

reasonably well. At the time, Anitech

commented saying, quote, "It looks like

Nvidia's standards have changed largely

thanks to AMD, and now the key players

in Nvidia's lineup are priced more

realistically end quote." AMD's

marketing noted that the card had quote

unparalleled anti-aliasing and

anisotropic filtering. End quote. with

the features tab noting DirectX 10.1

support giving us a timestamp of an era.

The officially launched 4870X2 spec

sheet lists a 750 MHz GPU clock, a 900

MHz memory clock, 800 shading units, 40

TMUs, 16 rocks, 1 GB of GDR5 memory, and

a 256bit bus width per GPU. And again,

that's with two GPUs total. So you're

getting 2 GB total of GDR5 across the

whole board. The card also has a 286

watt TDP PCIe 2.0 x6 interface, two DVI

ports, and one Svideo port for display.

Dual GPUs can't access the resources

across the GPUs. So, these specs are

reserved for each GPU individually.

Crossfire means that you could put two

of these together, which would give you

four GPUs in two slots. Theoretically,

AMD basically took two RV770 processors,

from what we could tell, from its

already released ATI HD 4870 nonX2s.

They doubled the memory of each from 512

megabytes to 1 GB per GPU and combined

them on a single PCB along with a PLX

PCIe switch between the GPUs. Following

the same strategy introduced in its HD

3800 series and confirmed by Anitech of

the time, AMD competed at the high end

by multiplying its GPUs, effectively

planting two on the same PCB rather than

the traditional approach of using a

single flagship GPU at the high end and

then cutting it down for variations for

the mid-range and maybe low-end market

segments, if not spinning out something

different entirely. Andy decided

strategy was to build perfect solutions

for $200 to $300, which would be $300 to

$450 today. Andy stuck with that

strategy later as well, like with the RX

480. AMD's single card approach for two

GPUs granted users some key advantages

over traditional Crossfire

configurations by requiring fewer PCIe

slots and fewer PCIe power cables in

addition to eliminating motherboard

support for Crossfire which is something

that reviewers of the time noted

regarding AMD's newfound plan of attack

and tech reported quote so does AMD's

approach invalidate Nvidia's big

monolithic GPU strategy not exactly

while it is true that two RV770s can

outperform a single GT200 in many cases.

You could also make the argument that

two GT200s could outperform anything

that AMD could possibly concoct. Three

and four-way crossfire scaling isn't

nearly as good as two-way. AMD strategy

makes sense for AMD, but it's

fundamentally no different than what

Nvidia is doing. AMD is simply targeting

a different initial market and scaling

up or down from there. End quote. Early

marketing for the HD 4870X2 heavily

pushed Crossfire Sideport, a new

technology planned to release with the

card with AMD claiming it increase the

total interconnect bandwidth from 6.8 GB

per second to 21.8 GB per second, a

massive increase. Tech report explained

how it was supposed to work. reporting,

quote, "AMD says the side port is

electrically similar to PCI Express, but

is simpler because it's only intended as

a peer-to-peer link between GPUs. This

link augments the bandwidth already

available via the X2's Crossfire Bridge

interface, CFBI, in the diagram, which

is only used to pass final frames from

one GPU to the next for compositing and

its PCI Express lanes. The sideep port

connection should help improve

performance in cases where multiGPU

applications have typically had

performance scaling problems, such as

when texture synchronization between the

GPUs becomes a problem. End quote.

Despite its initial marketing, AMD

disabled the feature at launch due to

minimal performance uplift despite

increased power requirements and costs

for AIB vendors as reported via Anitech.

On the card's release day, PC

Perspective noted, quote, "The current

plan is to offer a software update for

the card that will enable the technology

for better performance. The problem is

that AMD is allowing board vendors an

option to include the Sideport

technology or leave it out for a cost

savings. This means that there'll be

some HD 4870X2 cards on the market

that'll be able to use Sideport

technology and some that will not.

distinguishing between the two will be

nearly impossible unless the card

vendors explicitly indicate their card

status. End quote. As far as we could

tell, AMD never enabled Sideport with

driver updates and ultimately abandoned

the technology after its HD 4800 series

launch. At that time, it would also have

been in the fight for its life as a

company. The big challenge with dual GPU

cards is that if you have a lot of

memory between two GPUs, it doesn't mean

one GPU has all that memory. So if

you've got 48 gigabytes on the board in

total, but it's 24 per GPU, although

still useful, it's not like you have a

single contiguous pool of 48 gigabytes

of memory. And that was true back here

too where GPUA can't use GPUB's memory

on this card if GPUB isn't doing

anything useful with it. So that's

always been a challenge for these dual

GPU solutions. Now, the performance of

most dual GPUs is also largely dependent

on driver support, which is kind of

shoddy, and whether the game or the

application knows how to utilize both

devices. So, despite their intrigue and

really a lot of the desire to make these

types of cards good, because it'd be

awesome if you could have obviously a

dual GPU card and then you put two of

them, now you got four and two slots.

Despite all of that, they've just never

been common and they've become even less

common over time. The first multiGPU

cards that we could remember were 3DFX's

Voodoo 2 and Quantum 3D's Obsidian 2

X24, both released in 1998. It wasn't

until around 2006 with Nvidia's 7950 GX2

that they really started to pick up

steam. The dual GPU trend would remain

for the next decade or so on and off

with the HD 4870X2 launching amidst the

height of the trend. Two of the last

dual GPU cards we can remember from that

time are Nvidia's GTX Titanz and AMD's

R9 295X2,

both released in 2014. Nvidia also

earlier than that had the GTX 690 in

2012. Since then, most dual GPUs have

been restricted to workstation cards

like AMD's Radeon Pro Duo cards,

Huawei's Ascend 300i Duo, and Maxon's

Intel Pro B60 Duel. Now getting into

trying to use it. Unfortunately, that is

where we ran into the big problems. We

could get the card to boot in a separate

slot as long as video was put out by

something else. The system recognized

the card and even powered it on in

various platforms we assembled with

different operating systems in our

attempts to boot with it. But the card

had issues putting out display and also

had issues getting drivers installed

onto the system. Additionally, any

attempts to force install the drivers

last officially supported on Windows 8

would crash our system and even

corrupted the operating system on

multiple occasions. Although in some

instances, like the Windows 8 9.1

version, nothing of value would be lost.

The engineering sample that we purchased

from a viewer did arrive with some

relatively minor shipping damage along

the PCIe bracket. It doesn't really seem

to be enough to affect the card's

functionality in any way. It's just the

metal that was affected that we can see.

But maybe it explains some of the

display troubles we ran into if there's

some underlying damage we can't see. We

did try a lot of things to get it to

work. So we tried Windows 11, Windows

10, Windows 8 varying versions and uh

experimented with PCIe bifurcation

support, checked on the support on the

motherboard. So that's another

interesting challenge of these dual GPU

devices is you're taking the PCIe slot,

you're basically splitting it into each

GPU is going to get eight lanes. if

you're on a buy 16 and that will also

require some form of logic for the

motherboard to do that. So, we checked

all that stuff and we just could not get

this thing to work. Maybe it was an

engineering sample for a reason. It

turns on, the fan spins, that's about

it. All right, so now we're going to

take apart the RV 770 engineering

sample. We'll start with kind of a walk

around of the card to see what the

design looked like. Really simple stuff.

So, this is a longer card with a

standard height uh PCIe slot. So they're

not protruding past the slot here. They

were running a blower fan and that's

just going to push the air straight

through over both GPU cores and out.

That would mean that this one runs

warmer, which it would do anyway because

it's further back, but also it's going

to get the heat dumped on it from this

core. The rest of the cooler design, I

mean, this is this is about as bad as

you can get in terms of thermal design,

um, even back then. But if they were

designing these to be stacked multiple

in a system, then blower is still the

best way to do it. Just because if you

got another card right here, there's no

room for the axial fans to bring air in

and and cool the heat sink. So the big

question is going to be whether this is

two or 4 GB of memory in total because

two is what it should be. But GPUZ was

showing this as four. We think it's a

software issue. If this has four for

real though physically, then that would

have been another reason. Maybe this was

an engineering sample. So, we'll start

with actually disassembling it. Now, I'm

going to track these on the mod mat,

which if you want to grab one of these

anti-static mats we make, they're on

store. Nexus.net. There's a grid here.

We're going to be tracking the front and

the back screws as I take them out. Just

because sometimes these engineering

samples, they have different types of

screws in weird locations. So, we got

Phillips one. Let's go around the whole

back plate first.

It's interesting to see the fat

capacitors externally.

All right, that's a good start.

Uh, this is better than the MSI card I

just took apart in the Cyber Power

pre-built review where that card used a

plastic back plate. This is at least

metal. So, this is aluminum. They're

sinking into this. not going to do a

whole lot, but since there's memory

directly on the back, there's nothing

else cooling it, which is going to be a

challenge to keep the heat down on

those. So, they need to do at least

something. And this is kind of the bare

minimum. So, we've got four modules for

each GPU. There's going to be some on

the other side as well. So, these are

debug LEDs that might be unique to this

card, uh, because in engineering

samples, they'll often put additional

LEDs or hookups. There's some unoccupied

four pin connector spots on the PCB

here. Over here, there's captain tape

covering two dip switches. So, that's

going to be unique to engineering card.

It's they're labeled one and two and

they say on and K. Those would have been

used for something in the debug and

engineering process. Let's get the

cooler off. Get a smaller driver.

Definitely not fully That is not fully

torqued. I'm not sure if that's because

someone else took it apart before us or

if they just didn't assemble it all the

way.

So, these are spring retention, which is

pretty much how it's been done for a

long time with GPUs. And then there's

just some plastic bumpers here. These

are probably to prevent a short. So,

those are just going to stop the uh the

bracket here, which is most likely

steel, from bridging any components on

the board. I'm wondering if the DIP

switch enables or disables one of the

GPUs.

So, this specific module isn't in the

data sheet. It's really close, though.

There's some There's a few that there's

four that have the exact same character

strain, but then one or two letters are

changed, which might just be to do with

pre-production versus production. Um, so

the format seems to match the one

gigabit modules. There's eight bits in a

bite. That would mean if there's four on

this side and there's four on the flip

side, that'd be eight total modules per

GPU, which would be one gigabyte per

GPU, which is what we would expect. This

is supposed to be a two total gigabyte

card, one per core. Uh, it does detect

as four in GPUZ, but unfortunately, it

looks like we do not have a gold mine of

memory right now in 2026 on this card,

but I think we've seen enough here to to

confirm that. So, now I'm going to take

apart the rest of this.

Oh, that was

a little crusty.

Yeah, look at that. That doesn't look

like a good quality copper. Yeah, we've

got thermal paste applied to two GPU

cores. This is pretty cool. Is that a

PEX chip? That might be a PEX chip.

Yeah. So, this is a PLX. This is a

multiplexer

for PCIe. So, this is what allows the

two GPU cores to share their lanes

through a single PCIe slot. These were

expensive. So, I don't know what they

cost at this point in time in history,

but the last time I looked up the price

of these like cost, they were $30 to $50

at the time. It was when they were

they're really in demand for specific

applications. The IHS for this is the

size is larger than the GPU silicon.

Then we've got our other four memory

modules per core here. So, that's going

to be eight total front and back, same

memory. So, uh that is going to be a two

total gigabyte card.

And then for the rest of this, it looks

like they've maybe changed some caps or

moved some things around. So between

these two power connectors, if we look

at the reference images on Tech

Power-Ups website where they have some

board shots, uh, this region looks

different. So that may be a that may be

unique for the engineering sample here.

I'm actually not sure what this is. I'm

going to have to look that up. I might

look that up. Let's look up you or V.

Oh, it's VTEC.

>> Honda

Variable cam timing.

>> H. Is that what that is?

>> E tech. Yeah.

>> Multi-phase SMD coupled inductor. No,

it's a coupled inductor.

>> That's not right.

>> That's not Yeah. Ron Honda.

Oh, these apparently these were

best known for use on the uh 4850 and

4870s. I'm just looking at some data

sheet entry for it. So, it's a coupled

inductor. Uh, it's part of the VRM. If

you look straight down here,

those are going to be the inductors

themselves. So, it looks like it might

be a group of I don't know, three or

something. Maybe maybe more than that.

But, that's just a big fat inductor

that's all grouped into one thing.

That's kind of cool. I haven't seen that

or I haven't seen it in a long time. If

I have Oh, another coupled component.

Yeah, that's another one. Oh, nice.

Plastic. There's plastic in the thermal

pad. That's good.

that'll conduct. So, this is going to be

uh three fats to the three inductors in

the coupled inductor. So, without

probing it directly, uh it's probably a

three-phase. And then we've got a plus

one over here. Maybe that's for memory.

It's possible that this is for memory.

One of these is going to be a

controller. One's going to be for

memory. There's another three phases

over here. That is probably for the uh

the other GPU. All right. So, this is

pretty easy. We're just going to see if

it says anything on the die. It does.

Nice. All right. Let's see what it says.

ATI

says site for sore eyes.

There's the ATI logo. So, this says

diffused in Taiwan. Made in Taiwan. So,

they were still they were already doing

stuff there. It does say engineering

sample

0818G

and

it's a B3 stepping whatever that would

have meant for this particular card. So

that's cool. This should be just the

same exact thing.

This might be the first pace it's gotten

in 20 years when we redo it. So same

thing just twice flipped. And uh

rotating it like that would allow them

to kind of rotate the memory and the VRM

and everything so that they can just

mirror them. This has some very orange

copper heat sinks. Normally that's not a

great sign.

Uh that can be an indicator of plating.

In this case, I I thought it was copper

plated, but um some crusty paste.

There's your heat sink. We're going to

go through and just pull the shell apart

to look at what the fin stack looks like

inside. Good use of thermal pads.

They've got a large thermal pad right in

the middle that would have been

contacting that PLX

chip. That's for the fan.

They really wanted this held together.

Couple things. First of all, they drove

that screw right through the memory

contact patch, the heat sink. So, a

little bit of lacking coverage there,

but uh steel screw doesn't really sink

that much heat. This is an aluminum base

plate. This is a kind of traditional

sign. Look at that. Look right. It was

perfect.

Again, you can grab your mod mat on the

store for moments like this. Uh aluminum

base plate. There you go. That's not

bad. They've got some fat fins here. the

the uh that's going to help with the

PLX. So, that is sitting right on top of

the PLX chip. They did get warm from

what I remember. I don't know what the

power was for them, but throwing pad to

it. And just as a reminder, that's what

that looks like. So, that larger square

pad contacts that PLX chip in the

middle. This is going to help sink the

heat. Wider fins will reduce the

resistance as you're pushing air through

the tighter fin stack here and then the

tighter fin stack here. This is uh those

pins we were looking at earlier. So,

those are going to help with memory.

There's a little bit of contact directly

under it to some of the MOSFETs um some

of the memory components like that

blower fan and uh otherwise this is all

they were cooling it with. This thing

would have run pretty hot. So, this was

also in an era where AMD and Nvidia

weren't shamed enough yet to really

improve the cooling on their GPUs. So,

that's going to be it for the history

lesson on this unique ATI AMD card right

after AMD bought ATI before they kind of

shuttered the brand and before one of

AMD's worst moments in its history

financially. Uh, after this would have

been bulldozer, pile driver,

steamroller, all of those CPUs before

the Ryzen launch in 2017 and the AM4

launch in 2016. So, this was sort of AMD

was doing okay and they managed to buy

ATI. that was big acquisition but then

they hit those they hit the 2008 crisis

not a good time for AMD and uh Nvidia

was much smaller at that time. So

historically really interesting point in

time for this card because this came out

like just before massive drops in AMD's

stock price and before AMD had to do a

lot of cutbacks which is why I think you

see some of the technologies that were

supposed to be in cards like this never

really coming to fruition because they

probably made a bunch of cuts to their

tech ambitions and just went towards

sort of trying to survive through the

2008 uh crisis. So really interesting

stuff. So that's going to be it for this

one. if you have any other older GPUs or

CPUs or anything you're interested in

where we could do a historical recap

plus maybe a tear down or something and

if we can get it working, which

unfortunately we couldn't get this one

to actually run anything successfully

without a major crash or corruption or

something. But if we can get it working,

let us know what kinds of parts that

you're interested in from older eras.

We'll see if we can get them or if we

already have them and maybe do a history

breakdown of it and some testing. And

otherwise, if you want to see more like

this, check out our 3DFX Voodoo GPU that

we looked at previously. That one was

handmade by someone who salvaged GPUs

from 3dfx that were never officially

released. We'll link that below. Thanks

for watching. Subscribe for more. You

can go to store.gamersac.net to support

us directly or patreon.com/gamersac.

and we'll see you all next time.